Whamcloud - gitweb
LU-14540 o2iblnd: Use REMOTE_DROPPED for ECONNREFUSED
[fs/lustre-release.git] / lnet / klnds / o2iblnd / o2iblnd_cb.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lnet/klnds/o2iblnd/o2iblnd_cb.c
33  *
34  * Author: Eric Barton <eric@bartonsoftware.com>
35  */
36
37 #include "o2iblnd.h"
38
39 #define MAX_CONN_RACES_BEFORE_ABORT 20
40
41 static void kiblnd_peer_alive(struct kib_peer_ni *peer_ni);
42 static void kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
43                                        int error);
44 static void kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx,
45                                int type, int body_nob);
46 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
47                             int resid, struct kib_rdma_desc *dstrd, u64 dstcookie);
48 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
49 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
50
51 static void kiblnd_unmap_tx(struct kib_tx *tx);
52 static void kiblnd_check_sends_locked(struct kib_conn *conn);
53
54 void
55 kiblnd_tx_done(struct kib_tx *tx)
56 {
57         struct lnet_msg *lntmsg[2];
58         int         rc;
59         int         i;
60
61         LASSERT (!in_interrupt());
62         LASSERT (!tx->tx_queued);               /* mustn't be queued for sending */
63         LASSERT (tx->tx_sending == 0);          /* mustn't be awaiting sent callback */
64         LASSERT (!tx->tx_waiting);              /* mustn't be awaiting peer_ni response */
65         LASSERT (tx->tx_pool != NULL);
66
67         kiblnd_unmap_tx(tx);
68
69         /* tx may have up to 2 lnet msgs to finalise */
70         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
71         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
72         rc = tx->tx_status;
73
74         if (tx->tx_conn != NULL) {
75                 kiblnd_conn_decref(tx->tx_conn);
76                 tx->tx_conn = NULL;
77         }
78
79         tx->tx_nwrq = tx->tx_nsge = 0;
80         tx->tx_status = 0;
81
82         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
83
84         /* delay finalize until my descs have been freed */
85         for (i = 0; i < 2; i++) {
86                 if (lntmsg[i] == NULL)
87                         continue;
88
89                 /* propagate health status to LNet for requests */
90                 if (i == 0 && lntmsg[i])
91                         lntmsg[i]->msg_health_status = tx->tx_hstatus;
92
93                 lnet_finalize(lntmsg[i], rc);
94         }
95 }
96
97 void
98 kiblnd_txlist_done(struct list_head *txlist, int status,
99                    enum lnet_msg_hstatus hstatus)
100 {
101         struct kib_tx *tx;
102
103         while (!list_empty(txlist)) {
104                 tx = list_entry(txlist->next, struct kib_tx, tx_list);
105
106                 list_del(&tx->tx_list);
107                 /* complete now */
108                 tx->tx_waiting = 0;
109                 tx->tx_status = status;
110                 if (hstatus != LNET_MSG_STATUS_OK)
111                         tx->tx_hstatus = hstatus;
112                 kiblnd_tx_done(tx);
113         }
114 }
115
116 static struct kib_tx *
117 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
118 {
119         struct kib_net *net = ni->ni_data;
120         struct list_head *node;
121         struct kib_tx *tx;
122         struct kib_tx_poolset *tps;
123
124         tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
125         node = kiblnd_pool_alloc_node(&tps->tps_poolset);
126         if (node == NULL)
127                 return NULL;
128         tx = container_of(node, struct kib_tx, tx_list);
129
130         LASSERT (tx->tx_nwrq == 0);
131         LASSERT (!tx->tx_queued);
132         LASSERT (tx->tx_sending == 0);
133         LASSERT (!tx->tx_waiting);
134         LASSERT (tx->tx_status == 0);
135         LASSERT (tx->tx_conn == NULL);
136         LASSERT (tx->tx_lntmsg[0] == NULL);
137         LASSERT (tx->tx_lntmsg[1] == NULL);
138         LASSERT (tx->tx_nfrags == 0);
139
140         tx->tx_gaps = false;
141         tx->tx_hstatus = LNET_MSG_STATUS_OK;
142
143         return tx;
144 }
145
146 static void
147 kiblnd_drop_rx(struct kib_rx *rx)
148 {
149         struct kib_conn *conn = rx->rx_conn;
150         struct kib_sched_info *sched = conn->ibc_sched;
151         unsigned long flags;
152
153         spin_lock_irqsave(&sched->ibs_lock, flags);
154         LASSERT(conn->ibc_nrx > 0);
155         conn->ibc_nrx--;
156         spin_unlock_irqrestore(&sched->ibs_lock, flags);
157
158         kiblnd_conn_decref(conn);
159 }
160
161 int
162 kiblnd_post_rx(struct kib_rx *rx, int credit)
163 {
164         struct kib_conn *conn = rx->rx_conn;
165         struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
166         struct ib_recv_wr *bad_wrq = NULL;
167 #ifdef HAVE_IB_GET_DMA_MR
168         struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
169 #endif
170         int rc;
171
172         LASSERT (net != NULL);
173         LASSERT (!in_interrupt());
174         LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
175                  credit == IBLND_POSTRX_PEER_CREDIT ||
176                  credit == IBLND_POSTRX_RSRVD_CREDIT);
177 #ifdef HAVE_IB_GET_DMA_MR
178         LASSERT(mr != NULL);
179
180         rx->rx_sge.lkey   = mr->lkey;
181 #else
182         rx->rx_sge.lkey   = conn->ibc_hdev->ibh_pd->local_dma_lkey;
183 #endif
184         rx->rx_sge.addr   = rx->rx_msgaddr;
185         rx->rx_sge.length = IBLND_MSG_SIZE;
186
187         rx->rx_wrq.next = NULL;
188         rx->rx_wrq.sg_list = &rx->rx_sge;
189         rx->rx_wrq.num_sge = 1;
190         rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
191
192         LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
193         LASSERT (rx->rx_nob >= 0);              /* not posted */
194
195         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
196                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
197                 return 0;
198         }
199
200         rx->rx_nob = -1;                        /* flag posted */
201
202         /* NB: need an extra reference after ib_post_recv because we don't
203          * own this rx (and rx::rx_conn) anymore, LU-5678.
204          */
205         kiblnd_conn_addref(conn);
206 #ifdef HAVE_IB_POST_SEND_RECV_CONST
207         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq,
208                           (const struct ib_recv_wr **)&bad_wrq);
209 #else
210         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
211 #endif
212         if (unlikely(rc != 0)) {
213                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
214                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
215                 rx->rx_nob = 0;
216         }
217
218         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
219                 goto out;
220
221         if (unlikely(rc != 0)) {
222                 kiblnd_close_conn(conn, rc);
223                 kiblnd_drop_rx(rx);     /* No more posts for this rx */
224                 goto out;
225         }
226
227         if (credit == IBLND_POSTRX_NO_CREDIT)
228                 goto out;
229
230         spin_lock(&conn->ibc_lock);
231         if (credit == IBLND_POSTRX_PEER_CREDIT)
232                 conn->ibc_outstanding_credits++;
233         else
234                 conn->ibc_reserved_credits++;
235         kiblnd_check_sends_locked(conn);
236         spin_unlock(&conn->ibc_lock);
237
238 out:
239         kiblnd_conn_decref(conn);
240         return rc;
241 }
242
243 static struct kib_tx *
244 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, u64 cookie)
245 {
246         struct list_head *tmp;
247
248         list_for_each(tmp, &conn->ibc_active_txs) {
249                 struct kib_tx *tx = list_entry(tmp, struct kib_tx, tx_list);
250
251                 LASSERT(!tx->tx_queued);
252                 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
253
254                 if (tx->tx_cookie != cookie)
255                         continue;
256
257                 if (tx->tx_waiting &&
258                     tx->tx_msg->ibm_type == txtype)
259                         return tx;
260
261                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
262                       tx->tx_waiting ? "" : "NOT ",
263                       tx->tx_msg->ibm_type, txtype);
264         }
265         return NULL;
266 }
267
268 static void
269 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, u64 cookie)
270 {
271         struct kib_tx *tx;
272         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
273         int idle;
274
275         spin_lock(&conn->ibc_lock);
276
277         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
278         if (tx == NULL) {
279                 spin_unlock(&conn->ibc_lock);
280
281                 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
282                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
283                 kiblnd_close_conn(conn, -EPROTO);
284                 return;
285         }
286
287         if (tx->tx_status == 0) {               /* success so far */
288                 if (status < 0) {               /* failed? */
289                         tx->tx_status = status;
290                         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
291                 } else if (txtype == IBLND_MSG_GET_REQ) {
292                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
293                 }
294         }
295
296         tx->tx_waiting = 0;
297
298         idle = !tx->tx_queued && (tx->tx_sending == 0);
299         if (idle)
300                 list_del(&tx->tx_list);
301
302         spin_unlock(&conn->ibc_lock);
303
304         if (idle)
305                 kiblnd_tx_done(tx);
306 }
307
308 static void
309 kiblnd_send_completion(struct kib_conn *conn, int type, int status, u64 cookie)
310 {
311         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
312         struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
313
314         if (tx == NULL) {
315                 CERROR("Can't get tx for completion %x for %s\n",
316                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
317                 return;
318         }
319
320         tx->tx_msg->ibm_u.completion.ibcm_status = status;
321         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
322         kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
323
324         kiblnd_queue_tx(tx, conn);
325 }
326
327 static void
328 kiblnd_handle_rx(struct kib_rx *rx)
329 {
330         struct kib_msg *msg = rx->rx_msg;
331         struct kib_conn   *conn = rx->rx_conn;
332         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
333         int           credits = msg->ibm_credits;
334         struct kib_tx *tx;
335         int           rc = 0;
336         int           rc2;
337         int           post_credit;
338
339         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
340
341         CDEBUG (D_NET, "Received %x[%d] from %s\n",
342                 msg->ibm_type, credits,
343                 libcfs_nid2str(conn->ibc_peer->ibp_nid));
344
345         if (credits != 0) {
346                 /* Have I received credits that will let me send? */
347                 spin_lock(&conn->ibc_lock);
348
349                 if (conn->ibc_credits + credits >
350                     conn->ibc_queue_depth) {
351                         rc2 = conn->ibc_credits;
352                         spin_unlock(&conn->ibc_lock);
353
354                         CERROR("Bad credits from %s: %d + %d > %d\n",
355                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
356                                rc2, credits,
357                                conn->ibc_queue_depth);
358
359                         kiblnd_close_conn(conn, -EPROTO);
360                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
361                         return;
362                 }
363
364                 conn->ibc_credits += credits;
365
366                 /* This ensures the credit taken by NOOP can be returned */
367                 if (msg->ibm_type == IBLND_MSG_NOOP &&
368                     !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
369                         conn->ibc_outstanding_credits++;
370
371                 kiblnd_check_sends_locked(conn);
372                 spin_unlock(&conn->ibc_lock);
373         }
374
375         switch (msg->ibm_type) {
376         default:
377                 CERROR("Bad IBLND message type %x from %s\n",
378                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
379                 post_credit = IBLND_POSTRX_NO_CREDIT;
380                 rc = -EPROTO;
381                 break;
382
383         case IBLND_MSG_NOOP:
384                 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
385                         post_credit = IBLND_POSTRX_NO_CREDIT;
386                         break;
387                 }
388
389                 if (credits != 0) /* credit already posted */
390                         post_credit = IBLND_POSTRX_NO_CREDIT;
391                 else              /* a keepalive NOOP */
392                         post_credit = IBLND_POSTRX_PEER_CREDIT;
393                 break;
394
395         case IBLND_MSG_IMMEDIATE:
396                 post_credit = IBLND_POSTRX_DONT_POST;
397                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
398                                 msg->ibm_srcnid, rx, 0);
399                 if (rc < 0)                     /* repost on error */
400                         post_credit = IBLND_POSTRX_PEER_CREDIT;
401                 break;
402
403         case IBLND_MSG_PUT_REQ:
404                 post_credit = IBLND_POSTRX_DONT_POST;
405                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
406                                 msg->ibm_srcnid, rx, 1);
407                 if (rc < 0)                     /* repost on error */
408                         post_credit = IBLND_POSTRX_PEER_CREDIT;
409                 break;
410
411         case IBLND_MSG_PUT_NAK:
412                 CWARN ("PUT_NACK from %s\n",
413                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
414                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
415                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
416                                          msg->ibm_u.completion.ibcm_status,
417                                          msg->ibm_u.completion.ibcm_cookie);
418                 break;
419
420         case IBLND_MSG_PUT_ACK:
421                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
422
423                 spin_lock(&conn->ibc_lock);
424                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
425                                         msg->ibm_u.putack.ibpam_src_cookie);
426                 if (tx != NULL)
427                         list_del(&tx->tx_list);
428                 spin_unlock(&conn->ibc_lock);
429
430                 if (tx == NULL) {
431                         CERROR("Unmatched PUT_ACK from %s\n",
432                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
433                         rc = -EPROTO;
434                         break;
435                 }
436
437                 LASSERT (tx->tx_waiting);
438                 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
439                  * (a) I can overwrite tx_msg since my peer_ni has received it!
