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.sun.com/software/products/lustre/docs/GPLv2.pdf
19 *
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26 /*
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2012, Intel Corporation.
31 */
32 /*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 *
36 * lnet/klnds/o2iblnd/o2iblnd_cb.c
37 *
38 * Author: Eric Barton <eric@bartonsoftware.com>
39 */
40
41 #include "o2iblnd.h"
42
43 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
44
45 static void
kiblnd_tx_done(lnet_ni_t * ni,kib_tx_t * tx)46 kiblnd_tx_done(lnet_ni_t *ni, kib_tx_t *tx)
47 {
48 lnet_msg_t *lntmsg[2];
49 kib_net_t *net = ni->ni_data;
50 int rc;
51 int i;
52
53 LASSERT(net != NULL);
54 LASSERT(!in_interrupt());
55 LASSERT(!tx->tx_queued); /* mustn't be queued for sending */
56 LASSERT(tx->tx_sending == 0); /* mustn't be awaiting sent callback */
57 LASSERT(!tx->tx_waiting); /* mustn't be awaiting peer response */
58 LASSERT(tx->tx_pool != NULL);
59
60 kiblnd_unmap_tx(ni, tx);
61
62 /* tx may have up to 2 lnet msgs to finalise */
63 lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
64 lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
65 rc = tx->tx_status;
66
67 if (tx->tx_conn != NULL) {
68 LASSERT(ni == tx->tx_conn->ibc_peer->ibp_ni);
69
70 kiblnd_conn_decref(tx->tx_conn);
71 tx->tx_conn = NULL;
72 }
73
74 tx->tx_nwrq = 0;
75 tx->tx_status = 0;
76
77 kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
78
79 /* delay finalize until my descs have been freed */
80 for (i = 0; i < 2; i++) {
81 if (lntmsg[i] == NULL)
82 continue;
83
84 lnet_finalize(ni, lntmsg[i], rc);
85 }
86 }
87
88 void
kiblnd_txlist_done(lnet_ni_t * ni,struct list_head * txlist,int status)89 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
90 {
91 kib_tx_t *tx;
92
93 while (!list_empty(txlist)) {
94 tx = list_entry(txlist->next, kib_tx_t, tx_list);
95
96 list_del(&tx->tx_list);
97 /* complete now */
98 tx->tx_waiting = 0;
99 tx->tx_status = status;
100 kiblnd_tx_done(ni, tx);
101 }
102 }
103
104 static kib_tx_t *
kiblnd_get_idle_tx(lnet_ni_t * ni,lnet_nid_t target)105 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
106 {
107 kib_net_t *net = (kib_net_t *)ni->ni_data;
108 struct list_head *node;
109 kib_tx_t *tx;
110 kib_tx_poolset_t *tps;
111
112 tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
113 node = kiblnd_pool_alloc_node(&tps->tps_poolset);
114 if (node == NULL)
115 return NULL;
116 tx = container_of(node, kib_tx_t, tx_list);
117
118 LASSERT(tx->tx_nwrq == 0);
119 LASSERT(!tx->tx_queued);
120 LASSERT(tx->tx_sending == 0);
121 LASSERT(!tx->tx_waiting);
122 LASSERT(tx->tx_status == 0);
123 LASSERT(tx->tx_conn == NULL);
124 LASSERT(tx->tx_lntmsg[0] == NULL);
125 LASSERT(tx->tx_lntmsg[1] == NULL);
126 LASSERT(tx->tx_nfrags == 0);
127
128 return tx;
129 }
130
131 static void
kiblnd_drop_rx(kib_rx_t * rx)132 kiblnd_drop_rx(kib_rx_t *rx)
133 {
134 kib_conn_t *conn = rx->rx_conn;
135 struct kib_sched_info *sched = conn->ibc_sched;
136 unsigned long flags;
137
138 spin_lock_irqsave(&sched->ibs_lock, flags);
139 LASSERT(conn->ibc_nrx > 0);
140 conn->ibc_nrx--;
141 spin_unlock_irqrestore(&sched->ibs_lock, flags);
142
143 kiblnd_conn_decref(conn);
144 }
145
146 int
kiblnd_post_rx(kib_rx_t * rx,int credit)147 kiblnd_post_rx(kib_rx_t *rx, int credit)
148 {
149 kib_conn_t *conn = rx->rx_conn;
150 kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
151 struct ib_recv_wr *bad_wrq = NULL;
152 struct ib_mr *mr;
153 int rc;
154
155 LASSERT(net != NULL);
156 LASSERT(!in_interrupt());
157 LASSERT(credit == IBLND_POSTRX_NO_CREDIT ||
158 credit == IBLND_POSTRX_PEER_CREDIT ||
159 credit == IBLND_POSTRX_RSRVD_CREDIT);
160
161 mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
162 LASSERT(mr != NULL);
163
164 rx->rx_sge.lkey = mr->lkey;
165 rx->rx_sge.addr = rx->rx_msgaddr;
166 rx->rx_sge.length = IBLND_MSG_SIZE;
167
168 rx->rx_wrq.next = NULL;
169 rx->rx_wrq.sg_list = &rx->rx_sge;
170 rx->rx_wrq.num_sge = 1;
171 rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
172
173 LASSERT(conn->ibc_state >= IBLND_CONN_INIT);
174 LASSERT(rx->rx_nob >= 0); /* not posted */
175
176 if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
177 kiblnd_drop_rx(rx); /* No more posts for this rx */
178 return 0;
179 }
180
181 rx->rx_nob = -1; /* flag posted */
182
183 /* NB: need an extra reference after ib_post_recv because we don't
184 * own this rx (and rx::rx_conn) anymore, LU-5678.
185 */
186 kiblnd_conn_addref(conn);
187 rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
188 if (unlikely(rc != 0)) {
189 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
190 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
191 rx->rx_nob = 0;
192 }
193
194 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
195 goto out;
196
197 if (unlikely(rc != 0)) {
198 kiblnd_close_conn(conn, rc);
199 kiblnd_drop_rx(rx); /* No more posts for this rx */
200 goto out;
201 }
202
203 if (credit == IBLND_POSTRX_NO_CREDIT)
204 goto out;
205
206 spin_lock(&conn->ibc_lock);
207 if (credit == IBLND_POSTRX_PEER_CREDIT)
208 conn->ibc_outstanding_credits++;
209 else
210 conn->ibc_reserved_credits++;
211 spin_unlock(&conn->ibc_lock);
212
213 kiblnd_check_sends(conn);
214 out:
215 kiblnd_conn_decref(conn);
216 return rc;
217 }
218
219 static kib_tx_t *
kiblnd_find_waiting_tx_locked(kib_conn_t * conn,int txtype,__u64 cookie)220 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
221 {
222 struct list_head *tmp;
223
224 list_for_each(tmp, &conn->ibc_active_txs) {
225 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
226
227 LASSERT(!tx->tx_queued);
228 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
229
230 if (tx->tx_cookie != cookie)
231 continue;
232
233 if (tx->tx_waiting &&
234 tx->tx_msg->ibm_type == txtype)
235 return tx;
236
237 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
238 tx->tx_waiting ? "" : "NOT ",
239 tx->tx_msg->ibm_type, txtype);
240 }
241 return NULL;
242 }
243
244 static void
kiblnd_handle_completion(kib_conn_t * conn,int txtype,int status,__u64 cookie)245 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
246 {
247 kib_tx_t *tx;
248 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
249 int idle;
250
251 spin_lock(&conn->ibc_lock);
252
253 tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
254 if (tx == NULL) {
255 spin_unlock(&conn->ibc_lock);
256
257 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
258 txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
259 kiblnd_close_conn(conn, -EPROTO);
260 return;
261 }
262
263 if (tx->tx_status == 0) { /* success so far */
264 if (status < 0) /* failed? */
265 tx->tx_status = status;
266 else if (txtype == IBLND_MSG_GET_REQ)
267 lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
268 }
269
270 tx->tx_waiting = 0;
271
272 idle = !tx->tx_queued && (tx->tx_sending == 0);
273 if (idle)
274 list_del(&tx->tx_list);
275
276 spin_unlock(&conn->ibc_lock);
277
278 if (idle)
279 kiblnd_tx_done(ni, tx);
280 }
281
282 static void
kiblnd_send_completion(kib_conn_t * conn,int type,int status,__u64 cookie)283 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
284 {
285 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
286 kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
287
288 if (tx == NULL) {
289 CERROR("Can't get tx for completion %x for %s\n",
290 type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
291 return;
292 }
293
294 tx->tx_msg->ibm_u.completion.ibcm_status = status;
295 tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