440                  * (b) tx_waiting set tells tx_complete() it's not done. */
441
442                 tx->tx_nwrq = tx->tx_nsge = 0;  /* overwrite PUT_REQ */
443
444                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
445                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
446                                        &msg->ibm_u.putack.ibpam_rd,
447                                        msg->ibm_u.putack.ibpam_dst_cookie);
448                 if (rc2 < 0)
449                         CERROR("Can't setup rdma for PUT to %s: %d\n",
450                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
451
452                 spin_lock(&conn->ibc_lock);
453                 tx->tx_waiting = 0;     /* clear waiting and queue atomically */
454                 kiblnd_queue_tx_locked(tx, conn);
455                 spin_unlock(&conn->ibc_lock);
456                 break;
457
458         case IBLND_MSG_PUT_DONE:
459                 post_credit = IBLND_POSTRX_PEER_CREDIT;
460                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
461                                          msg->ibm_u.completion.ibcm_status,
462                                          msg->ibm_u.completion.ibcm_cookie);
463                 break;
464
465         case IBLND_MSG_GET_REQ:
466                 post_credit = IBLND_POSTRX_DONT_POST;
467                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
468                                 msg->ibm_srcnid, rx, 1);
469                 if (rc < 0)                     /* repost on error */
470                         post_credit = IBLND_POSTRX_PEER_CREDIT;
471                 break;
472
473         case IBLND_MSG_GET_DONE:
474                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
475                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
476                                          msg->ibm_u.completion.ibcm_status,
477                                          msg->ibm_u.completion.ibcm_cookie);
478                 break;
479         }
480
481         if (rc < 0)                             /* protocol error */
482                 kiblnd_close_conn(conn, rc);
483
484         if (post_credit != IBLND_POSTRX_DONT_POST)
485                 kiblnd_post_rx(rx, post_credit);
486 }
487
488 static void
489 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
490 {
491         struct kib_msg *msg = rx->rx_msg;
492         struct kib_conn   *conn = rx->rx_conn;
493         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
494         struct kib_net *net = ni->ni_data;
495         int rc;
496         int err = -EIO;
497
498         LASSERT (net != NULL);
499         LASSERT (rx->rx_nob < 0);               /* was posted */
500         rx->rx_nob = 0;                         /* isn't now */
501
502         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
503                 goto ignore;
504
505         if (status != IB_WC_SUCCESS) {
506                 CNETERR("Rx from %s failed: %d\n",
507                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
508                 goto failed;
509         }
510
511         LASSERT (nob >= 0);
512         rx->rx_nob = nob;
513
514         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
515         if (rc != 0) {
516                 CERROR ("Error %d unpacking rx from %s\n",
517                         rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
518                 goto failed;
519         }
520
521         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
522             msg->ibm_dstnid != ni->ni_nid ||
523             msg->ibm_srcstamp != conn->ibc_incarnation ||
524             msg->ibm_dststamp != net->ibn_incarnation) {
525                 CERROR ("Stale rx from %s\n",
526                         libcfs_nid2str(conn->ibc_peer->ibp_nid));
527                 err = -ESTALE;
528                 goto failed;
529         }
530
531         /* set time last known alive */
532         kiblnd_peer_alive(conn->ibc_peer);
533
534         /* racing with connection establishment/teardown! */
535
536         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
537                 rwlock_t  *g_lock = &kiblnd_data.kib_global_lock;
538                 unsigned long  flags;
539
540                 write_lock_irqsave(g_lock, flags);
541                 /* must check holding global lock to eliminate race */
542                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
543                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
544                         write_unlock_irqrestore(g_lock, flags);
545                         return;
546                 }
547                 write_unlock_irqrestore(g_lock, flags);
548         }
549         kiblnd_handle_rx(rx);
550         return;
551
552  failed:
553         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
554         kiblnd_close_conn(conn, err);
555  ignore:
556         kiblnd_drop_rx(rx);                     /* Don't re-post rx. */
557 }
558
559 static int
560 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx,
561                   struct kib_rdma_desc *rd, u32 nob)
562 {
563         struct kib_hca_dev *hdev;
564         struct kib_dev *dev;
565         struct kib_fmr_poolset *fps;
566         int                     cpt;
567         int                     rc;
568         int i;
569
570         LASSERT(tx->tx_pool != NULL);
571         LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
572
573         dev = net->ibn_dev;
574         hdev = tx->tx_pool->tpo_hdev;
575         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
576
577         /*
578          * If we're dealing with FastReg, but the device doesn't
579          * support GAPS and the tx has GAPS, then there is no real point
580          * in trying to map the memory, because it'll just fail. So
581          * preemptively fail with an appropriate message
582          */
583         if ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) &&
584             !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
585             tx->tx_gaps) {
586                 CERROR("Using FastReg with no GAPS support, but tx has gaps. "
587                        "Try setting use_fastreg_gaps to 1\n");
588                 return -EPROTONOSUPPORT;
589         }
590
591 #ifdef HAVE_FMR_POOL_API
592         /*
593          * FMR does not support gaps but the tx has gaps then
594          * we should make sure that the number of fragments we'll be sending
595          * over fits within the number of fragments negotiated on the
596          * connection, otherwise, we won't be able to RDMA the data.
597          * We need to maintain the number of fragments negotiation on the
598          * connection for backwards compatibility.
599          */
600         if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
601                 if (tx->tx_conn &&
602                     tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
603                         CERROR("TX number of frags (%d) is <= than connection"
604                                " number of frags (%d). Consider setting peer's"
605                                " map_on_demand to 256\n", tx->tx_nfrags,
606                                tx->tx_conn->ibc_max_frags);
607                         return -EFBIG;
608                 }
609         }
610 #endif
611
612         fps = net->ibn_fmr_ps[cpt];
613         rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
614         if (rc != 0) {
615                 CERROR("Can't map %u bytes (%u/%u)s: %d\n", nob,
616                        tx->tx_nfrags, rd->rd_nfrags, rc);
617                 return rc;
618         }
619
620         /*
621          * If rd is not tx_rd, it's going to get sent to a peer_ni, who will
622          * need the rkey
623          */
624         rd->rd_key = tx->tx_fmr.fmr_key;
625         /*
626          * for FastReg or FMR with no gaps we can accumulate all
627          * the fragments in one FastReg or FMR fragment.
628          */
629         if (
630 #ifdef HAVE_FMR_POOL_API
631             ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
632              && !tx->tx_gaps) ||
633 #endif
634             (dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)) {
635                 /* FMR requires zero based address */
636 #ifdef HAVE_FMR_POOL_API
637                 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
638                         rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
639 #endif
640                 rd->rd_frags[0].rf_nob = nob;
641                 rd->rd_nfrags = 1;
642         } else {
643                 /*
644                  * We're transmitting with gaps using FMR.
645                  * We'll need to use multiple fragments and identify the
646                  * zero based address of each fragment.
647                  */
648                 for (i = 0; i < rd->rd_nfrags; i++) {
649                         rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
650                         rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
651                 }
652         }
653
654         return 0;
655 }
656
657 static void
658 kiblnd_unmap_tx(struct kib_tx *tx)
659 {
660         if (
661 #ifdef HAVE_FMR_POOL_API
662                 tx->tx_fmr.fmr_pfmr ||
663 #endif
664                 tx->tx_fmr.fmr_frd)
665                 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
666
667         if (tx->tx_nfrags != 0) {
668                 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
669                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
670                 tx->tx_nfrags = 0;
671         }
672 }
673
674 #ifdef HAVE_IB_GET_DMA_MR
675 static struct ib_mr *
676 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd)
677 {
678         struct kib_net *net = ni->ni_data;
679         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
680         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
681
682         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
683
684         /*
685          * if map-on-demand is turned on and the device supports
686          * either FMR or FastReg then use that. Otherwise use global
687          * memory regions. If that's not available either, then you're
688          * dead in the water and fail the operation.
689          */
690         if (tunables->lnd_map_on_demand &&
691             (net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED
692 #ifdef HAVE_FMR_POOL_API
693              || net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED
694 #endif
695         ))
696                 return NULL;
697
698         /*
699          * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
700          * in the call chain. The mapping will fail with appropriate error
701          * message.
702          */
703         return hdev->ibh_mrs;
704 }
705 #endif
706
707 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
708                          struct kib_rdma_desc *rd, int nfrags)
709 {
710         struct kib_net *net = ni->ni_data;
711         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
712 #ifdef HAVE_IB_GET_DMA_MR
713         struct ib_mr *mr = NULL;
714 #endif
715         __u32 nob;
716         int i;
717
718         /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
719          * RDMA sink */
720         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
721         tx->tx_nfrags = nfrags;
722
723         rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
724                                           tx->tx_nfrags, tx->tx_dmadir);
725
726         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
727                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
728                         hdev->ibh_ibdev, &tx->tx_frags[i]);
729                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
730                         hdev->ibh_ibdev, &tx->tx_frags[i]);
731                 nob += rd->rd_frags[i].rf_nob;
732         }
733
734 #ifdef HAVE_IB_GET_DMA_MR
735         mr = kiblnd_find_rd_dma_mr(ni, rd);
736         if (mr != NULL) {
737                 /* found pre-mapping MR */
738                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
739                 return 0;
740         }
741 #endif
742
743         if (net->ibn_fmr_ps != NULL)
744                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
745
746         return -EINVAL;
747 }
748
749 static int kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
750                                 struct kib_rdma_desc *rd, int nkiov,
751                                 struct bio_vec *kiov, int offset, int nob)
752 {
753         struct kib_net *net = ni->ni_data;
754         struct scatterlist *sg;
755         int                 fragnob;
756         int                 max_nkiov;
757
758         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
759
760         LASSERT(nob > 0);
761         LASSERT(nkiov > 0);
762         LASSERT(net != NULL);
763
764         while (offset >= kiov->bv_len) {
765                 offset -= kiov->bv_len;
766                 nkiov--;
767                 kiov++;
768                 LASSERT(nkiov > 0);
769         }
770
771         max_nkiov = nkiov;
772
773         sg = tx->tx_frags;
774         do {
775                 LASSERT(nkiov > 0);
776
777                 fragnob = min((int)(kiov->bv_len - offset), nob);
778
779                 /*
780                  * We're allowed to start at a non-aligned page offset in
781                  * the first fragment and end at a non-aligned page offset
782                  * in the last fragment.
783                  */
784                 if ((fragnob < (int)(kiov->bv_len - offset)) &&
785                     nkiov < max_nkiov && nob > fragnob) {
786                         CDEBUG(D_NET, "fragnob %d < available page %d: with"
787                                       " remaining %d kiovs with %d nob left\n",
788                                fragnob, (int)(kiov->bv_len - offset),
789                                nkiov, nob);
790                         tx->tx_gaps = true;
791                 }
792
793                 sg_set_page(sg, kiov->bv_page, fragnob,
794                             kiov->bv_offset + offset);
795                 sg = sg_next(sg);
796                 if (!sg) {
797                         CERROR("lacking enough sg entries to map tx\n");
798                         return -EFAULT;
799                 }
800
801                 offset = 0;
802                 kiov++;
803                 nkiov--;
804                 nob -= fragnob;
805         } while (nob > 0);
806
807         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
808 }
809
810 static int
811 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
812 __must_hold(&conn->ibc_lock)
813 {
814         struct kib_msg *msg = tx->tx_msg;
815         struct kib_peer_ni *peer_ni = conn->ibc_peer;
816         struct lnet_ni *ni = peer_ni->ibp_ni;
817         int ver = conn->ibc_version;
818         int rc;
819         int done;
820
821         LASSERT(tx->tx_queued);
822         /* We rely on this for QP sizing */
823         LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
824         LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
825
826         LASSERT(credit == 0 || credit == 1);
827         LASSERT(conn->ibc_outstanding_credits >= 0);
828         LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
829         LASSERT(conn->ibc_credits >= 0);
830         LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
831
832         if (conn->ibc_nsends_posted ==
833             kiblnd_concurrent_sends(ver, ni)) {
834                 /* tx completions outstanding... */
835                 CDEBUG(D_NET, "%s: posted enough\n",
836                        libcfs_nid2str(peer_ni->ibp_nid));
837                 return -EAGAIN;
838         }
839
840         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
841                 CDEBUG(D_NET, "%s: no credits\n",
842                        libcfs_nid2str(peer_ni->ibp_nid));
843                 return -EAGAIN;
844         }
845
846         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
847             conn->ibc_credits == 1 &&   /* last credit reserved */
848             msg->ibm_type != IBLND_MSG_NOOP) {      /* for NOOP */
849                 CDEBUG(D_NET, "%s: not using last credit\n",
850                        libcfs_nid2str(peer_ni->ibp_nid));
851                 return -EAGAIN;
852         }
853
854         /* NB don't drop ibc_lock before bumping tx_sending */
855         list_del(&tx->tx_list);
856         tx->tx_queued = 0;
857
858         if (msg->ibm_type == IBLND_MSG_NOOP &&
859             (!kiblnd_need_noop(conn) ||     /* redundant NOOP */
860              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
861               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
862                 /* OK to drop when posted enough NOOPs, since
863                  * kiblnd_check_sends_locked will queue NOOP again when
864                  * posted NOOPs complete */
865                 spin_unlock(&conn->ibc_lock);
866                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
867                 kiblnd_tx_done(tx);
868                 spin_lock(&conn->ibc_lock);
869                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
870                        libcfs_nid2str(peer_ni->ibp_nid),
871                        conn->ibc_noops_posted);
872                 return 0;
873         }
874
875         kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
876                         peer_ni->ibp_nid, conn->ibc_incarnation);
877
878         conn->ibc_credits -= credit;
879         conn->ibc_outstanding_credits = 0;
880         conn->ibc_nsends_posted++;
881         if (msg->ibm_type == IBLND_MSG_NOOP)
882                 conn->ibc_noops_posted++;
883
884         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
885          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
886          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
887          * and then re-queued here.  It's (just) possible that
888          * tx_sending is non-zero if we've not done the tx_complete()
889          * from the first send; hence the ++ rather than = below. */
890         tx->tx_sending++;
891         list_add(&tx->tx_list, &conn->ibc_active_txs);
892
893         /* I'm still holding ibc_lock! */
894         if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
895                 rc = -ECONNABORTED;
896         } else if (tx->tx_pool->tpo_pool.po_failed ||
897                  conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
898                 /* close_conn will launch failover */
899                 rc = -ENETDOWN;
900         } else {
901                 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
902                 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
903                 struct ib_send_wr *wr  = &tx->tx_wrq[0].wr;
904
905                 if (frd != NULL) {
906                         if (!frd->frd_valid) {
907                                 wr = &frd->frd_inv_wr.wr;
908                                 wr->next = &frd->frd_fastreg_wr.wr;
909                         } else {
910                                 wr = &frd->frd_fastreg_wr.wr;
911                         }
912                         frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
913                 }
914
915                 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
916                          "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
917                          bad->wr_id, bad->opcode, bad->send_flags,
918                          libcfs_nid2str(conn->ibc_peer->ibp_nid));
919
920                 bad = NULL;
921                 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
922                         rc = -EINVAL;
923                 else
924 #ifdef HAVE_IB_POST_SEND_RECV_CONST
925                         rc = ib_post_send(conn->ibc_cmid->qp, wr,
926                                           (const struct ib_send_wr **)&bad);
927 #else
928                         rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
929 #endif
930         }
931
932         conn->ibc_last_send = ktime_get();
933
934         if (rc == 0)
935                 return 0;
936
937         /* NB credits are transferred in the actual
938          * message, which can only be the last work item */
939         conn->ibc_credits += credit;
940         conn->ibc_outstanding_credits += msg->ibm_credits;
941         conn->ibc_nsends_posted--;
942         if (msg->ibm_type == IBLND_MSG_NOOP)
943                 conn->ibc_noops_posted--;
944
945         tx->tx_status = rc;
946         tx->tx_waiting = 0;
947         tx->tx_sending--;
948
949         done = (tx->tx_sending == 0);
950         if (done)
951                 list_del(&tx->tx_list);
952
953         spin_unlock(&conn->ibc_lock);
954
955         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
956                 CERROR("Error %d posting transmit to %s\n",
957                        rc, libcfs_nid2str(peer_ni->ibp_nid));
958         else
959                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
960                        rc, libcfs_nid2str(peer_ni->ibp_nid));
961
962         kiblnd_close_conn(conn, rc);
963
964         if (done)
965                 kiblnd_tx_done(tx);
966
967         spin_lock(&conn->ibc_lock);
968
969         return -EIO;
970 }
971
972 static void
973 kiblnd_check_sends_locked(struct kib_conn *conn)
974 {
975         int ver = conn->ibc_version;
976         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
977         struct kib_tx *tx;
978
979         /* Don't send anything until after the connection is established */
980         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
981                 CDEBUG(D_NET, "%s too soon\n",
982                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
983                 return;
984         }
985
986         LASSERT(conn->ibc_nsends_posted <=
987                 kiblnd_concurrent_sends(ver, ni));
988         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
989                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
990         LASSERT (conn->ibc_reserved_credits >= 0);
991
992         while (conn->ibc_reserved_credits > 0 &&
993                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
994                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
995                                 struct kib_tx, tx_list);
996                 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
997                 conn->ibc_reserved_credits--;
998         }
999
1000         if (kiblnd_need_noop(conn)) {
1001                 spin_unlock(&conn->ibc_lock);
1002
1003                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1004                 if (tx != NULL)
1005                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
1006
1007                 spin_lock(&conn->ibc_lock);
1008                 if (tx != NULL)
1009                         kiblnd_queue_tx_locked(tx, conn);
1010         }
1011
1012         for (;;) {
1013                 int credit;
1014
1015                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1016                         credit = 0;
1017                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
1018                                         struct kib_tx, tx_list);
1019                 } else if (!list_empty(&conn->ibc_tx_noops)) {
1020                         LASSERT (!IBLND_OOB_CAPABLE(ver));
1021                         credit = 1;
1022                         tx = list_entry(conn->ibc_tx_noops.next,
1023                                         struct kib_tx, tx_list);
1024                 } else if (!list_empty(&conn->ibc_tx_queue)) {
1025                         credit = 1;
1026                         tx = list_entry(conn->ibc_tx_queue.next,
1027                                         struct kib_tx, tx_list);
1028                 } else
1029                         break;
1030
1031                 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1032                         break;
1033         }
1034 }
1035
1036 static void
1037 kiblnd_tx_complete(struct kib_tx *tx, int status)
1038 {
1039         int           failed = (status != IB_WC_SUCCESS);
1040         struct kib_conn   *conn = tx->tx_conn;
1041         int           idle;
1042
1043         if (tx->tx_sending <= 0) {
1044                 CERROR("Received an event on a freed tx: %p status %d\n",
1045                        tx, tx->tx_status);
1046                 return;
1047         }
1048
1049         if (failed) {
1050                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1051                         CNETERR("Tx -> %s cookie %#llx"
1052                                 " sending %d waiting %d: failed %d\n",
1053                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1054                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1055                                 status);
1056
1057                 kiblnd_close_conn(conn, -EIO);
1058         } else {
1059                 kiblnd_peer_alive(conn->ibc_peer);
1060         }
1061
1062         spin_lock(&conn->ibc_lock);
1063
1064         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
1065          * gets to free it, which also drops its ref on 'conn'. */
1066
1067         tx->tx_sending--;
1068         conn->ibc_nsends_posted--;
1069         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1070                 conn->ibc_noops_posted--;
1071
1072         if (failed) {
1073                 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1074                 tx->tx_waiting = 0;             /* don't wait for peer_ni */
1075                 tx->tx_status = -EIO;
1076         }
1077
1078         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
1079                !tx->tx_waiting &&               /* Not waiting for peer_ni */
1080                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
1081         if (idle)
1082                 list_del(&tx->tx_list);
1083
1084         kiblnd_check_sends_locked(conn);
1085         spin_unlock(&conn->ibc_lock);
1086
1087         if (idle)
1088                 kiblnd_tx_done(tx);
1089 }
1090
1091 static void
1092 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1093                    int body_nob)
1094 {
1095         struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1096         struct ib_sge *sge = &tx->tx_msgsge;
1097         struct ib_rdma_wr *wrq;
1098         int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1099 #ifdef HAVE_IB_GET_DMA_MR
1100         struct ib_mr *mr = hdev->ibh_mrs;
1101 #endif
1102
1103         LASSERT(tx->tx_nwrq >= 0);
1104         LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1105         LASSERT(nob <= IBLND_MSG_SIZE);
1106 #ifdef HAVE_IB_GET_DMA_MR
1107         LASSERT(mr != NULL);
1108 #endif
1109
1110         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1111
1112 #ifdef HAVE_IB_GET_DMA_MR
1113         sge->lkey   = mr->lkey;
1114 #else
1115         sge->lkey   = hdev->ibh_pd->local_dma_lkey;
1116 #endif
1117         sge->addr   = tx->tx_msgaddr;
1118         sge->length = nob;
1119
1120         wrq = &tx->tx_wrq[tx->tx_nwrq];
1121         memset(wrq, 0, sizeof(*wrq));
1122
1123         wrq->wr.next            = NULL;
1124         wrq->wr.wr_id           = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1125         wrq->wr.sg_list         = sge;
1126         wrq->wr.num_sge         = 1;
1127         wrq->wr.opcode          = IB_WR_SEND;
1128         wrq->wr.send_flags      = IB_SEND_SIGNALED;
1129
1130         tx->tx_nwrq++;
1131 }
1132
1133 static int
1134 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1135                  int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1136 {
1137         struct kib_msg *ibmsg = tx->tx_msg;
1138         struct kib_rdma_desc *srcrd = tx->tx_rd;
1139         struct ib_rdma_wr *wrq = NULL;
1140         struct ib_sge     *sge;
1141         int                rc  = resid;
1142         int                srcidx;
1143         int                dstidx;
1144         int                sge_nob;
1145         int                wrq_sge;
1146
1147         LASSERT(!in_interrupt());
1148         LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1149         LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1150
1151         for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1152              resid > 0; resid -= sge_nob) {
1153                 int     prev = dstidx;
1154
1155                 if (srcidx >= srcrd->rd_nfrags) {
1156                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1157                         rc = -EPROTO;
1158                         break;
1159                 }
1160
1161                 if (dstidx >= dstrd->rd_nfrags) {
1162                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1163                         rc = -EPROTO;
1164                         break;
1165                 }
1166
1167                 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1168                         CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1169                                "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1170                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1171                                conn->ibc_max_frags,
1172                                srcidx, srcrd->rd_nfrags,
1173                                dstidx, dstrd->rd_nfrags);
1174                         rc = -EMSGSIZE;
1175                         break;
1176                 }
1177
1178                 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1179                                kiblnd_rd_frag_size(dstrd, dstidx),
1180                                resid);
1181
1182                 sge = &tx->tx_sge[tx->tx_nsge];