296 kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
297
298 kiblnd_queue_tx(tx, conn);
299 }
300
301 static void
kiblnd_handle_rx(kib_rx_t * rx)302 kiblnd_handle_rx(kib_rx_t *rx)
303 {
304 kib_msg_t *msg = rx->rx_msg;
305 kib_conn_t *conn = rx->rx_conn;
306 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
307 int credits = msg->ibm_credits;
308 kib_tx_t *tx;
309 int rc = 0;
310 int rc2;
311 int post_credit;
312
313 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
314
315 CDEBUG(D_NET, "Received %x[%d] from %s\n",
316 msg->ibm_type, credits,
317 libcfs_nid2str(conn->ibc_peer->ibp_nid));
318
319 if (credits != 0) {
320 /* Have I received credits that will let me send? */
321 spin_lock(&conn->ibc_lock);
322
323 if (conn->ibc_credits + credits >
324 IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
325 rc2 = conn->ibc_credits;
326 spin_unlock(&conn->ibc_lock);
327
328 CERROR("Bad credits from %s: %d + %d > %d\n",
329 libcfs_nid2str(conn->ibc_peer->ibp_nid),
330 rc2, credits,
331 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
332
333 kiblnd_close_conn(conn, -EPROTO);
334 kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
335 return;
336 }
337
338 conn->ibc_credits += credits;
339
340 /* This ensures the credit taken by NOOP can be returned */
341 if (msg->ibm_type == IBLND_MSG_NOOP &&
342 !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
343 conn->ibc_outstanding_credits++;
344
345 spin_unlock(&conn->ibc_lock);
346 kiblnd_check_sends(conn);
347 }
348
349 switch (msg->ibm_type) {
350 default:
351 CERROR("Bad IBLND message type %x from %s\n",
352 msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
353 post_credit = IBLND_POSTRX_NO_CREDIT;
354 rc = -EPROTO;
355 break;
356
357 case IBLND_MSG_NOOP:
358 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
359 post_credit = IBLND_POSTRX_NO_CREDIT;
360 break;
361 }
362
363 if (credits != 0) /* credit already posted */
364 post_credit = IBLND_POSTRX_NO_CREDIT;
365 else /* a keepalive NOOP */
366 post_credit = IBLND_POSTRX_PEER_CREDIT;
367 break;
368
369 case IBLND_MSG_IMMEDIATE:
370 post_credit = IBLND_POSTRX_DONT_POST;
371 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
372 msg->ibm_srcnid, rx, 0);
373 if (rc < 0) /* repost on error */
374 post_credit = IBLND_POSTRX_PEER_CREDIT;
375 break;
376
377 case IBLND_MSG_PUT_REQ:
378 post_credit = IBLND_POSTRX_DONT_POST;
379 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
380 msg->ibm_srcnid, rx, 1);
381 if (rc < 0) /* repost on error */
382 post_credit = IBLND_POSTRX_PEER_CREDIT;
383 break;
384
385 case IBLND_MSG_PUT_NAK:
386 CWARN("PUT_NACK from %s\n",
387 libcfs_nid2str(conn->ibc_peer->ibp_nid));
388 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
389 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
390 msg->ibm_u.completion.ibcm_status,
391 msg->ibm_u.completion.ibcm_cookie);
392 break;
393
394 case IBLND_MSG_PUT_ACK:
395 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
396
397 spin_lock(&conn->ibc_lock);
398 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
399 msg->ibm_u.putack.ibpam_src_cookie);
400 if (tx != NULL)
401 list_del(&tx->tx_list);
402 spin_unlock(&conn->ibc_lock);
403
404 if (tx == NULL) {
405 CERROR("Unmatched PUT_ACK from %s\n",
406 libcfs_nid2str(conn->ibc_peer->ibp_nid));
407 rc = -EPROTO;
408 break;
409 }
410
411 LASSERT(tx->tx_waiting);
412 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
413 * (a) I can overwrite tx_msg since my peer has received it!
414 * (b) tx_waiting set tells tx_complete() it's not done. */
415
416 tx->tx_nwrq = 0; /* overwrite PUT_REQ */
417
418 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
419 kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
420 &msg->ibm_u.putack.ibpam_rd,
421 msg->ibm_u.putack.ibpam_dst_cookie);
422 if (rc2 < 0)
423 CERROR("Can't setup rdma for PUT to %s: %d\n",
424 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
425
426 spin_lock(&conn->ibc_lock);
427 tx->tx_waiting = 0; /* clear waiting and queue atomically */
428 kiblnd_queue_tx_locked(tx, conn);
429 spin_unlock(&conn->ibc_lock);
430 break;
431
432 case IBLND_MSG_PUT_DONE:
433 post_credit = IBLND_POSTRX_PEER_CREDIT;
434 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
435 msg->ibm_u.completion.ibcm_status,
436 msg->ibm_u.completion.ibcm_cookie);
437 break;
438
439 case IBLND_MSG_GET_REQ:
440 post_credit = IBLND_POSTRX_DONT_POST;
441 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
442 msg->ibm_srcnid, rx, 1);
443 if (rc < 0) /* repost on error */
444 post_credit = IBLND_POSTRX_PEER_CREDIT;
445 break;
446
447 case IBLND_MSG_GET_DONE:
448 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
449 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
450 msg->ibm_u.completion.ibcm_status,
451 msg->ibm_u.completion.ibcm_cookie);
452 break;
453 }
454
455 if (rc < 0) /* protocol error */
456 kiblnd_close_conn(conn, rc);
457
458 if (post_credit != IBLND_POSTRX_DONT_POST)
459 kiblnd_post_rx(rx, post_credit);
460 }
461
462 static void
kiblnd_rx_complete(kib_rx_t * rx,int status,int nob)463 kiblnd_rx_complete(kib_rx_t *rx, int status, int nob)
464 {
465 kib_msg_t *msg = rx->rx_msg;
466 kib_conn_t *conn = rx->rx_conn;
467 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
468 kib_net_t *net = ni->ni_data;
469 int rc;
470 int err = -EIO;
471
472 LASSERT(net != NULL);
473 LASSERT(rx->rx_nob < 0); /* was posted */
474 rx->rx_nob = 0; /* isn't now */
475
476 if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
477 goto ignore;
478
479 if (status != IB_WC_SUCCESS) {
480 CNETERR("Rx from %s failed: %d\n",
481 libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
482 goto failed;
483 }
484
485 LASSERT(nob >= 0);
486 rx->rx_nob = nob;
487
488 rc = kiblnd_unpack_msg(msg, rx->rx_nob);
489 if (rc != 0) {
490 CERROR("Error %d unpacking rx from %s\n",
491 rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
492 goto failed;
493 }
494
495 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
496 msg->ibm_dstnid != ni->ni_nid ||
497 msg->ibm_srcstamp != conn->ibc_incarnation ||
498 msg->ibm_dststamp != net->ibn_incarnation) {
499 CERROR("Stale rx from %s\n",
500 libcfs_nid2str(conn->ibc_peer->ibp_nid));
501 err = -ESTALE;
502 goto failed;
503 }
504
505 /* set time last known alive */
506 kiblnd_peer_alive(conn->ibc_peer);
507
508 /* racing with connection establishment/teardown! */
509
510 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
511 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
512 unsigned long flags;
513
514 write_lock_irqsave(g_lock, flags);
515 /* must check holding global lock to eliminate race */
516 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
517 list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
518 write_unlock_irqrestore(g_lock, flags);
519 return;
520 }
521 write_unlock_irqrestore(g_lock, flags);
522 }
523 kiblnd_handle_rx(rx);
524 return;
525
526 failed:
527 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
528 kiblnd_close_conn(conn, err);
529 ignore:
530 kiblnd_drop_rx(rx); /* Don't re-post rx. */
531 }
532
533 static struct page *
kiblnd_kvaddr_to_page(unsigned long vaddr)534 kiblnd_kvaddr_to_page(unsigned long vaddr)
535 {
536 struct page *page;
537
538 if (is_vmalloc_addr((void *)vaddr)) {
539 page = vmalloc_to_page((void *)vaddr);
540 LASSERT(page != NULL);
541 return page;
542 }
543 #ifdef CONFIG_HIGHMEM
544 if (vaddr >= PKMAP_BASE &&
545 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
546 /* No highmem pages only used for bulk (kiov) I/O */
547 CERROR("find page for address in highmem\n");
548 LBUG();
549 }
550 #endif
551 page = virt_to_page(vaddr);
552 LASSERT(page != NULL);
553 return page;
554 }
555
556 static int
kiblnd_fmr_map_tx(kib_net_t * net,kib_tx_t * tx,kib_rdma_desc_t * rd,int nob)557 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
558 {
559 kib_hca_dev_t *hdev;
560 __u64 *pages = tx->tx_pages;
561 kib_fmr_poolset_t *fps;
562 int npages;
563 int size;
564 int cpt;
565 int rc;
566 int i;
567
568 LASSERT(tx->tx_pool != NULL);
569 LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
570
571 hdev = tx->tx_pool->tpo_hdev;
572
573 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
574 for (size = 0; size < rd->rd_frags[i].rf_nob;
575 size += hdev->ibh_page_size) {
576 pages[npages++] = (rd->rd_frags[i].rf_addr &
577 hdev->ibh_page_mask) + size;
578 }
579 }
580
581 cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
582
583 fps = net->ibn_fmr_ps[cpt];
584 rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
585 if (rc != 0) {
586 CERROR("Can't map %d pages: %d\n", npages, rc);
587 return rc;
588 }
589
590 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
591 * the rkey */
592 rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
593 tx->fmr.fmr_pfmr->fmr->lkey;
594 rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
595 rd->rd_frags[0].rf_nob = nob;
596 rd->rd_nfrags = 1;
597
598 return 0;
599 }
600
kiblnd_unmap_tx(lnet_ni_t * ni,kib_tx_t * tx)601 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
602 {
603 kib_net_t *net = ni->ni_data;
604
605 LASSERT(net != NULL);
606
607 if (net->ibn_fmr_ps && tx->fmr.fmr_pfmr) {
608 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
609 tx->fmr.fmr_pfmr = NULL;
610 }
611
612 if (tx->tx_nfrags != 0) {
613 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
614 tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
615 tx->tx_nfrags = 0;
616 }
617 }
618
kiblnd_map_tx(lnet_ni_t * ni,kib_tx_t * tx,kib_rdma_desc_t * rd,int nfrags)619 static int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
620 int nfrags)
621 {
622 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
623 kib_net_t *net = ni->ni_data;
624 struct ib_mr *mr = NULL;
625 __u32 nob;
626 int i;
627
628 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
629 * RDMA sink */
630 tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
631 tx->tx_nfrags = nfrags;
632
633 rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
634 tx->tx_nfrags, tx->tx_dmadir);
635
636 for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
637 rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len(
638 hdev->ibh_ibdev, &tx->tx_frags[i]);
639 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
640 hdev->ibh_ibdev, &tx->tx_frags[i]);
641 nob += rd->rd_frags[i].rf_nob;
642 }
643
644 /* looking for pre-mapping MR */
645 mr = kiblnd_find_rd_dma_mr(hdev, rd);
646 if (mr != NULL) {
647 /* found pre-mapping MR */
648 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
649 return 0;
650 }
651
652 if (net->ibn_fmr_ps != NULL)
653 return kiblnd_fmr_map_tx(net, tx, rd, nob);
654
655 return -EINVAL;
656 }
657
658 static int
kiblnd_setup_rd_iov(lnet_ni_t * ni,kib_tx_t * tx,kib_rdma_desc_t * rd,unsigned int niov,struct kvec * iov,int offset,int nob)659 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
660 unsigned int niov, struct kvec *iov, int offset, int nob)
661 {
662 kib_net_t *net = ni->ni_data;
663 struct page *page;
664 struct scatterlist *sg;
665 unsigned long vaddr;
666 int fragnob;
667 int page_offset;
668
669 LASSERT(nob > 0);
670 LASSERT(niov > 0);
671 LASSERT(net != NULL);
672
673 while (offset >= iov->iov_len) {
674 offset -= iov->iov_len;
675 niov--;
676 iov++;
677 LASSERT(niov > 0);
678 }
679
680 sg = tx->tx_frags;
681 do {
682 LASSERT(niov > 0);
683
684 vaddr = ((unsigned long)iov->iov_base) + offset;
685 page_offset = vaddr & (PAGE_SIZE - 1);
686 page = kiblnd_kvaddr_to_page(vaddr);
687 if (page == NULL) {
688 CERROR("Can't find page\n");
689 return -EFAULT;
690 }
691
692 fragnob = min((int)(iov->iov_len - offset), nob);
693 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
694
695 sg_set_page(sg, page, fragnob, page_offset);
696 sg++;
697
698 if (offset + fragnob < iov->iov_len) {
699 offset += fragnob;
700 } else {
701 offset = 0;
702 iov++;
703 niov--;
704 }
705 nob -= fragnob;
706 } while (nob > 0);
707
708 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
709 }
710
711 static int
kiblnd_setup_rd_kiov(lnet_ni_t * ni,kib_tx_t * tx,kib_rdma_desc_t * rd,int nkiov,lnet_kiov_t * kiov,int offset,int nob)712 kiblnd_setup_rd_kiov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
713 int nkiov, lnet_kiov_t *kiov, int offset, int nob)
714 {
715 kib_net_t *net = ni->ni_data;
716 struct scatterlist *sg;
717 int fragnob;
718
719 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
720
721 LASSERT(nob > 0);
722 LASSERT(nkiov > 0);
723 LASSERT(net != NULL);
724
725 while (offset >= kiov->kiov_len) {
726 offset -= kiov->kiov_len;
727 nkiov--;
728 kiov++;
729 LASSERT(nkiov > 0);
730 }
731
732 sg = tx->tx_frags;
733 do {
734 LASSERT(nkiov > 0);
735
736 fragnob = min((int)(kiov->kiov_len - offset), nob);
737
738 sg_set_page(sg, kiov->kiov_page, fragnob,
739 kiov->kiov_offset + offset);
740 sg++;
741
742 offset = 0;
743 kiov++;
744 nkiov--;
745 nob -= fragnob;
746 } while (nob > 0);
747
748 return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
749 }
750
751 static int
kiblnd_post_tx_locked(kib_conn_t * conn,kib_tx_t * tx,int credit)752 kiblnd_post_tx_locked(kib_conn_t *conn, kib_tx_t *tx, int credit)
753 __releases(conn->ibc_lock)
754 __acquires(conn->ibc_lock)
755 {
756 kib_msg_t *msg = tx->tx_msg;
757 kib_peer_t *peer = conn->ibc_peer;
758 int ver = conn->ibc_version;
759 int rc;
760 int done;
761 struct ib_send_wr *bad_wrq;
762
763 LASSERT(tx->tx_queued);
764 /* We rely on this for QP sizing */
765 LASSERT(tx->tx_nwrq > 0);
766 LASSERT(tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
767
768 LASSERT(credit == 0 || credit == 1);
769 LASSERT(conn->ibc_outstanding_credits >= 0);
770 LASSERT(conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
771 LASSERT(conn->ibc_credits >= 0);
772 LASSERT(conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
773
774 if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
775 /* tx completions outstanding... */
776 CDEBUG(D_NET, "%s: posted enough\n",
777 libcfs_nid2str(peer->ibp_nid));
778 return -EAGAIN;
779 }
780
781 if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
782 CDEBUG(D_NET, "%s: no credits\n",
783 libcfs_nid2str(peer->ibp_nid));
784 return -EAGAIN;
785 }
786
787 if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
788 conn->ibc_credits == 1 && /* last credit reserved */
789 msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
790 CDEBUG(D_NET, "%s: not using last credit\n",
791 libcfs_nid2str(peer->ibp_nid));
792 return -EAGAIN;
793 }
794
795 /* NB don't drop ibc_lock before bumping tx_sending */
796 list_del(&tx->tx_list);
797 tx->tx_queued = 0;
798
799 if (msg->ibm_type == IBLND_MSG_NOOP &&
800 (!kiblnd_need_noop(conn) || /* redundant NOOP */
801 (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
802 conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
803 /* OK to drop when posted enough NOOPs, since
804 * kiblnd_check_sends will queue NOOP again when
805 * posted NOOPs complete */
806 spin_unlock(&conn->ibc_lock);
807 kiblnd_tx_done(peer->ibp_ni, tx);
808 spin_lock(&conn->ibc_lock);
809 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
810 libcfs_nid2str(peer->ibp_nid),