1183                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1184                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1185                 sge->length = sge_nob;
1186
1187                 if (wrq_sge == 0) {
1188                         wrq = &tx->tx_wrq[tx->tx_nwrq];
1189
1190                         wrq->wr.next    = &(wrq + 1)->wr;
1191                         wrq->wr.wr_id   = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1192                         wrq->wr.sg_list = sge;
1193                         wrq->wr.opcode  = IB_WR_RDMA_WRITE;
1194                         wrq->wr.send_flags = 0;
1195
1196 #ifdef HAVE_IB_RDMA_WR
1197                         wrq->remote_addr        = kiblnd_rd_frag_addr(dstrd,
1198                                                                       dstidx);
1199                         wrq->rkey               = kiblnd_rd_frag_key(dstrd,
1200                                                                      dstidx);
1201 #else
1202                         wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1203                                                                         dstidx);
1204                         wrq->wr.wr.rdma.rkey    = kiblnd_rd_frag_key(dstrd,
1205                                                                      dstidx);
1206 #endif
1207                 }
1208
1209                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1210                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1211
1212                 wrq_sge++;
1213                 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1214                         tx->tx_nwrq++;
1215                         wrq->wr.num_sge = wrq_sge;
1216                         wrq_sge = 0;
1217                 }
1218                 tx->tx_nsge++;
1219         }
1220
1221         if (rc < 0)     /* no RDMA if completing with failure */
1222                 tx->tx_nwrq = tx->tx_nsge = 0;
1223
1224         ibmsg->ibm_u.completion.ibcm_status = rc;
1225         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1226         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1227                            type, sizeof(struct kib_completion_msg));
1228
1229         return rc;
1230 }
1231
1232 static void
1233 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1234 {
1235         struct list_head *q;
1236         s64 timeout_ns;
1237
1238         LASSERT(tx->tx_nwrq > 0);       /* work items set up */
1239         LASSERT(!tx->tx_queued);        /* not queued for sending already */
1240         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1241
1242         if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1243                 tx->tx_status = -ECONNABORTED;
1244                 tx->tx_waiting = 0;
1245                 if (tx->tx_conn != NULL) {
1246                         /* PUT_DONE first attached to conn as a PUT_REQ */
1247                         LASSERT(tx->tx_conn == conn);
1248                         LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1249                         tx->tx_conn = NULL;
1250                         kiblnd_conn_decref(conn);
1251                 }
1252                 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1253
1254                 return;
1255         }
1256
1257         timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1258         tx->tx_queued = 1;
1259         tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1260
1261         if (tx->tx_conn == NULL) {
1262                 kiblnd_conn_addref(conn);
1263                 tx->tx_conn = conn;
1264                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1265         } else {
1266                 /* PUT_DONE first attached to conn as a PUT_REQ */
1267                 LASSERT (tx->tx_conn == conn);
1268                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1269         }
1270
1271         switch (tx->tx_msg->ibm_type) {
1272         default:
1273                 LBUG();
1274
1275         case IBLND_MSG_PUT_REQ:
1276         case IBLND_MSG_GET_REQ:
1277                 q = &conn->ibc_tx_queue_rsrvd;
1278                 break;
1279
1280         case IBLND_MSG_PUT_NAK:
1281         case IBLND_MSG_PUT_ACK:
1282         case IBLND_MSG_PUT_DONE:
1283         case IBLND_MSG_GET_DONE:
1284                 q = &conn->ibc_tx_queue_nocred;
1285                 break;
1286
1287         case IBLND_MSG_NOOP:
1288                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1289                         q = &conn->ibc_tx_queue_nocred;
1290                 else
1291                         q = &conn->ibc_tx_noops;
1292                 break;
1293
1294         case IBLND_MSG_IMMEDIATE:
1295                 q = &conn->ibc_tx_queue;
1296                 break;
1297         }
1298
1299         list_add_tail(&tx->tx_list, q);
1300 }
1301
1302 static void
1303 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1304 {
1305         spin_lock(&conn->ibc_lock);
1306         kiblnd_queue_tx_locked(tx, conn);
1307         kiblnd_check_sends_locked(conn);
1308         spin_unlock(&conn->ibc_lock);
1309 }
1310
1311 static int
1312 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1313                         struct sockaddr_in *srcaddr,
1314                         struct sockaddr_in *dstaddr,
1315                         int timeout_ms)
1316 {
1317         unsigned short port;
1318         int rc;
1319
1320         /* allow the port to be reused */
1321         rc = rdma_set_reuseaddr(cmid, 1);
1322         if (rc != 0) {
1323                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1324                 return rc;
1325         }
1326
1327         /* look for a free privileged port */
1328         for (port = PROT_SOCK-1; port > 0; port--) {
1329                 srcaddr->sin_port = htons(port);
1330                 rc = rdma_resolve_addr(cmid,
1331                                        (struct sockaddr *)srcaddr,
1332                                        (struct sockaddr *)dstaddr,
1333                                        timeout_ms);
1334                 if (rc == 0) {
1335                         CDEBUG(D_NET, "bound to port %hu\n", port);
1336                         return 0;
1337                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1338                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1339                                port, rc);
1340                 } else {
1341                         return rc;
1342                 }
1343         }
1344
1345         CERROR("cannot bind to a free privileged port: rc = %d\n", rc);
1346
1347         return rc;
1348 }
1349
1350 static int
1351 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1352                     struct sockaddr_in *srcaddr,
1353                     struct sockaddr_in *dstaddr,
1354                     int timeout_ms)
1355 {
1356         const struct cred *old_creds = NULL;
1357         struct cred *new_creds;
1358         int rc;
1359
1360         if (!capable(CAP_NET_BIND_SERVICE)) {
1361                 new_creds = prepare_kernel_cred(NULL);
1362                 if (!new_creds)
1363                         return -ENOMEM;
1364
1365                 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1366                 old_creds = override_creds(new_creds);
1367         }
1368
1369         rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1370
1371         if (old_creds)
1372                 revert_creds(old_creds);
1373
1374         return rc;
1375 }
1376
1377 static void
1378 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1379 {
1380         struct rdma_cm_id *cmid;
1381         struct kib_dev *dev;
1382         struct kib_net *net = peer_ni->ibp_ni->ni_data;
1383         struct sockaddr_in srcaddr;
1384         struct sockaddr_in dstaddr;
1385         int rc;
1386
1387         LASSERT (net != NULL);
1388         LASSERT (peer_ni->ibp_connecting > 0);
1389
1390         cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1391                                      kiblnd_cm_callback, peer_ni,
1392                                      RDMA_PS_TCP, IB_QPT_RC);
1393
1394         if (IS_ERR(cmid)) {
1395                 CERROR("Can't create CMID for %s: %ld\n",
1396                        libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1397                 rc = PTR_ERR(cmid);
1398                 goto failed;
1399         }
1400
1401         dev = net->ibn_dev;
1402         memset(&srcaddr, 0, sizeof(srcaddr));
1403         srcaddr.sin_family = AF_INET;
1404         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1405
1406         memset(&dstaddr, 0, sizeof(dstaddr));
1407         dstaddr.sin_family = AF_INET;
1408         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1409         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1410
1411         kiblnd_peer_addref(peer_ni);               /* cmid's ref */
1412
1413         if (*kiblnd_tunables.kib_use_priv_port) {
1414                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1415                                          kiblnd_timeout() * 1000);
1416         } else {
1417                 rc = rdma_resolve_addr(cmid,
1418                                        (struct sockaddr *)&srcaddr,
1419                                        (struct sockaddr *)&dstaddr,
1420                                        kiblnd_timeout() * 1000);
1421         }
1422         if (rc != 0) {
1423                 /* Can't initiate address resolution:  */
1424                 CERROR("Can't resolve addr for %s: %d\n",
1425                        libcfs_nid2str(peer_ni->ibp_nid), rc);
1426                 goto failed2;
1427         }
1428
1429         return;
1430
1431  failed2:
1432         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1433         kiblnd_peer_decref(peer_ni);               /* cmid's ref */
1434         rdma_destroy_id(cmid);
1435         return;
1436  failed:
1437         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1438 }
1439
1440 bool
1441 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1442 {
1443         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1444         char *reason = NULL;
1445         LIST_HEAD(txs);
1446         unsigned long flags;
1447
1448         write_lock_irqsave(glock, flags);
1449         if (peer_ni->ibp_reconnecting == 0) {
1450                 if (peer_ni->ibp_accepting)
1451                         reason = "accepting";
1452                 else if (peer_ni->ibp_connecting)
1453                         reason = "connecting";
1454                 else if (!list_empty(&peer_ni->ibp_conns))
1455                         reason = "connected";
1456                 else /* connected then closed */
1457                         reason = "closed";
1458
1459                 goto no_reconnect;
1460         }
1461
1462         if (peer_ni->ibp_accepting)
1463                 CNETERR("Detecting race between accepting and reconnecting\n");
1464         peer_ni->ibp_reconnecting--;
1465
1466         if (!kiblnd_peer_active(peer_ni)) {
1467                 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1468                 reason = "unlinked";
1469                 goto no_reconnect;
1470         }
1471
1472         peer_ni->ibp_connecting++;
1473         peer_ni->ibp_reconnected++;
1474
1475         write_unlock_irqrestore(glock, flags);
1476
1477         kiblnd_connect_peer(peer_ni);
1478         return true;
1479
1480  no_reconnect:
1481         write_unlock_irqrestore(glock, flags);
1482
1483         CWARN("Abort reconnection of %s: %s\n",
1484               libcfs_nid2str(peer_ni->ibp_nid), reason);
1485         kiblnd_txlist_done(&txs, -ECONNABORTED,
1486                            LNET_MSG_STATUS_LOCAL_ABORTED);
1487         return false;
1488 }
1489
1490 void
1491 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1492 {
1493         struct kib_peer_ni *peer_ni;
1494         struct kib_peer_ni *peer2;
1495         struct kib_conn *conn;
1496         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1497         unsigned long flags;
1498         int rc;
1499         int i;
1500         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1501
1502         /* If I get here, I've committed to send, so I complete the tx with
1503          * failure on any problems
1504          */
1505
1506         LASSERT(!tx || !tx->tx_conn);     /* only set when assigned a conn */
1507         LASSERT(!tx || tx->tx_nwrq > 0);  /* work items have been set up */
1508
1509         /* First time, just use a read lock since I expect to find my peer_ni
1510          * connected
1511          */
1512         read_lock_irqsave(g_lock, flags);
1513
1514         peer_ni = kiblnd_find_peer_locked(ni, nid);
1515         if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1516                 /* Found a peer_ni with an established connection */
1517                 conn = kiblnd_get_conn_locked(peer_ni);
1518                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1519
1520                 read_unlock_irqrestore(g_lock, flags);
1521
1522                 if (tx != NULL)
1523                         kiblnd_queue_tx(tx, conn);
1524                 kiblnd_conn_decref(conn); /* ...to here */
1525                 return;
1526         }
1527
1528         read_unlock(g_lock);
1529         /* Re-try with a write lock */
1530         write_lock(g_lock);
1531
1532         peer_ni = kiblnd_find_peer_locked(ni, nid);
1533         if (peer_ni != NULL) {
1534                 if (list_empty(&peer_ni->ibp_conns)) {
1535                         /* found a peer_ni, but it's still connecting... */
1536                         LASSERT(kiblnd_peer_connecting(peer_ni));
1537                         if (tx != NULL)
1538                                 list_add_tail(&tx->tx_list,
1539                                               &peer_ni->ibp_tx_queue);
1540                         write_unlock_irqrestore(g_lock, flags);
1541                 } else {
1542                         conn = kiblnd_get_conn_locked(peer_ni);
1543                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1544
1545                         write_unlock_irqrestore(g_lock, flags);
1546
1547                         if (tx != NULL)
1548                                 kiblnd_queue_tx(tx, conn);
1549                         kiblnd_conn_decref(conn); /* ...to here */
1550                 }
1551                 return;
1552         }
1553
1554         write_unlock_irqrestore(g_lock, flags);
1555
1556         /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1557         rc = kiblnd_create_peer(ni, &peer_ni, nid);
1558         if (rc != 0) {
1559                 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1560                 if (tx != NULL) {
1561                         tx->tx_status = -EHOSTUNREACH;
1562                         tx->tx_waiting = 0;
1563                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1564                         kiblnd_tx_done(tx);
1565                 }
1566                 return;
1567         }
1568
1569         write_lock_irqsave(g_lock, flags);
1570
1571         peer2 = kiblnd_find_peer_locked(ni, nid);
1572         if (peer2 != NULL) {
1573                 if (list_empty(&peer2->ibp_conns)) {
1574                         /* found a peer_ni, but it's still connecting... */
1575                         LASSERT(kiblnd_peer_connecting(peer2));
1576                         if (tx != NULL)
1577                                 list_add_tail(&tx->tx_list,
1578                                               &peer2->ibp_tx_queue);
1579                         write_unlock_irqrestore(g_lock, flags);
1580                 } else {
1581                         conn = kiblnd_get_conn_locked(peer2);
1582                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1583
1584                         write_unlock_irqrestore(g_lock, flags);
1585
1586                         if (tx != NULL)
1587                                 kiblnd_queue_tx(tx, conn);
1588                         kiblnd_conn_decref(conn); /* ...to here */
1589                 }
1590
1591                 kiblnd_peer_decref(peer_ni);
1592                 return;
1593         }
1594
1595         /* Brand new peer_ni */
1596         LASSERT(peer_ni->ibp_connecting == 0);