811 conn->ibc_noops_posted);
812 return 0;
813 }
814
815 kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
816 peer->ibp_nid, conn->ibc_incarnation);
817
818 conn->ibc_credits -= credit;
819 conn->ibc_outstanding_credits = 0;
820 conn->ibc_nsends_posted++;
821 if (msg->ibm_type == IBLND_MSG_NOOP)
822 conn->ibc_noops_posted++;
823
824 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
825 * PUT. If so, it was first queued here as a PUT_REQ, sent and
826 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
827 * and then re-queued here. It's (just) possible that
828 * tx_sending is non-zero if we've not done the tx_complete()
829 * from the first send; hence the ++ rather than = below. */
830 tx->tx_sending++;
831 list_add(&tx->tx_list, &conn->ibc_active_txs);
832
833 /* I'm still holding ibc_lock! */
834 if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
835 rc = -ECONNABORTED;
836 } else if (tx->tx_pool->tpo_pool.po_failed ||
837 conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
838 /* close_conn will launch failover */
839 rc = -ENETDOWN;
840 } else {
841 rc = ib_post_send(conn->ibc_cmid->qp, &tx->tx_wrq->wr, &bad_wrq);
842 }
843
844 conn->ibc_last_send = jiffies;
845
846 if (rc == 0)
847 return 0;
848
849 /* NB credits are transferred in the actual
850 * message, which can only be the last work item */
851 conn->ibc_credits += credit;
852 conn->ibc_outstanding_credits += msg->ibm_credits;
853 conn->ibc_nsends_posted--;
854 if (msg->ibm_type == IBLND_MSG_NOOP)
855 conn->ibc_noops_posted--;
856
857 tx->tx_status = rc;
858 tx->tx_waiting = 0;
859 tx->tx_sending--;
860
861 done = (tx->tx_sending == 0);
862 if (done)
863 list_del(&tx->tx_list);
864
865 spin_unlock(&conn->ibc_lock);
866
867 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
868 CERROR("Error %d posting transmit to %s\n",
869 rc, libcfs_nid2str(peer->ibp_nid));
870 else
871 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
872 rc, libcfs_nid2str(peer->ibp_nid));
873
874 kiblnd_close_conn(conn, rc);
875
876 if (done)
877 kiblnd_tx_done(peer->ibp_ni, tx);
878
879 spin_lock(&conn->ibc_lock);
880
881 return -EIO;
882 }
883
884 void
kiblnd_check_sends(kib_conn_t * conn)885 kiblnd_check_sends(kib_conn_t *conn)
886 {
887 int ver = conn->ibc_version;
888 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
889 kib_tx_t *tx;
890
891 /* Don't send anything until after the connection is established */
892 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
893 CDEBUG(D_NET, "%s too soon\n",
894 libcfs_nid2str(conn->ibc_peer->ibp_nid));
895 return;
896 }
897
898 spin_lock(&conn->ibc_lock);
899
900 LASSERT(conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
901 LASSERT(!IBLND_OOB_CAPABLE(ver) ||
902 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
903 LASSERT(conn->ibc_reserved_credits >= 0);
904
905 while (conn->ibc_reserved_credits > 0 &&
906 !list_empty(&conn->ibc_tx_queue_rsrvd)) {
907 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
908 kib_tx_t, tx_list);
909 list_del(&tx->tx_list);
910 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
911 conn->ibc_reserved_credits--;
912 }
913
914 if (kiblnd_need_noop(conn)) {
915 spin_unlock(&conn->ibc_lock);
916
917 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
918 if (tx != NULL)
919 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
920
921 spin_lock(&conn->ibc_lock);
922 if (tx != NULL)
923 kiblnd_queue_tx_locked(tx, conn);
924 }
925
926 kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
927
928 for (;;) {
929 int credit;
930
931 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
932 credit = 0;
933 tx = list_entry(conn->ibc_tx_queue_nocred.next,
934 kib_tx_t, tx_list);
935 } else if (!list_empty(&conn->ibc_tx_noops)) {
936 LASSERT(!IBLND_OOB_CAPABLE(ver));
937 credit = 1;
938 tx = list_entry(conn->ibc_tx_noops.next,
939 kib_tx_t, tx_list);
940 } else if (!list_empty(&conn->ibc_tx_queue)) {
941 credit = 1;
942 tx = list_entry(conn->ibc_tx_queue.next,
943 kib_tx_t, tx_list);
944 } else
945 break;
946
947 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
948 break;
949 }
950
951 spin_unlock(&conn->ibc_lock);
952
953 kiblnd_conn_decref(conn); /* ...until here */
954 }
955
956 static void
kiblnd_tx_complete(kib_tx_t * tx,int status)957 kiblnd_tx_complete(kib_tx_t *tx, int status)
958 {
959 int failed = (status != IB_WC_SUCCESS);
960 kib_conn_t *conn = tx->tx_conn;
961 int idle;
962
963 LASSERT(tx->tx_sending > 0);
964
965 if (failed) {
966 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
967 CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
968 libcfs_nid2str(conn->ibc_peer->ibp_nid),
969 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
970 status);
971
972 kiblnd_close_conn(conn, -EIO);
973 } else {
974 kiblnd_peer_alive(conn->ibc_peer);
975 }
976
977 spin_lock(&conn->ibc_lock);
978
979 /* I could be racing with rdma completion. Whoever makes 'tx' idle
980 * gets to free it, which also drops its ref on 'conn'. */
981
982 tx->tx_sending--;
983 conn->ibc_nsends_posted--;
984 if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
985 conn->ibc_noops_posted--;
986
987 if (failed) {
988 tx->tx_waiting = 0; /* don't wait for peer */
989 tx->tx_status = -EIO;
990 }
991
992 idle = (tx->tx_sending == 0) && /* This is the final callback */
993 !tx->tx_waiting && /* Not waiting for peer */
994 !tx->tx_queued; /* Not re-queued (PUT_DONE) */
995 if (idle)
996 list_del(&tx->tx_list);
997
998 kiblnd_conn_addref(conn); /* 1 ref for me.... */
999
1000 spin_unlock(&conn->ibc_lock);
1001
1002 if (idle)
1003 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1004
1005 kiblnd_check_sends(conn);
1006
1007 kiblnd_conn_decref(conn); /* ...until here */
1008 }
1009
1010 void
kiblnd_init_tx_msg(lnet_ni_t * ni,kib_tx_t * tx,int type,int body_nob)1011 kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1012 {
1013 kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1014 struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1015 struct ib_rdma_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1016 int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1017 struct ib_mr *mr;
1018
1019 LASSERT(tx->tx_nwrq >= 0);
1020 LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1021 LASSERT(nob <= IBLND_MSG_SIZE);
1022
1023 kiblnd_init_msg(tx->tx_msg, type, body_nob);
1024
1025 mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1026 LASSERT(mr != NULL);
1027
1028 sge->lkey = mr->lkey;
1029 sge->addr = tx->tx_msgaddr;
1030 sge->length = nob;
1031
1032 memset(wrq, 0, sizeof(*wrq));
1033
1034 wrq->wr.next = NULL;
1035 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1036 wrq->wr.sg_list = sge;
1037 wrq->wr.num_sge = 1;
1038 wrq->wr.opcode = IB_WR_SEND;
1039 wrq->wr.send_flags = IB_SEND_SIGNALED;
1040
1041 tx->tx_nwrq++;
1042 }
1043
1044 int
kiblnd_init_rdma(kib_conn_t * conn,kib_tx_t * tx,int type,int resid,kib_rdma_desc_t * dstrd,__u64 dstcookie)1045 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1046 int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1047 {
1048 kib_msg_t *ibmsg = tx->tx_msg;
1049 kib_rdma_desc_t *srcrd = tx->tx_rd;
1050 struct ib_sge *sge = &tx->tx_sge[0];
1051 struct ib_rdma_wr *wrq = &tx->tx_wrq[0], *next;
1052 int rc = resid;
1053 int srcidx;
1054 int dstidx;
1055 int wrknob;
1056
1057 LASSERT(!in_interrupt());
1058 LASSERT(tx->tx_nwrq == 0);
1059 LASSERT(type == IBLND_MSG_GET_DONE ||
1060 type == IBLND_MSG_PUT_DONE);
1061
1062 srcidx = dstidx = 0;
1063
1064 while (resid > 0) {
1065 if (srcidx >= srcrd->rd_nfrags) {