1597         tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1598         peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1599
1600         /* always called with a ref on ni, which prevents ni being shutdown */
1601         LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1602
1603         if (tx != NULL)
1604                 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1605
1606         kiblnd_peer_addref(peer_ni);
1607         hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
1608
1609         write_unlock_irqrestore(g_lock, flags);
1610
1611         for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1612                 kiblnd_connect_peer(peer_ni);
1613         kiblnd_peer_decref(peer_ni);
1614 }
1615
1616 int
1617 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1618 {
1619         struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1620         int               type = lntmsg->msg_type;
1621         struct lnet_process_id target = lntmsg->msg_target;
1622         int               target_is_router = lntmsg->msg_target_is_router;
1623         int               routing = lntmsg->msg_routing;
1624         unsigned int      payload_niov = lntmsg->msg_niov;
1625         struct bio_vec   *payload_kiov = lntmsg->msg_kiov;
1626         unsigned int      payload_offset = lntmsg->msg_offset;
1627         unsigned int      payload_nob = lntmsg->msg_len;
1628         struct kib_msg *ibmsg;
1629         struct kib_rdma_desc *rd;
1630         struct kib_tx *tx;
1631         int               nob;
1632         int               rc;
1633
1634         /* NB 'private' is different depending on what we're sending.... */
1635
1636         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1637                payload_nob, payload_niov, libcfs_id2str(target));
1638
1639         LASSERT (payload_nob == 0 || payload_niov > 0);
1640         LASSERT (payload_niov <= LNET_MAX_IOV);
1641
1642         /* Thread context */
1643         LASSERT (!in_interrupt());
1644
1645         switch (type) {
1646         default:
1647                 LBUG();
1648                 return (-EIO);
1649
1650         case LNET_MSG_ACK:
1651                 LASSERT (payload_nob == 0);
1652                 break;
1653
1654         case LNET_MSG_GET:
1655                 if (routing || target_is_router)
1656                         break;                  /* send IMMEDIATE */
1657
1658                 /* is the REPLY message too small for RDMA? */
1659                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1660                 if (nob <= IBLND_MSG_SIZE)
1661                         break;                  /* send IMMEDIATE */
1662
1663                 tx = kiblnd_get_idle_tx(ni, target.nid);
1664                 if (tx == NULL) {
1665                         CERROR("Can't allocate txd for GET to %s\n",
1666                                libcfs_nid2str(target.nid));
1667                         return -ENOMEM;
1668                 }
1669
1670                 ibmsg = tx->tx_msg;
1671                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1672                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1673                                           lntmsg->msg_md->md_niov,
1674                                           lntmsg->msg_md->md_kiov,
1675                                           0, lntmsg->msg_md->md_length);
1676                 if (rc != 0) {
1677                         CERROR("Can't setup GET sink for %s: %d\n",
1678                                libcfs_nid2str(target.nid), rc);
1679                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1680                         kiblnd_tx_done(tx);
1681                         return -EIO;
1682                 }
1683
1684                 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1685                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1686                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1687
1688                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1689
1690                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1691                 if (tx->tx_lntmsg[1] == NULL) {
1692                         CERROR("Can't create reply for GET -> %s\n",
1693                                libcfs_nid2str(target.nid));
1694                         kiblnd_tx_done(tx);
1695                         return -EIO;
1696                 }
1697
1698                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1699                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1700                 kiblnd_launch_tx(ni, tx, target.nid);
1701                 return 0;
1702
1703         case LNET_MSG_REPLY:
1704         case LNET_MSG_PUT:
1705                 /* Is the payload small enough not to need RDMA? */
1706                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1707                 if (nob <= IBLND_MSG_SIZE)
1708                         break;                  /* send IMMEDIATE */
1709
1710                 tx = kiblnd_get_idle_tx(ni, target.nid);
1711                 if (tx == NULL) {
1712                         CERROR("Can't allocate %s txd for %s\n",
1713                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1714                                libcfs_nid2str(target.nid));
1715                         return -ENOMEM;
1716                 }
1717
1718                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1719                                           payload_niov, payload_kiov,
1720                                           payload_offset, payload_nob);
1721                 if (rc != 0) {
1722                         CERROR("Can't setup PUT src for %s: %d\n",
1723                                libcfs_nid2str(target.nid), rc);
1724                         kiblnd_tx_done(tx);
1725                         return -EIO;
1726                 }
1727
1728                 ibmsg = tx->tx_msg;
1729                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1730                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1731                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1732                                    sizeof(struct kib_putreq_msg));
1733
1734                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1735                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1736                 kiblnd_launch_tx(ni, tx, target.nid);
1737                 return 0;
1738         }
1739
1740         /* send IMMEDIATE */
1741         LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1742                 <= IBLND_MSG_SIZE);
1743
1744         tx = kiblnd_get_idle_tx(ni, target.nid);
1745         if (tx == NULL) {
1746                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1747                         type, libcfs_nid2str(target.nid));
1748                 return -ENOMEM;
1749         }
1750
1751         ibmsg = tx->tx_msg;
1752         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1753
1754         lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1755                             offsetof(struct kib_msg,
1756                                      ibm_u.immediate.ibim_payload),
1757                             payload_niov, payload_kiov,
1758                             payload_offset, payload_nob);
1759
1760         nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1761         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1762
1763         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1764         kiblnd_launch_tx(ni, tx, target.nid);
1765         return 0;
1766 }
1767
1768 static void
1769 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1770 {
1771         struct lnet_process_id target = lntmsg->msg_target;
1772         unsigned int niov = lntmsg->msg_niov;
1773         struct bio_vec *kiov = lntmsg->msg_kiov;
1774         unsigned int offset = lntmsg->msg_offset;
1775         unsigned int nob = lntmsg->msg_len;
1776         struct kib_tx *tx;
1777         int rc;
1778
1779         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1780         if (tx == NULL) {
1781                 CERROR("Can't get tx for REPLY to %s\n",
1782                        libcfs_nid2str(target.nid));
1783                 goto failed_0;
1784         }
1785
1786         if (nob == 0)
1787                 rc = 0;
1788         else
1789                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1790                                           niov, kiov, offset, nob);
1791
1792         if (rc != 0) {
1793                 CERROR("Can't setup GET src for %s: %d\n",
1794                        libcfs_nid2str(target.nid), rc);
1795                 goto failed_1;
1796         }
1797
1798         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1799                               IBLND_MSG_GET_DONE, nob,
1800                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1801                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1802         if (rc < 0) {
1803                 CERROR("Can't setup rdma for GET from %s: %d\n",
1804                        libcfs_nid2str(target.nid), rc);
1805                 goto failed_1;
1806         }
1807
1808         if (nob == 0) {
1809                 /* No RDMA: local completion may happen now! */
1810                 lnet_finalize(lntmsg, 0);
1811         } else {
1812                 /* RDMA: lnet_finalize(lntmsg) when it
1813                  * completes */
1814                 tx->tx_lntmsg[0] = lntmsg;
1815         }
1816
1817         kiblnd_queue_tx(tx, rx->rx_conn);
1818         return;
1819
1820
1821 failed_1:
1822         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1823         kiblnd_tx_done(tx);
1824 failed_0:
1825         lnet_finalize(lntmsg, -EIO);
1826 }
1827
1828 int
1829 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1830             int delayed, unsigned int niov, struct bio_vec *kiov,
1831             unsigned int offset, unsigned int mlen, unsigned int rlen)
1832 {
1833         struct kib_rx *rx = private;
1834         struct kib_msg *rxmsg = rx->rx_msg;
1835         struct kib_conn *conn = rx->rx_conn;
1836         struct kib_tx *tx;
1837         __u64        ibprm_cookie;
1838         int          nob;
1839         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1840         int          rc = 0;
1841
1842         LASSERT (mlen <= rlen);
1843         LASSERT (!in_interrupt());
1844
1845         switch (rxmsg->ibm_type) {
1846         default:
1847                 LBUG();
1848
1849         case IBLND_MSG_IMMEDIATE:
1850                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1851                 if (nob > rx->rx_nob) {
1852                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1853                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1854                                 nob, rx->rx_nob);
1855                         rc = -EPROTO;
1856                         break;
1857                 }
1858
1859                 lnet_copy_flat2kiov(niov, kiov, offset,
1860                                     IBLND_MSG_SIZE, rxmsg,
1861                                     offsetof(struct kib_msg,
1862                                              ibm_u.immediate.ibim_payload),
1863                                     mlen);
1864                 lnet_finalize(lntmsg, 0);
1865                 break;
1866
1867         case IBLND_MSG_PUT_REQ: {
1868                 struct kib_msg  *txmsg;
1869                 struct kib_rdma_desc *rd;
1870                 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1871
1872                 if (mlen == 0) {
1873                         lnet_finalize(lntmsg, 0);
1874                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1875                                                0, ibprm_cookie);
1876                         break;
1877                 }
1878
1879                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1880                 if (tx == NULL) {
1881                         CERROR("Can't allocate tx for %s\n",
1882                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1883                         /* Not replying will break the connection */
1884                         rc = -ENOMEM;
1885                         break;
1886                 }
1887
1888                 txmsg = tx->tx_msg;
1889                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1890                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1891                                           niov, kiov, offset, mlen);
1892                 if (rc != 0) {
1893                         CERROR("Can't setup PUT sink for %s: %d\n",
1894                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1895                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1896                         kiblnd_tx_done(tx);
1897                         /* tell peer_ni it's over */
1898                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1899                                                rc, ibprm_cookie);
1900                         break;
1901                 }
1902
1903                 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1904                 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1905                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1906
1907                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1908
1909                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1910                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1911                 kiblnd_queue_tx(tx, conn);
1912
1913                 /* reposted buffer reserved for PUT_DONE */
1914                 post_credit = IBLND_POSTRX_NO_CREDIT;
1915                 break;
1916                 }
1917
1918         case IBLND_MSG_GET_REQ:
1919                 if (lntmsg != NULL) {
1920                         /* Optimized GET; RDMA lntmsg's payload */
1921                         kiblnd_reply(ni, rx, lntmsg);
1922                 } else {
1923                         /* GET didn't match anything */
1924                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1925                                                -ENODATA,
1926                                                rxmsg->ibm_u.get.ibgm_cookie);
1927                 }
1928                 break;
1929         }
1930
1931         kiblnd_post_rx(rx, post_credit);
1932         return rc;
1933 }
1934
1935 int
1936 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1937 {
1938         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1939
1940         if (IS_ERR(task))
1941                 return PTR_ERR(task);
1942
1943         atomic_inc(&kiblnd_data.kib_nthreads);
1944         return 0;
1945 }
1946
1947 static void
1948 kiblnd_thread_fini (void)
1949 {
1950         atomic_dec (&kiblnd_data.kib_nthreads);
1951 }
1952
1953 static void
1954 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1955 {
1956         /* This is racy, but everyone's only writing ktime_get_seconds() */
1957         peer_ni->ibp_last_alive = ktime_get_seconds();
1958         smp_mb();
1959 }
1960
1961 static void
1962 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1963 {
1964         int           error = 0;
1965         time64_t last_alive = 0;
1966         unsigned long flags;
1967
1968         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1969
1970         if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1971                 error = peer_ni->ibp_error;
1972                 peer_ni->ibp_error = 0;
1973
1974                 last_alive = peer_ni->ibp_last_alive;
1975         }
1976
1977         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1978
1979         if (error != 0)
1980                 lnet_notify(peer_ni->ibp_ni,
1981                             peer_ni->ibp_nid, false, false, last_alive);
1982 }
1983
1984 void
1985 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1986 {
1987         /* This just does the immediate housekeeping.  'error' is zero for a
1988          * normal shutdown which can happen only after the connection has been
1989          * established.  If the connection is established, schedule the
1990          * connection to be finished off by the connd.  Otherwise the connd is
1991          * already dealing with it (either to set it up or tear it down).
1992          * Caller holds kib_global_lock exclusively in irq context */
1993         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1994         struct kib_dev *dev;
1995         unsigned long flags;
1996
1997         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1998
1999         if (error != 0 && conn->ibc_comms_error == 0)
2000                 conn->ibc_comms_error = error;
2001
2002         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
2003                 return; /* already being handled  */
2004
2005         if (error == 0 &&
2006             list_empty(&conn->ibc_tx_noops) &&
2007             list_empty(&conn->ibc_tx_queue) &&
2008             list_empty(&conn->ibc_tx_queue_rsrvd) &&
2009             list_empty(&conn->ibc_tx_queue_nocred) &&
2010             list_empty(&conn->ibc_active_txs)) {
2011                 CDEBUG(D_NET, "closing conn to %s\n", 
2012                        libcfs_nid2str(peer_ni->ibp_nid));
2013         } else {
2014                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2015                        libcfs_nid2str(peer_ni->ibp_nid), error,
2016                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2017                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2018                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
2019                                                 "" : "(sending_rsrvd)",
2020                        list_empty(&conn->ibc_tx_queue_nocred) ?
2021                                                  "" : "(sending_nocred)",
2022                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2023         }
2024
2025         dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2026         if (peer_ni->ibp_next_conn == conn)
2027                 /* clear next_conn so it won't be used */
2028                 peer_ni->ibp_next_conn = NULL;
2029         list_del(&conn->ibc_list);
2030         /* connd (see below) takes over ibc_list's ref */
2031
2032         if (list_empty(&peer_ni->ibp_conns) &&    /* no more conns */
2033             kiblnd_peer_active(peer_ni)) {         /* still in peer_ni table */
2034                 kiblnd_unlink_peer_locked(peer_ni);
2035
2036                 /* set/clear error on last conn */
2037                 peer_ni->ibp_error = conn->ibc_comms_error;
2038         }
2039
2040         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2041
2042         if (error != 0 &&
2043             kiblnd_dev_can_failover(dev)) {
2044                 list_add_tail(&dev->ibd_fail_list,
2045                               &kiblnd_data.kib_failed_devs);
2046                 wake_up(&kiblnd_data.kib_failover_waitq);
2047         }
2048
2049         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2050
2051         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2052         wake_up(&kiblnd_data.kib_connd_waitq);
2053
2054         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2055 }
2056
2057 void
2058 kiblnd_close_conn(struct kib_conn *conn, int error)
2059 {
2060         unsigned long flags;
2061
2062         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2063
2064         kiblnd_close_conn_locked(conn, error);
2065
2066         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2067 }
2068
2069 static void
2070 kiblnd_handle_early_rxs(struct kib_conn *conn)
2071 {
2072         unsigned long flags;
2073         struct kib_rx *rx;
2074
2075         LASSERT(!in_interrupt());
2076         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2077
2078         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2079         while (!list_empty(&conn->ibc_early_rxs)) {
2080                 rx = list_entry(conn->ibc_early_rxs.next,
2081                                 struct kib_rx, rx_list);
2082                 list_del(&rx->rx_list);
2083                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2084
2085                 kiblnd_handle_rx(rx);
2086
2087                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2088         }
2089         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2090 }
2091
2092 void
2093 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2094 {
2095         LIST_HEAD(zombies);
2096         struct kib_tx *nxt;
2097         struct kib_tx *tx;
2098
2099         spin_lock(&conn->ibc_lock);
2100
2101         list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2102                 if (txs == &conn->ibc_active_txs) {
2103                         LASSERT(!tx->tx_queued);
2104                         LASSERT(tx->tx_waiting ||
2105                                 tx->tx_sending != 0);
2106                         if (conn->ibc_comms_error == -ETIMEDOUT) {
2107                                 if (tx->tx_waiting && !tx->tx_sending)
2108                                         tx->tx_hstatus =
2109                                           LNET_MSG_STATUS_REMOTE_TIMEOUT;
2110                                 else if (tx->tx_sending)
2111                                         tx->tx_hstatus =
2112                                           LNET_MSG_STATUS_NETWORK_TIMEOUT;
2113                         }
2114                 } else {
2115                         LASSERT(tx->tx_queued);
2116                         if (conn->ibc_comms_error == -ETIMEDOUT)
2117                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2118                         else
2119                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2120                 }
2121
2122                 tx->tx_status = -ECONNABORTED;
2123                 tx->tx_waiting = 0;
2124
2125                 /*
2126                  * TODO: This makes an assumption that
2127                  * kiblnd_tx_complete() will be called for each tx. If
2128                  * that event is dropped we could end up with stale
2129                  * connections floating around. We'd like to deal with
2130                  * that in a better way.
2131                  *
2132                  * Also that means we can exceed the timeout by many
2133                  * seconds.
2134                  */
2135                 if (tx->tx_sending == 0) {
2136                         tx->tx_queued = 0;
2137                         list_move(&tx->tx_list, &zombies);
2138                 } else {
2139                         /* keep tx until cq destroy */
2140                         list_move(&tx->tx_list, &conn->ibc_zombie_txs);
2141                         conn->ibc_waits ++;
2142                 }
2143         }
2144
2145         spin_unlock(&conn->ibc_lock);
2146
2147         /*
2148          * aborting transmits occurs when finalizing the connection.
2149          * The connection is finalized on error.
2150          * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2151          * override the value already set in tx->tx_hstatus above.
2152          */
2153         kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2154 }
2155
2156 static bool
2157 kiblnd_tx_may_discard(struct kib_conn *conn)
2158 {
2159         bool rc = false;
2160         struct kib_tx *nxt;
2161         struct kib_tx *tx;
2162
2163         spin_lock(&conn->ibc_lock);
2164
2165         list_for_each_entry_safe(tx, nxt, &conn->ibc_zombie_txs, tx_list) {
2166                 if (tx->tx_sending > 0 && tx->tx_lntmsg[0] &&
2167                     lnet_md_discarded(tx->tx_lntmsg[0]->msg_md)) {
2168                         tx->tx_sending --;
2169                         if (tx->tx_sending == 0) {
2170                                 kiblnd_conn_decref(tx->tx_conn);
2171                                 tx->tx_conn = NULL;
2172                                 rc = true;
2173                         }
2174                 }
2175         }
2176
2177         spin_unlock(&conn->ibc_lock);
2178         return rc;
2179 }
2180
2181 static void
2182 kiblnd_finalise_conn(struct kib_conn *conn)
2183 {
2184         LASSERT (!in_interrupt());
2185         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2186
2187         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
2188          * for connections that didn't get as far as being connected, because
2189          * rdma_disconnect() does this for free. */
2190         kiblnd_abort_receives(conn);
2191
2192         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2193
2194         /* Complete all tx descs not waiting for sends to complete.
2195          * NB we should be safe from RDMA now that the QP has changed state */
2196
2197         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2198         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2199         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2200         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2201         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2202
2203         kiblnd_handle_early_rxs(conn);
2204 }
2205
2206 static void
2207 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2208                            int error)
2209 {
2210         LIST_HEAD(zombies);
2211         unsigned long flags;
2212         enum lnet_msg_hstatus hstatus;
2213
2214         LASSERT(error != 0);
2215         LASSERT(!in_interrupt());
2216
2217         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2218
2219         if (active) {
2220                 LASSERT(peer_ni->ibp_connecting > 0);
2221                 peer_ni->ibp_connecting--;
2222         } else {
2223                 LASSERT (peer_ni->ibp_accepting > 0);
2224                 peer_ni->ibp_accepting--;
2225         }
2226
2227         if (kiblnd_peer_connecting(peer_ni)) {
2228                 /* another connection attempt under way... */
2229                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2230                                         flags);
2231                 return;
2232         }
2233
2234         peer_ni->ibp_reconnected = 0;
2235         if (list_empty(&peer_ni->ibp_conns)) {
2236                 /* Take peer_ni's blocked transmits to complete with error */
2237                 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2238
2239                 if (kiblnd_peer_active(peer_ni))
2240                         kiblnd_unlink_peer_locked(peer_ni);
2241
2242                 peer_ni->ibp_error = error;
2243         } else {
2244                 /* Can't have blocked transmits if there are connections */
2245                 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2246         }
2247
2248         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2249
2250         kiblnd_peer_notify(peer_ni);
2251
2252         if (list_empty(&zombies))
2253                 return;
2254
2255         CNETERR("Deleting messages for %s: connection failed\n",
2256                 libcfs_nid2str(peer_ni->ibp_nid));
2257
2258         switch (error) {
2259         case -EHOSTUNREACH:
2260         case -ETIMEDOUT:
2261                 hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
2262                 break;
2263         case -ECONNREFUSED:
2264                 hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
2265                 break;
2266         default:
2267                 hstatus = LNET_MSG_STATUS_LOCAL_DROPPED;
2268                 break;
2269         }
2270
2271         kiblnd_txlist_done(&zombies, error, hstatus);
2272 }
2273
2274 static void
2275 kiblnd_connreq_done(struct kib_conn *conn, int status)
2276 {
2277         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2278         struct kib_tx *tx;
2279         LIST_HEAD(txs);
2280         unsigned long    flags;
2281         int              active;
2282
2283         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2284
2285         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2286                libcfs_nid2str(peer_ni->ibp_nid), active,
2287                conn->ibc_version, status);
2288
2289         LASSERT (!in_interrupt());
2290         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2291                   peer_ni->ibp_connecting > 0) ||
2292                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2293                   peer_ni->ibp_accepting > 0));
2294
2295         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2296         conn->ibc_connvars = NULL;
2297
2298         if (status != 0) {
2299                 /* failed to establish connection */
2300                 kiblnd_peer_connect_failed(peer_ni, active, status);
2301                 kiblnd_finalise_conn(conn);
2302                 return;
2303         }
2304
2305         /* connection established */
2306         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2307
2308         /* reset retry count */
2309         peer_ni->ibp_retries = 0;
2310
2311         conn->ibc_last_send = ktime_get();
2312         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2313         kiblnd_peer_alive(peer_ni);
2314
2315         /* Add conn to peer_ni's list and nuke any dangling conns from a different
2316          * peer_ni instance... */
2317         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2318         list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2319         peer_ni->ibp_reconnected = 0;
2320         if (active)
2321                 peer_ni->ibp_connecting--;
2322         else
2323                 peer_ni->ibp_accepting--;
2324
2325         if (peer_ni->ibp_version == 0) {
2326                 peer_ni->ibp_version     = conn->ibc_version;
2327                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2328         }
2329
2330         if (peer_ni->ibp_version     != conn->ibc_version ||
2331             peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2332                 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2333                                                 conn->ibc_incarnation);
2334                 peer_ni->ibp_version     = conn->ibc_version;
2335                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2336         }
2337
2338         /* grab pending txs while I have the lock */
2339         list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2340
2341         if (!kiblnd_peer_active(peer_ni) ||        /* peer_ni has been deleted */
2342             conn->ibc_comms_error != 0) {       /* error has happened already */
2343
2344                 /* start to shut down connection */
2345                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2346                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2347
2348                 kiblnd_txlist_done(&txs, -ECONNABORTED,
2349                                    LNET_MSG_STATUS_LOCAL_ERROR);
2350
2351                 return;
2352         }
2353
2354         /* +1 ref for myself, this connection is visible to other threads
2355          * now, refcount of peer:ibp_conns can be released by connection
2356          * close from either a different thread, or the calling of
2357          * kiblnd_check_sends_locked() below. See bz21911 for details.
2358          */
2359         kiblnd_conn_addref(conn);
2360         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2361
2362         /* Schedule blocked txs
2363          * Note: if we are running with conns_per_peer > 1, these blocked
2364          * txs will all get scheduled to the first connection which gets
2365          * scheduled.  We won't be using round robin on this first batch.