1066 CERROR("Src buffer exhausted: %d frags\n", srcidx);
1067 rc = -EPROTO;
1068 break;
1069 }
1070
1071 if (dstidx == dstrd->rd_nfrags) {
1072 CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1073 rc = -EPROTO;
1074 break;
1075 }
1076
1077 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1078 CERROR("RDMA too fragmented for %s (%d): %d/%d src %d/%d dst frags\n",
1079 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1080 IBLND_RDMA_FRAGS(conn->ibc_version),
1081 srcidx, srcrd->rd_nfrags,
1082 dstidx, dstrd->rd_nfrags);
1083 rc = -EMSGSIZE;
1084 break;
1085 }
1086
1087 wrknob = min(min(kiblnd_rd_frag_size(srcrd, srcidx),
1088 kiblnd_rd_frag_size(dstrd, dstidx)),
1089 (__u32) resid);
1090
1091 sge = &tx->tx_sge[tx->tx_nwrq];
1092 sge->addr = kiblnd_rd_frag_addr(srcrd, srcidx);
1093 sge->lkey = kiblnd_rd_frag_key(srcrd, srcidx);
1094 sge->length = wrknob;
1095
1096 wrq = &tx->tx_wrq[tx->tx_nwrq];
1097 next = wrq + 1;
1098
1099 wrq->wr.next = &next->wr;
1100 wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1101 wrq->wr.sg_list = sge;
1102 wrq->wr.num_sge = 1;
1103 wrq->wr.opcode = IB_WR_RDMA_WRITE;
1104 wrq->wr.send_flags = 0;
1105
1106 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1107 wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
1108
1109 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1110 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1111
1112 resid -= wrknob;
1113
1114 tx->tx_nwrq++;
1115 wrq++;
1116 sge++;
1117 }
1118
1119 if (rc < 0) /* no RDMA if completing with failure */
1120 tx->tx_nwrq = 0;
1121
1122 ibmsg->ibm_u.completion.ibcm_status = rc;
1123 ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1124 kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1125 type, sizeof(kib_completion_msg_t));
1126
1127 return rc;
1128 }
1129
1130 void
kiblnd_queue_tx_locked(kib_tx_t * tx,kib_conn_t * conn)1131 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1132 {
1133 struct list_head *q;
1134
1135 LASSERT(tx->tx_nwrq > 0); /* work items set up */
1136 LASSERT(!tx->tx_queued); /* not queued for sending already */
1137 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1138
1139 tx->tx_queued = 1;
1140 tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1141
1142 if (tx->tx_conn == NULL) {
1143 kiblnd_conn_addref(conn);
1144 tx->tx_conn = conn;
1145 LASSERT(tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1146 } else {
1147 /* PUT_DONE first attached to conn as a PUT_REQ */
1148 LASSERT(tx->tx_conn == conn);
1149 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1150 }
1151
1152 switch (tx->tx_msg->ibm_type) {
1153 default:
1154 LBUG();
1155
1156 case IBLND_MSG_PUT_REQ:
1157 case IBLND_MSG_GET_REQ:
1158 q = &conn->ibc_tx_queue_rsrvd;
1159 break;
1160
1161 case IBLND_MSG_PUT_NAK:
1162 case IBLND_MSG_PUT_ACK:
1163 case IBLND_MSG_PUT_DONE:
1164 case IBLND_MSG_GET_DONE:
1165 q = &conn->ibc_tx_queue_nocred;
1166 break;
1167
1168 case IBLND_MSG_NOOP:
1169 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1170 q = &conn->ibc_tx_queue_nocred;
1171 else
1172 q = &conn->ibc_tx_noops;
1173 break;
1174
1175 case IBLND_MSG_IMMEDIATE:
1176 q = &conn->ibc_tx_queue;
1177 break;
1178 }
1179
1180 list_add_tail(&tx->tx_list, q);
1181 }
1182
1183 void
kiblnd_queue_tx(kib_tx_t * tx,kib_conn_t * conn)1184 kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn)
1185 {
1186 spin_lock(&conn->ibc_lock);
1187 kiblnd_queue_tx_locked(tx, conn);
1188 spin_unlock(&conn->ibc_lock);
1189
1190 kiblnd_check_sends(conn);
1191 }
1192
kiblnd_resolve_addr(struct rdma_cm_id * cmid,struct sockaddr_in * srcaddr,struct sockaddr_in * dstaddr,int timeout_ms)1193 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1194 struct sockaddr_in *srcaddr,
1195 struct sockaddr_in *dstaddr,
1196 int timeout_ms)
1197 {
1198 unsigned short port;
1199 int rc;
1200
1201 /* allow the port to be reused */
1202 rc = rdma_set_reuseaddr(cmid, 1);
1203 if (rc != 0) {
1204 CERROR("Unable to set reuse on cmid: %d\n", rc);
1205 return rc;
1206 }
1207
1208 /* look for a free privileged port */
1209 for (port = PROT_SOCK-1; port > 0; port--) {
1210 srcaddr->sin_port = htons(port);
1211 rc = rdma_resolve_addr(cmid,
1212 (struct sockaddr *)srcaddr,
1213 (struct sockaddr *)dstaddr,
1214 timeout_ms);
1215 if (rc == 0) {
1216 CDEBUG(D_NET, "bound to port %hu\n", port);
1217 return 0;
1218 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1219 CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1220 port, rc);
1221 } else {
1222 return rc;
1223 }
1224 }
1225
1226 CERROR("Failed to bind to a free privileged port\n");
1227 return rc;
1228 }
1229
1230 static void
kiblnd_connect_peer(kib_peer_t * peer)1231 kiblnd_connect_peer(kib_peer_t *peer)
1232 {
1233 struct rdma_cm_id *cmid;
1234 kib_dev_t *dev;
1235 kib_net_t *net = peer->ibp_ni->ni_data;
1236 struct sockaddr_in srcaddr;
1237 struct sockaddr_in dstaddr;
1238 int rc;
1239
1240 LASSERT(net != NULL);
1241 LASSERT(peer->ibp_connecting > 0);
1242
1243 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1244 IB_QPT_RC);
1245
1246 if (IS_ERR(cmid)) {
1247 CERROR("Can't create CMID for %s: %ld\n",
1248 libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1249 rc = PTR_ERR(cmid);
1250 goto failed;
1251 }
1252
1253 dev = net->ibn_dev;
1254 memset(&srcaddr, 0, sizeof(srcaddr));
1255 srcaddr.sin_family = AF_INET;
1256 srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1257
1258 memset(&dstaddr, 0, sizeof(dstaddr));
1259 dstaddr.sin_family = AF_INET;
1260 dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1261 dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1262
1263 kiblnd_peer_addref(peer); /* cmid's ref */
1264
1265 if (*kiblnd_tunables.kib_use_priv_port) {
1266 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1267 *kiblnd_tunables.kib_timeout * 1000);
1268 } else {
1269 rc = rdma_resolve_addr(cmid,
1270 (struct sockaddr *)&srcaddr,
1271 (struct sockaddr *)&dstaddr,
1272 *kiblnd_tunables.kib_timeout * 1000);
1273 }
1274 if (rc != 0) {
1275 /* Can't initiate address resolution: */
1276 CERROR("Can't resolve addr for %s: %d\n",
1277 libcfs_nid2str(peer->ibp_nid), rc);
1278 goto failed2;
1279 }
1280
1281 LASSERT(cmid->device != NULL);
1282 CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1283 libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1284 &dev->ibd_ifip, cmid->device->name);
1285
1286 return;
1287
1288 failed2:
1289 kiblnd_peer_decref(peer); /* cmid's ref */
1290 rdma_destroy_id(cmid);
1291 failed:
1292 kiblnd_peer_connect_failed(peer, 1, rc);
1293 }
1294
1295 void
kiblnd_launch_tx(lnet_ni_t * ni,kib_tx_t * tx,lnet_nid_t nid)1296 kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1297 {
1298 kib_peer_t *peer;
1299 kib_peer_t *peer2;
1300 kib_conn_t *conn;
1301 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1302 unsigned long flags;
1303 int rc;
1304
1305 /* If I get here, I've committed to send, so I complete the tx with
1306 * failure on any problems */
1307
1308 LASSERT(tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1309 LASSERT(tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
1310
1311 /* First time, just use a read lock since I expect to find my peer
1312 * connected */
1313 read_lock_irqsave(g_lock, flags);
1314
1315 peer = kiblnd_find_peer_locked(nid);
1316 if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1317 /* Found a peer with an established connection */
1318 conn = kiblnd_get_conn_locked(peer);
1319 kiblnd_conn_addref(conn); /* 1 ref for me... */
1320
1321 read_unlock_irqrestore(g_lock, flags);
1322