2366          */
2367         spin_lock(&conn->ibc_lock);
2368         while (!list_empty(&txs)) {
2369                 tx = list_entry(txs.next, struct kib_tx, tx_list);
2370                 list_del(&tx->tx_list);
2371
2372                 kiblnd_queue_tx_locked(tx, conn);
2373         }
2374         kiblnd_check_sends_locked(conn);
2375         spin_unlock(&conn->ibc_lock);
2376
2377         /* schedule blocked rxs */
2378         kiblnd_handle_early_rxs(conn);
2379         kiblnd_conn_decref(conn);
2380 }
2381
2382 static void
2383 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2384 {
2385         int          rc;
2386
2387 #ifdef HAVE_RDMA_REJECT_4ARGS
2388         rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2389 #else
2390         rc = rdma_reject(cmid, rej, sizeof(*rej));
2391 #endif
2392
2393         if (rc != 0)
2394                 CWARN("Error %d sending reject\n", rc);
2395 }
2396
2397 static int
2398 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2399 {
2400         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2401         struct kib_msg *reqmsg = priv;
2402         struct kib_msg *ackmsg;
2403         struct kib_dev *ibdev;
2404         struct kib_peer_ni *peer_ni;
2405         struct kib_peer_ni *peer2;
2406         struct kib_conn *conn;
2407         struct lnet_ni *ni = NULL;
2408         struct kib_net *net = NULL;
2409         lnet_nid_t nid;
2410         struct rdma_conn_param cp;
2411         struct kib_rej rej;
2412         int version = IBLND_MSG_VERSION;
2413         unsigned long flags;
2414         int rc;
2415         struct sockaddr_in *peer_addr;
2416
2417         LASSERT(!in_interrupt());
2418         /* cmid inherits 'context' from the corresponding listener id */
2419         ibdev = cmid->context;
2420         LASSERT(ibdev);
2421
2422         memset(&rej, 0, sizeof(rej));
2423         rej.ibr_magic                = IBLND_MSG_MAGIC;
2424         rej.ibr_why                  = IBLND_REJECT_FATAL;
2425         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2426
2427         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2428         if (*kiblnd_tunables.kib_require_priv_port &&
2429             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2430                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2431                 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2432                        &ip, ntohs(peer_addr->sin_port));
2433                 goto failed;
2434         }
2435
2436         if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2437                 CERROR("Short connection request\n");
2438                 goto failed;
2439         }
2440
2441         /* Future protocol version compatibility support!  If the
2442          * o2iblnd-specific protocol changes, or when LNET unifies
2443          * protocols over all LNDs, the initial connection will
2444          * negotiate a protocol version.  I trap this here to avoid
2445          * console errors; the reject tells the peer_ni which protocol I
2446          * speak. */
2447         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2448             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2449                 goto failed;
2450         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2451             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2452             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2453                 goto failed;
2454         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2455             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2456             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2457                 goto failed;
2458
2459         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2460         if (rc != 0) {
2461                 CERROR("Can't parse connection request: %d\n", rc);
2462                 goto failed;
2463         }
2464
2465         nid = reqmsg->ibm_srcnid;
2466         ni  = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2467
2468         if (ni != NULL) {
2469                 net = (struct kib_net *)ni->ni_data;
2470                 rej.ibr_incarnation = net->ibn_incarnation;
2471         }
2472
2473         if (ni == NULL ||                         /* no matching net */
2474             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2475             net->ibn_dev != ibdev) {              /* wrong device */
2476                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n", libcfs_nid2str(nid),
2477                        ni ? libcfs_nid2str(ni->ni_nid) : "NA",
2478                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2479                        &ibdev->ibd_ifip,
2480                        libcfs_nid2str(reqmsg->ibm_dstnid));
2481
2482                 goto failed;
2483         }
2484
2485         /* check time stamp as soon as possible */
2486         if (reqmsg->ibm_dststamp != 0 &&
2487             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2488                 CWARN("Stale connection request\n");
2489                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2490                 goto failed;
2491         }
2492
2493         /* I can accept peer_ni's version */
2494         version = reqmsg->ibm_version;
2495
2496         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2497                 CERROR("Unexpected connreq msg type: %x from %s\n",
2498                        reqmsg->ibm_type, libcfs_nid2str(nid));
2499                 goto failed;
2500         }
2501
2502         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2503             kiblnd_msg_queue_size(version, ni)) {
2504                 CERROR("Can't accept conn from %s, queue depth too large:  %d (<=%d wanted)\n",
2505                        libcfs_nid2str(nid),
2506                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2507                        kiblnd_msg_queue_size(version, ni));
2508
2509                 if (version == IBLND_MSG_VERSION)
2510                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2511
2512                 goto failed;
2513         }
2514
2515         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2516             IBLND_MAX_RDMA_FRAGS) {
2517                 CWARN("Can't accept conn from %s (version %x): max_frags %d too large (%d wanted)\n",
2518                       libcfs_nid2str(nid), version,
2519                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2520                       IBLND_MAX_RDMA_FRAGS);
2521
2522                 if (version >= IBLND_MSG_VERSION)
2523                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2524
2525                 goto failed;
2526         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2527                    IBLND_MAX_RDMA_FRAGS &&
2528                    net->ibn_fmr_ps == NULL) {
2529                 CWARN("Can't accept conn from %s (version %x): max_frags %d incompatible without FMR pool (%d wanted)\n",
2530                       libcfs_nid2str(nid), version,
2531                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2532                       IBLND_MAX_RDMA_FRAGS);
2533
2534                 if (version == IBLND_MSG_VERSION)
2535                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2536
2537                 goto failed;
2538         }
2539
2540         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2541                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2542                        libcfs_nid2str(nid),
2543                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2544                        IBLND_MSG_SIZE);
2545                 goto failed;
2546         }
2547
2548         /* assume 'nid' is a new peer_ni; create  */
2549         rc = kiblnd_create_peer(ni, &peer_ni, nid);
2550         if (rc != 0) {
2551                 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2552                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2553                 goto failed;
2554         }
2555
2556         /* We have validated the peer's parameters so use those */
2557         peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2558         peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2559
2560         write_lock_irqsave(g_lock, flags);
2561
2562         peer2 = kiblnd_find_peer_locked(ni, nid);
2563         if (peer2 != NULL) {
2564                 if (peer2->ibp_version == 0) {
2565                         peer2->ibp_version     = version;
2566                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2567                 }
2568
2569                 /* not the guy I've talked with */
2570                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2571                     peer2->ibp_version     != version) {
2572                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2573
2574                         if (kiblnd_peer_active(peer2)) {
2575                                 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2576                                 peer2->ibp_version = version;
2577                         }
2578                         write_unlock_irqrestore(g_lock, flags);
2579
2580                         CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2581                               libcfs_nid2str(nid), peer2->ibp_version, version,
2582                               peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2583
2584                         kiblnd_peer_decref(peer_ni);
2585                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2586                         goto failed;
2587                 }
2588
2589                 /* Tie-break connection race in favour of the higher NID.
2590                  * If we keep running into a race condition multiple times,
2591                  * we have to assume that the connection attempt with the
2592                  * higher NID is stuck in a connecting state and will never
2593                  * recover.  As such, we pass through this if-block and let
2594                  * the lower NID connection win so we can move forward.
2595                  */
2596                 if (peer2->ibp_connecting != 0 &&
2597                     nid < ni->ni_nid && peer2->ibp_races <
2598                     MAX_CONN_RACES_BEFORE_ABORT) {
2599                         peer2->ibp_races++;
2600                         write_unlock_irqrestore(g_lock, flags);
2601
2602                         CDEBUG(D_NET, "Conn race %s\n",
2603                                libcfs_nid2str(peer2->ibp_nid));
2604
2605                         kiblnd_peer_decref(peer_ni);
2606                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2607                         goto failed;
2608                 }
2609                 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2610                         CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2611                                 libcfs_nid2str(peer2->ibp_nid),
2612                                 MAX_CONN_RACES_BEFORE_ABORT);
2613                 /*
2614                  * passive connection is allowed even this peer_ni is waiting for
2615                  * reconnection.
2616                  */
2617                 peer2->ibp_reconnecting = 0;
2618                 peer2->ibp_races = 0;
2619                 peer2->ibp_accepting++;
2620                 kiblnd_peer_addref(peer2);
2621
2622                 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2623                  * so copy validated parameters since we now know what the
2624                  * peer_ni's limits are */
2625                 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2626                 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2627
2628                 write_unlock_irqrestore(g_lock, flags);
2629                 kiblnd_peer_decref(peer_ni);
2630                 peer_ni = peer2;
2631         } else {
2632                 /* Brand new peer_ni */
2633                 LASSERT(peer_ni->ibp_accepting == 0);
2634                 LASSERT(peer_ni->ibp_version == 0 &&
2635                         peer_ni->ibp_incarnation == 0);
2636
2637                 peer_ni->ibp_accepting   = 1;
2638                 peer_ni->ibp_version     = version;
2639                 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2640
2641                 /* I have a ref on ni that prevents it being shutdown */
2642                 LASSERT(net->ibn_shutdown == 0);
2643
2644                 kiblnd_peer_addref(peer_ni);
2645                 hash_add(kiblnd_data.kib_peers, &peer_ni->ibp_list, nid);
2646
2647                 write_unlock_irqrestore(g_lock, flags);
2648         }
2649
2650         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT,
2651                                   version);
2652         if (!conn) {
2653                 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2654                 kiblnd_peer_decref(peer_ni);
2655                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2656                 goto failed;
2657         }
2658
2659         /* conn now "owns" cmid, so I return success from here on to ensure the
2660          * CM callback doesn't destroy cmid.
2661          */
2662         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2663         conn->ibc_credits          = conn->ibc_queue_depth;
2664         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2665         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2666                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2667
2668         ackmsg = &conn->ibc_connvars->cv_msg;
2669         memset(ackmsg, 0, sizeof(*ackmsg));
2670
2671         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2672                         sizeof(ackmsg->ibm_u.connparams));
2673         ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2674         ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2675         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2676
2677         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2678
2679         memset(&cp, 0, sizeof(cp));
2680         cp.private_data        = ackmsg;
2681         cp.private_data_len    = ackmsg->ibm_nob;
2682         cp.responder_resources = 0;            /* No atomic ops or RDMA reads */
2683         cp.initiator_depth     = 0;
2684         cp.flow_control        = 1;
2685         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2686         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2687
2688         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2689
2690         rc = rdma_accept(cmid, &cp);
2691         if (rc != 0) {
2692                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2693                 rej.ibr_version = version;
2694                 rej.ibr_why     = IBLND_REJECT_FATAL;
2695
2696                 kiblnd_reject(cmid, &rej);
2697                 kiblnd_connreq_done(conn, rc);
2698                 kiblnd_conn_decref(conn);
2699         }
2700
2701         lnet_ni_decref(ni);
2702         return 0;
2703
2704  failed:
2705         if (ni != NULL) {
2706                 rej.ibr_cp.ibcp_queue_depth =
2707                         kiblnd_msg_queue_size(version, ni);
2708                 rej.ibr_cp.ibcp_max_frags   = IBLND_MAX_RDMA_FRAGS;
2709                 lnet_ni_decref(ni);
2710         }
2711
2712         rej.ibr_version = version;
2713         kiblnd_reject(cmid, &rej);
2714
2715         return -ECONNREFUSED;
2716 }
2717
2718 static void
2719 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2720                        u64 incarnation, int why, struct kib_connparams *cp)
2721 {
2722         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
2723         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2724         char            *reason;
2725         int              msg_size = IBLND_MSG_SIZE;
2726         int              frag_num = -1;
2727         int              queue_dep = -1;
2728         bool             reconnect;
2729         unsigned long    flags;
2730
2731         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2732         LASSERT(peer_ni->ibp_connecting > 0);   /* 'conn' at least */
2733
2734         if (cp) {
2735                 msg_size        = cp->ibcp_max_msg_size;
2736                 frag_num        = cp->ibcp_max_frags;
2737                 queue_dep       = cp->ibcp_queue_depth;
2738         }
2739
2740         write_lock_irqsave(glock, flags);
2741         /* retry connection if it's still needed and no other connection
2742          * attempts (active or passive) are in progress
2743          * NB: reconnect is still needed even when ibp_tx_queue is
2744          * empty if ibp_version != version because reconnect may be
2745          * initiated.