1323 if (tx != NULL)
1324 kiblnd_queue_tx(tx, conn);
1325 kiblnd_conn_decref(conn); /* ...to here */
1326 return;
1327 }
1328
1329 read_unlock(g_lock);
1330 /* Re-try with a write lock */
1331 write_lock(g_lock);
1332
1333 peer = kiblnd_find_peer_locked(nid);
1334 if (peer != NULL) {
1335 if (list_empty(&peer->ibp_conns)) {
1336 /* found a peer, but it's still connecting... */
1337 LASSERT(peer->ibp_connecting != 0 ||
1338 peer->ibp_accepting != 0);
1339 if (tx != NULL)
1340 list_add_tail(&tx->tx_list,
1341 &peer->ibp_tx_queue);
1342 write_unlock_irqrestore(g_lock, flags);
1343 } else {
1344 conn = kiblnd_get_conn_locked(peer);
1345 kiblnd_conn_addref(conn); /* 1 ref for me... */
1346
1347 write_unlock_irqrestore(g_lock, flags);
1348
1349 if (tx != NULL)
1350 kiblnd_queue_tx(tx, conn);
1351 kiblnd_conn_decref(conn); /* ...to here */
1352 }
1353 return;
1354 }
1355
1356 write_unlock_irqrestore(g_lock, flags);
1357
1358 /* Allocate a peer ready to add to the peer table and retry */
1359 rc = kiblnd_create_peer(ni, &peer, nid);
1360 if (rc != 0) {
1361 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1362 if (tx != NULL) {
1363 tx->tx_status = -EHOSTUNREACH;
1364 tx->tx_waiting = 0;
1365 kiblnd_tx_done(ni, tx);
1366 }
1367 return;
1368 }
1369
1370 write_lock_irqsave(g_lock, flags);
1371
1372 peer2 = kiblnd_find_peer_locked(nid);
1373 if (peer2 != NULL) {
1374 if (list_empty(&peer2->ibp_conns)) {
1375 /* found a peer, but it's still connecting... */
1376 LASSERT(peer2->ibp_connecting != 0 ||
1377 peer2->ibp_accepting != 0);
1378 if (tx != NULL)
1379 list_add_tail(&tx->tx_list,
1380 &peer2->ibp_tx_queue);
1381 write_unlock_irqrestore(g_lock, flags);
1382 } else {
1383 conn = kiblnd_get_conn_locked(peer2);
1384 kiblnd_conn_addref(conn); /* 1 ref for me... */
1385
1386 write_unlock_irqrestore(g_lock, flags);
1387
1388 if (tx != NULL)
1389 kiblnd_queue_tx(tx, conn);
1390 kiblnd_conn_decref(conn); /* ...to here */
1391 }
1392
1393 kiblnd_peer_decref(peer);
1394 return;
1395 }
1396
1397 /* Brand new peer */
1398 LASSERT(peer->ibp_connecting == 0);
1399 peer->ibp_connecting = 1;
1400
1401 /* always called with a ref on ni, which prevents ni being shutdown */
1402 LASSERT(((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1403
1404 if (tx != NULL)
1405 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1406
1407 kiblnd_peer_addref(peer);
1408 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1409
1410 write_unlock_irqrestore(g_lock, flags);
1411
1412 kiblnd_connect_peer(peer);
1413 kiblnd_peer_decref(peer);
1414 }
1415
1416 int
kiblnd_send(lnet_ni_t * ni,void * private,lnet_msg_t * lntmsg)1417 kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1418 {
1419 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1420 int type = lntmsg->msg_type;
1421 lnet_process_id_t target = lntmsg->msg_target;
1422 int target_is_router = lntmsg->msg_target_is_router;
1423 int routing = lntmsg->msg_routing;
1424 unsigned int payload_niov = lntmsg->msg_niov;
1425 struct kvec *payload_iov = lntmsg->msg_iov;
1426 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1427 unsigned int payload_offset = lntmsg->msg_offset;
1428 unsigned int payload_nob = lntmsg->msg_len;
1429 kib_msg_t *ibmsg;
1430 kib_rdma_desc_t *rd;
1431 kib_tx_t *tx;
1432 int nob;
1433 int rc;
1434
1435 /* NB 'private' is different depending on what we're sending.... */
1436
1437 CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1438 payload_nob, payload_niov, libcfs_id2str(target));
1439
1440 LASSERT(payload_nob == 0 || payload_niov > 0);
1441 LASSERT(payload_niov <= LNET_MAX_IOV);
1442
1443 /* Thread context */
1444 LASSERT(!in_interrupt());
1445 /* payload is either all vaddrs or all pages */
1446 LASSERT(!(payload_kiov != NULL && payload_iov != NULL));
1447
1448 switch (type) {
1449 default:
1450 LBUG();
1451 return -EIO;
1452
1453 case LNET_MSG_ACK:
1454 LASSERT(payload_nob == 0);
1455 break;
1456
1457 case LNET_MSG_GET:
1458 if (routing || target_is_router)
1459 break; /* send IMMEDIATE */
1460
1461 /* is the REPLY message too small for RDMA? */
1462 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1463 if (nob <= IBLND_MSG_SIZE)
1464 break; /* send IMMEDIATE */
1465
1466 tx = kiblnd_get_idle_tx(ni, target.nid);
1467 if (tx == NULL) {
1468 CERROR("Can't allocate txd for GET to %s\n",
1469 libcfs_nid2str(target.nid));
1470 return -ENOMEM;
1471 }
1472
1473 ibmsg = tx->tx_msg;
1474 rd = &ibmsg->ibm_u.get.ibgm_rd;
1475 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1476 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1477 lntmsg->msg_md->md_niov,
1478 lntmsg->msg_md->md_iov.iov,
1479 0, lntmsg->msg_md->md_length);
1480 else
1481 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1482 lntmsg->msg_md->md_niov,
1483 lntmsg->msg_md->md_iov.kiov,
1484 0, lntmsg->msg_md->md_length);
1485 if (rc != 0) {
1486 CERROR("Can't setup GET sink for %s: %d\n",
1487 libcfs_nid2str(target.nid), rc);
1488 kiblnd_tx_done(ni, tx);
1489 return -EIO;
1490 }
1491
1492 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1493 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1494 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1495
1496 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1497
1498 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1499 if (tx->tx_lntmsg[1] == NULL) {
1500 CERROR("Can't create reply for GET -> %s\n",
1501 libcfs_nid2str(target.nid));
1502 kiblnd_tx_done(ni, tx);
1503 return -EIO;
1504 }
1505
1506 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */
1507 tx->tx_waiting = 1; /* waiting for GET_DONE */
1508 kiblnd_launch_tx(ni, tx, target.nid);
1509 return 0;
1510
1511 case LNET_MSG_REPLY:
1512 case LNET_MSG_PUT:
1513 /* Is the payload small enough not to need RDMA? */
1514 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1515 if (nob <= IBLND_MSG_SIZE)
1516 break; /* send IMMEDIATE */
1517
1518 tx = kiblnd_get_idle_tx(ni, target.nid);
1519 if (tx == NULL) {
1520 CERROR("Can't allocate %s txd for %s\n",
1521 type == LNET_MSG_PUT ? "PUT" : "REPLY",
1522 libcfs_nid2str(target.nid));
1523 return -ENOMEM;
1524 }
1525
1526 if (payload_kiov == NULL)
1527 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1528 payload_niov, payload_iov,
1529 payload_offset, payload_nob);
1530 else
1531 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1532 payload_niov, payload_kiov,
1533 payload_offset, payload_nob);
1534 if (rc != 0) {
1535 CERROR("Can't setup PUT src for %s: %d\n",
1536 libcfs_nid2str(target.nid), rc);
1537 kiblnd_tx_done(ni, tx);
1538 return -EIO;
1539 }
1540
1541 ibmsg = tx->tx_msg;
1542 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1543 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1544 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1545
1546 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1547 tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */
1548 kiblnd_launch_tx(ni, tx, target.nid);
1549 return 0;
1550 }
1551
1552 /* send IMMEDIATE */
1553
1554 LASSERT(offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1555 <= IBLND_MSG_SIZE);
1556
1557 tx = kiblnd_get_idle_tx(ni, target.nid);
1558 if (tx == NULL) {
1559 CERROR("Can't send %d to %s: tx descs exhausted\n",
1560 type, libcfs_nid2str(target.nid));
1561 return -ENOMEM;
1562 }
1563
1564 ibmsg = tx->tx_msg;
1565 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1566
1567 if (payload_kiov != NULL)
1568 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1569 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1570 payload_niov, payload_kiov,
1571 payload_offset, payload_nob);
1572 else
1573 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1574 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1575 payload_niov, payload_iov,
1576 payload_offset, payload_nob);
1577
1578 nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1579 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1580
1581 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1582 kiblnd_launch_tx(ni, tx, target.nid);
1583 return 0;
1584 }
1585
1586 static void
kiblnd_reply(lnet_ni_t * ni,kib_rx_t * rx,lnet_msg_t * lntmsg)1587 kiblnd_reply(lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1588 {
1589 lnet_process_id_t target = lntmsg->msg_target;
1590 unsigned int niov = lntmsg->msg_niov;
1591 struct kvec *iov = lntmsg->msg_iov;
1592 lnet_kiov_t *kiov = lntmsg->msg_kiov;
1593 unsigned int offset = lntmsg->msg_offset;
1594 unsigned int nob = lntmsg->msg_len;
1595 kib_tx_t *tx;
1596 int rc;
1597
1598 tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1599 if (tx == NULL) {
1600 CERROR("Can't get tx for REPLY to %s\n",
1601 libcfs_nid2str(target.nid));
1602 goto failed_0;
1603 }
1604
1605 if (nob == 0)
1606 rc = 0;
1607 else if (kiov == NULL)
1608 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1609 niov, iov, offset, nob);
1610 else
1611 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1612 niov, kiov, offset, nob);
1613
1614 if (rc != 0) {
1615 CERROR("Can't setup GET src for %s: %d\n",
1616 libcfs_nid2str(target.nid), rc);
1617 goto failed_1;
1618 }
1619
1620 rc = kiblnd_init_rdma(rx->rx_conn, tx,
1621 IBLND_MSG_GET_DONE, nob,
1622 &rx->rx_msg->ibm_u.get.ibgm_rd,
1623 rx->rx_msg->ibm_u.get.ibgm_cookie);
1624 if (rc < 0) {
1625 CERROR("Can't setup rdma for GET from %s: %d\n",
1626 libcfs_nid2str(target.nid), rc);
1627 goto failed_1;
1628 }
1629
1630 if (nob == 0) {
1631 /* No RDMA: local completion may happen now! */
1632 lnet_finalize(ni, lntmsg, 0);
1633 } else {
1634 /* RDMA: lnet_finalize(lntmsg) when it
1635 * completes */
1636 tx->tx_lntmsg[0] = lntmsg;
1637 }
1638
1639 kiblnd_queue_tx(tx, rx->rx_conn);
1640 return;
1641
1642 failed_1:
1643 kiblnd_tx_done(ni, tx);
1644 failed_0:
1645 lnet_finalize(ni, lntmsg, -EIO);
1646 }
1647
1648 int
kiblnd_recv(lnet_ni_t * ni,void * private,lnet_msg_t * lntmsg,int delayed,unsigned int niov,struct kvec * iov,lnet_kiov_t * kiov,unsigned int offset,unsigned int mlen,unsigned int rlen)1649 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1650 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1651 unsigned int offset, unsigned int mlen, unsigned int rlen)
1652 {
1653 kib_rx_t *rx = private;
1654 kib_msg_t *rxmsg = rx->rx_msg;
1655 kib_conn_t *conn = rx->rx_conn;
1656 kib_tx_t *tx;
1657 int nob;
1658 int post_credit = IBLND_POSTRX_PEER_CREDIT;
1659 int rc = 0;
1660
1661 LASSERT(mlen <= rlen);
1662 LASSERT(!in_interrupt());
1663 /* Either all pages or all vaddrs */
1664 LASSERT(!(kiov != NULL && iov != NULL));
1665
1666 switch (rxmsg->ibm_type) {
1667 default:
1668 LBUG();
1669
1670 case IBLND_MSG_IMMEDIATE:
1671 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1672 if (nob > rx->rx_nob) {
1673 CERROR("Immediate message from %s too big: %d(%d)\n",
1674 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1675 nob, rx->rx_nob);
1676 rc = -EPROTO;
1677 break;
1678 }
1679
1680 if (kiov != NULL)
1681 lnet_copy_flat2kiov(niov, kiov, offset,
1682 IBLND_MSG_SIZE, rxmsg,
1683 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1684 mlen);
1685 else
1686 lnet_copy_flat2iov(niov, iov, offset,
1687 IBLND_MSG_SIZE, rxmsg,
1688 offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1689 mlen);
1690 lnet_finalize(ni, lntmsg, 0);
1691 break;
1692
1693 case IBLND_MSG_PUT_REQ: {
1694 kib_msg_t *txmsg;
1695 kib_rdma_desc_t *rd;
1696
1697 if (mlen == 0) {
1698 lnet_finalize(ni, lntmsg, 0);
1699 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1700 rxmsg->ibm_u.putreq.ibprm_cookie);
1701 break;
1702 }
1703
1704 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1705 if (tx == NULL) {
1706 CERROR("Can't allocate tx for %s\n",
1707 libcfs_nid2str(conn->ibc_peer->ibp_nid));
1708 /* Not replying will break the connection */
1709 rc = -ENOMEM;
1710 break;
1711 }
1712
1713 txmsg = tx->tx_msg;
1714 rd = &txmsg->ibm_u.putack.ibpam_rd;
1715 if (kiov == NULL)
1716 rc = kiblnd_setup_rd_iov(ni, tx, rd,
1717 niov, iov, offset, mlen);
1718 else
1719 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1720 niov, kiov, offset, mlen);
1721 if (rc != 0) {
1722 CERROR("Can't setup PUT sink for %s: %d\n",
1723 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1724 kiblnd_tx_done(ni, tx);
1725 /* tell peer it's over */
1726 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1727 rxmsg->ibm_u.putreq.ibprm_cookie);
1728 break;
1729 }
1730
1731 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1732 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1733 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1734
1735 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1736
1737 tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */
1738 tx->tx_waiting = 1; /* waiting for PUT_DONE */
1739 kiblnd_queue_tx(tx, conn);
1740
1741 /* reposted buffer reserved for PUT_DONE */
1742 post_credit = IBLND_POSTRX_NO_CREDIT;
1743 break;
1744 }
1745
1746 case IBLND_MSG_GET_REQ:
1747 if (lntmsg != NULL) {
1748 /* Optimized GET; RDMA lntmsg's payload */
1749 kiblnd_reply(ni, rx, lntmsg);
1750 } else {
1751 /* GET didn't match anything */
1752 kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1753 -ENODATA,
1754 rxmsg->ibm_u.get.ibgm_cookie);
1755 }
1756 break;
1757 }
1758
1759 kiblnd_post_rx(rx, post_credit);
1760 return rc;
1761 }
1762
1763 int
kiblnd_thread_start(int (* fn)(void * arg),void * arg,char * name)1764 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1765 {
1766 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1767
1768 if (IS_ERR(task))
1769 return PTR_ERR(task);
1770
1771 atomic_inc(&kiblnd_data.kib_nthreads);
1772 return 0;
1773 }
1774
1775 static void
kiblnd_thread_fini(void)1776 kiblnd_thread_fini(void)
1777 {
1778 atomic_dec(&kiblnd_data.kib_nthreads);
1779 }
1780
1781 void
kiblnd_peer_alive(kib_peer_t * peer)1782 kiblnd_peer_alive(kib_peer_t *peer)
1783 {
1784 /* This is racy, but everyone's only writing cfs_time_current() */
1785 peer->ibp_last_alive = cfs_time_current();
1786 mb();
1787 }
1788
1789 static void
kiblnd_peer_notify(kib_peer_t * peer)1790 kiblnd_peer_notify(kib_peer_t *peer)
1791 {
1792 int error = 0;
1793 unsigned long last_alive = 0;
1794 unsigned long flags;
1795
1796 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1797
1798 if (list_empty(&peer->ibp_conns) &&
1799 peer->ibp_accepting == 0 &&
1800 peer->ibp_connecting == 0 &&
1801 peer->ibp_error != 0) {
1802 error = peer->ibp_error;
1803 peer->ibp_error = 0;
1804
1805 last_alive = peer->ibp_last_alive;
1806 }
1807
1808 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1809
1810 if (error != 0)
1811 lnet_notify(peer->ibp_ni,
1812 peer->ibp_nid, 0, last_alive);
1813 }
1814
1815 void
kiblnd_close_conn_locked(kib_conn_t * conn,int error)1816 kiblnd_close_conn_locked(kib_conn_t *conn, int error)
1817 {
1818 /* This just does the immediate housekeeping. 'error' is zero for a
1819 * normal shutdown which can happen only after the connection has been
1820 * established. If the connection is established, schedule the
1821 * connection to be finished off by the connd. Otherwise the connd is
1822 * already dealing with it (either to set it up or tear it down).