2746          */
2747         reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2748                      peer_ni->ibp_version != version) &&
2749                     peer_ni->ibp_connecting &&
2750                     peer_ni->ibp_accepting == 0;
2751         if (!reconnect) {
2752                 reason = "no need";
2753                 goto out;
2754         }
2755
2756         if (peer_ni->ibp_retries > *kiblnd_tunables.kib_retry_count) {
2757                 reason = "retry count exceeded due to no listener";
2758                 goto out;
2759         }
2760
2761         switch (why) {
2762         default:
2763                 reason = "Unknown";
2764                 break;
2765
2766         case IBLND_REJECT_RDMA_FRAGS: {
2767                 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2768
2769                 if (!cp) {
2770                         reason = "can't negotiate max frags";
2771                         goto out;
2772                 }
2773                 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2774 #ifdef HAVE_IB_GET_DMA_MR
2775                 /*
2776                  * This check only makes sense if the kernel supports global
2777                  * memory registration. Otherwise, map_on_demand will never == 0
2778                  */
2779                 if (!tunables->lnd_map_on_demand) {
2780                         reason = "map_on_demand must be enabled";
2781                         goto out;
2782                 }
2783 #endif
2784                 if (conn->ibc_max_frags <= frag_num) {
2785                         reason = "unsupported max frags";
2786                         goto out;
2787                 }
2788
2789                 peer_ni->ibp_max_frags = frag_num;
2790                 reason = "rdma fragments";
2791                 break;
2792         }
2793         case IBLND_REJECT_MSG_QUEUE_SIZE:
2794                 if (!cp) {
2795                         reason = "can't negotiate queue depth";
2796                         goto out;
2797                 }
2798                 if (conn->ibc_queue_depth <= queue_dep) {
2799                         reason = "unsupported queue depth";
2800                         goto out;
2801                 }
2802
2803                 peer_ni->ibp_queue_depth = queue_dep;
2804                 reason = "queue depth";
2805                 break;
2806
2807         case IBLND_REJECT_CONN_STALE:
2808                 reason = "stale";
2809                 break;
2810
2811         case IBLND_REJECT_CONN_RACE:
2812                 reason = "conn race";
2813                 break;
2814
2815         case IBLND_REJECT_CONN_UNCOMPAT:
2816                 reason = "version negotiation";
2817                 break;
2818
2819         case IBLND_REJECT_INVALID_SRV_ID:
2820                 reason = "invalid service id";
2821                 break;
2822         }
2823
2824         conn->ibc_reconnect = 1;
2825         peer_ni->ibp_reconnecting++;
2826         peer_ni->ibp_version = version;
2827         if (incarnation != 0)
2828                 peer_ni->ibp_incarnation = incarnation;
2829  out:
2830         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2831
2832         CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2833                 libcfs_nid2str(peer_ni->ibp_nid),
2834                 reconnect ? "reconnect" : "don't reconnect",
2835                 reason, IBLND_MSG_VERSION, version, msg_size,
2836                 conn->ibc_queue_depth, queue_dep,
2837                 conn->ibc_max_frags, frag_num);
2838         /*
2839          * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2840          * while destroying the zombie
2841          */
2842 }
2843
2844 static void
2845 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2846 {
2847         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2848
2849         LASSERT (!in_interrupt());
2850         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2851
2852         switch (reason) {
2853         case IB_CM_REJ_STALE_CONN:
2854                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2855                                        IBLND_REJECT_CONN_STALE, NULL);
2856                 break;
2857
2858         case IB_CM_REJ_INVALID_SERVICE_ID:
2859                 peer_ni->ibp_retries++;
2860                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2861                                        IBLND_REJECT_INVALID_SRV_ID, NULL);
2862                 CNETERR("%s rejected: no listener at %d\n",
2863                         libcfs_nid2str(peer_ni->ibp_nid),
2864                         *kiblnd_tunables.kib_service);
2865                 break;
2866
2867         case IB_CM_REJ_CONSUMER_DEFINED:
2868                 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2869                         struct kib_rej *rej = priv;
2870                         struct kib_connparams *cp = NULL;
2871                         bool flip = false;
2872                         __u64 incarnation = -1;
2873
2874                         /* NB. default incarnation is -1 because:
2875                          * a) V1 will ignore dst incarnation in connreq.
2876                          * b) V2 will provide incarnation while rejecting me,
2877                          *    -1 will be overwrote.
2878                          *
2879                          * if I try to connect to a V1 peer_ni with V2 protocol,
2880                          * it rejected me then upgrade to V2, I have no idea
2881                          * about the upgrading and try to reconnect with V1,
2882                          * in this case upgraded V2 can find out I'm trying to
2883                          * talk to the old guy and reject me(incarnation is -1).
2884                          */
2885
2886                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2887                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2888                                 __swab32s(&rej->ibr_magic);
2889                                 __swab16s(&rej->ibr_version);
2890                                 flip = true;
2891                         }
2892
2893                         if (priv_nob >= sizeof(struct kib_rej) &&
2894                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2895                                 /* priv_nob is always 148 in current version
2896                                  * of OFED, so we still need to check version.
2897                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2898                                  */
2899                                 cp = &rej->ibr_cp;
2900
2901                                 if (flip) {
2902                                         __swab64s(&rej->ibr_incarnation);
2903                                         __swab16s(&cp->ibcp_queue_depth);
2904                                         __swab16s(&cp->ibcp_max_frags);
2905                                         __swab32s(&cp->ibcp_max_msg_size);
2906                                 }
2907
2908                                 incarnation = rej->ibr_incarnation;
2909                         }
2910
2911                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2912                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2913                                 CERROR("%s rejected: consumer defined fatal error\n",
2914                                        libcfs_nid2str(peer_ni->ibp_nid));
2915                                 break;
2916                         }
2917
2918                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2919                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2920                                 CERROR("%s rejected: o2iblnd version %x error\n",
2921                                        libcfs_nid2str(peer_ni->ibp_nid),
2922                                        rej->ibr_version);
2923                                 break;
2924                         }
2925
2926                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2927                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2928                                 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2929                                        libcfs_nid2str(peer_ni->ibp_nid),
2930                                        rej->ibr_version);
2931
2932                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2933                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2934                         }
2935
2936                         switch (rej->ibr_why) {
2937                         case IBLND_REJECT_CONN_RACE:
2938                         case IBLND_REJECT_CONN_STALE:
2939                         case IBLND_REJECT_CONN_UNCOMPAT:
2940                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2941                         case IBLND_REJECT_RDMA_FRAGS:
2942                                 kiblnd_check_reconnect(conn, rej->ibr_version,
2943                                                        incarnation,
2944                                                        rej->ibr_why, cp);
2945                                 break;
2946
2947                         case IBLND_REJECT_NO_RESOURCES:
2948                                 CERROR("%s rejected: o2iblnd no resources\n",
2949                                        libcfs_nid2str(peer_ni->ibp_nid));
2950                                 break;
2951
2952                         case IBLND_REJECT_FATAL:
2953                                 CERROR("%s rejected: o2iblnd fatal error\n",
2954                                        libcfs_nid2str(peer_ni->ibp_nid));
2955                                 break;
2956
2957                         default:
2958                                 CERROR("%s rejected: o2iblnd reason %d\n",
2959                                        libcfs_nid2str(peer_ni->ibp_nid),
2960                                        rej->ibr_why);
2961                                 break;
2962                         }
2963                         break;
2964                 }
2965                 /* fall through */
2966         default:
2967                 CNETERR("%s rejected: reason %d, size %d\n",
2968                         libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2969                 break;
2970         }
2971
2972         kiblnd_connreq_done(conn, -ECONNREFUSED);
2973 }
2974
2975 static void
2976 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2977 {
2978         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2979         struct lnet_ni *ni = peer_ni->ibp_ni;
2980         struct kib_net *net = ni->ni_data;
2981         struct kib_msg *msg = priv;
2982         int            ver  = conn->ibc_version;
2983         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2984         unsigned long  flags;
2985
2986         LASSERT (net != NULL);
2987
2988         if (rc != 0) {
2989                 CERROR("Can't unpack connack from %s: %d\n",
2990                        libcfs_nid2str(peer_ni->ibp_nid), rc);
2991                 goto failed;
2992         }
2993
2994         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2995                 CERROR("Unexpected message %d from %s\n",
2996                        msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2997                 rc = -EPROTO;
2998                 goto failed;
2999         }
3000
3001         if (ver != msg->ibm_version) {
3002                 CERROR("%s replied version %x is different with "
3003                        "requested version %x\n",
3004                        libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
3005                 rc = -EPROTO;
3006                 goto failed;
3007         }
3008
3009         if (msg->ibm_u.connparams.ibcp_queue_depth >
3010             conn->ibc_queue_depth) {
3011                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
3012                        libcfs_nid2str(peer_ni->ibp_nid),
3013                        msg->ibm_u.connparams.ibcp_queue_depth,
3014                        conn->ibc_queue_depth);
3015                 rc = -EPROTO;
3016                 goto failed;
3017         }
3018
3019         if (msg->ibm_u.connparams.ibcp_max_frags >
3020             conn->ibc_max_frags) {
3021                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
3022                        libcfs_nid2str(peer_ni->ibp_nid),
3023                        msg->ibm_u.connparams.ibcp_max_frags,
3024                        conn->ibc_max_frags);
3025                 rc = -EPROTO;
3026                 goto failed;
3027         }
3028
3029         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
3030                 CERROR("%s max message size %d too big (%d max)\n",
3031                        libcfs_nid2str(peer_ni->ibp_nid),
3032                        msg->ibm_u.connparams.ibcp_max_msg_size,
3033                        IBLND_MSG_SIZE);
3034                 rc = -EPROTO;
3035                 goto failed;
3036         }
3037
3038         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3039         if (msg->ibm_dstnid == ni->ni_nid &&
3040             msg->ibm_dststamp == net->ibn_incarnation)
3041                 rc = 0;
3042         else
3043                 rc = -ESTALE;
3044         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3045
3046         if (rc != 0) {
3047                 CERROR("Bad connection reply from %s, rc = %d, "
3048                        "version: %x max_frags: %d\n",
3049                        libcfs_nid2str(peer_ni->ibp_nid), rc,
3050                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
3051                 goto failed;
3052         }
3053
3054         conn->ibc_incarnation      = msg->ibm_srcstamp;
3055         conn->ibc_credits          = msg->ibm_u.connparams.ibcp_queue_depth;
3056         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3057         conn->ibc_queue_depth      = msg->ibm_u.connparams.ibcp_queue_depth;
3058         conn->ibc_max_frags        = msg->ibm_u.connparams.ibcp_max_frags;
3059         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3060                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3061
3062         kiblnd_connreq_done(conn, 0);
3063         return;
3064
3065  failed:
3066         /* NB My QP has already established itself, so I handle anything going
3067          * wrong here by setting ibc_comms_error.
3068          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3069          * immediately tears it down. */
3070
3071         LASSERT (rc != 0);
3072         conn->ibc_comms_error = rc;
3073         kiblnd_connreq_done(conn, 0);
3074 }
3075
3076 static int
3077 kiblnd_active_connect(struct rdma_cm_id *cmid)
3078 {
3079         struct kib_peer_ni *peer_ni = cmid->context;
3080         struct kib_conn *conn;
3081         struct kib_msg *msg;
3082         struct rdma_conn_param cp;
3083         int                      version;
3084         __u64                    incarnation;
3085         unsigned long            flags;
3086         int                      rc;
3087
3088         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3089
3090         incarnation = peer_ni->ibp_incarnation;
3091         version     = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3092                                                  peer_ni->ibp_version;
3093
3094         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3095
3096         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3097                                   version);
3098         if (conn == NULL) {
3099                 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3100                 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3101                 return -ENOMEM;
3102         }
3103
3104         /* conn "owns" cmid now, so I return success from here on to ensure the
3105          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3106          * on peer_ni */
3107
3108         msg = &conn->ibc_connvars->cv_msg;
3109
3110         memset(msg, 0, sizeof(*msg));
3111         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3112         msg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
3113         msg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
3114         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3115