1823 * Caller holds kib_global_lock exclusively in irq context */
1824 kib_peer_t *peer = conn->ibc_peer;
1825 kib_dev_t *dev;
1826 unsigned long flags;
1827
1828 LASSERT(error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1829
1830 if (error != 0 && conn->ibc_comms_error == 0)
1831 conn->ibc_comms_error = error;
1832
1833 if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1834 return; /* already being handled */
1835
1836 if (error == 0 &&
1837 list_empty(&conn->ibc_tx_noops) &&
1838 list_empty(&conn->ibc_tx_queue) &&
1839 list_empty(&conn->ibc_tx_queue_rsrvd) &&
1840 list_empty(&conn->ibc_tx_queue_nocred) &&
1841 list_empty(&conn->ibc_active_txs)) {
1842 CDEBUG(D_NET, "closing conn to %s\n",
1843 libcfs_nid2str(peer->ibp_nid));
1844 } else {
1845 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1846 libcfs_nid2str(peer->ibp_nid), error,
1847 list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1848 list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1849 list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1850 list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1851 list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1852 }
1853
1854 dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1855 list_del(&conn->ibc_list);
1856 /* connd (see below) takes over ibc_list's ref */
1857
1858 if (list_empty(&peer->ibp_conns) && /* no more conns */
1859 kiblnd_peer_active(peer)) { /* still in peer table */
1860 kiblnd_unlink_peer_locked(peer);
1861
1862 /* set/clear error on last conn */
1863 peer->ibp_error = conn->ibc_comms_error;
1864 }
1865
1866 kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1867
1868 if (error != 0 &&
1869 kiblnd_dev_can_failover(dev)) {
1870 list_add_tail(&dev->ibd_fail_list,
1871 &kiblnd_data.kib_failed_devs);
1872 wake_up(&kiblnd_data.kib_failover_waitq);
1873 }
1874
1875 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1876
1877 list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1878 wake_up(&kiblnd_data.kib_connd_waitq);
1879
1880 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1881 }
1882
1883 void
kiblnd_close_conn(kib_conn_t * conn,int error)1884 kiblnd_close_conn(kib_conn_t *conn, int error)
1885 {
1886 unsigned long flags;
1887
1888 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1889
1890 kiblnd_close_conn_locked(conn, error);
1891
1892 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1893 }
1894
1895 static void
kiblnd_handle_early_rxs(kib_conn_t * conn)1896 kiblnd_handle_early_rxs(kib_conn_t *conn)
1897 {
1898 unsigned long flags;
1899 kib_rx_t *rx;
1900 kib_rx_t *tmp;
1901
1902 LASSERT(!in_interrupt());
1903 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1904
1905 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1906 list_for_each_entry_safe(rx, tmp, &conn->ibc_early_rxs, rx_list) {
1907 list_del(&rx->rx_list);
1908 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1909
1910 kiblnd_handle_rx(rx);
1911
1912 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1913 }
1914 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1915 }
1916
1917 static void
kiblnd_abort_txs(kib_conn_t * conn,struct list_head * txs)1918 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1919 {
1920 LIST_HEAD(zombies);
1921 struct list_head *tmp;
1922 struct list_head *nxt;
1923 kib_tx_t *tx;
1924
1925 spin_lock(&conn->ibc_lock);
1926
1927 list_for_each_safe(tmp, nxt, txs) {
1928 tx = list_entry(tmp, kib_tx_t, tx_list);
1929
1930 if (txs == &conn->ibc_active_txs) {
1931 LASSERT(!tx->tx_queued);
1932 LASSERT(tx->tx_waiting ||
1933 tx->tx_sending != 0);
1934 } else {
1935 LASSERT(tx->tx_queued);
1936 }
1937
1938 tx->tx_status = -ECONNABORTED;
1939 tx->tx_waiting = 0;
1940
1941 if (tx->tx_sending == 0) {
1942 tx->tx_queued = 0;
1943 list_del(&tx->tx_list);
1944 list_add(&tx->tx_list, &zombies);
1945 }
1946 }
1947
1948 spin_unlock(&conn->ibc_lock);
1949
1950 kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1951 }
1952
1953 static void
kiblnd_finalise_conn(kib_conn_t * conn)1954 kiblnd_finalise_conn(kib_conn_t *conn)
1955 {
1956 LASSERT(!in_interrupt());
1957 LASSERT(conn->ibc_state > IBLND_CONN_INIT);
1958
1959 kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1960
1961 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
1962 * for connections that didn't get as far as being connected, because
1963 * rdma_disconnect() does this for free. */
1964 kiblnd_abort_receives(conn);
1965
1966 /* Complete all tx descs not waiting for sends to complete.
1967 * NB we should be safe from RDMA now that the QP has changed state */
1968
1969 kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
1970 kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1971 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1972 kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1973 kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1974
1975 kiblnd_handle_early_rxs(conn);
1976 }
1977
1978 void
kiblnd_peer_connect_failed(kib_peer_t * peer,int active,int error)1979 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
1980 {
1981 LIST_HEAD(zombies);
1982 unsigned long flags;
1983
1984 LASSERT(error != 0);
1985 LASSERT(!in_interrupt());
1986
1987 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1988
1989 if (active) {
1990 LASSERT(peer->ibp_connecting > 0);
1991 peer->ibp_connecting--;
1992 } else {
1993 LASSERT(peer->ibp_accepting > 0);
1994 peer->ibp_accepting--;
1995 }
1996
1997 if (peer->ibp_connecting != 0 ||
1998 peer->ibp_accepting != 0) {
1999 /* another connection attempt under way... */
2000 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2001 flags);
2002 return;
2003 }
2004
2005 if (list_empty(&peer->ibp_conns)) {
2006 /* Take peer's blocked transmits to complete with error */
2007 list_add(&zombies, &peer->ibp_tx_queue);
2008 list_del_init(&peer->ibp_tx_queue);
2009
2010 if (kiblnd_peer_active(peer))
2011 kiblnd_unlink_peer_locked(peer);
2012
2013 peer->ibp_error = error;
2014 } else {
2015 /* Can't have blocked transmits if there are connections */
2016 LASSERT(list_empty(&peer->ibp_tx_queue));
2017 }
2018
2019 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2020
2021 kiblnd_peer_notify(peer);
2022
2023 if (list_empty(&zombies))
2024 return;
2025
2026 CNETERR("Deleting messages for %s: connection failed\n",
2027 libcfs_nid2str(peer->ibp_nid));
2028
2029 kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2030 }
2031
2032 void
kiblnd_connreq_done(kib_conn_t * conn,int status)2033 kiblnd_connreq_done(kib_conn_t *conn, int status)
2034 {
2035 kib_peer_t *peer = conn->ibc_peer;
2036 kib_tx_t *tx;
2037 kib_tx_t *tmp;
2038 struct list_head txs;
2039 unsigned long flags;
2040 int active;
2041
2042 active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2043
2044 CDEBUG(D_NET, "%s: active(%d), version(%x), status(%d)\n",
2045 libcfs_nid2str(peer->ibp_nid), active,
2046 conn->ibc_version, status);
2047
2048 LASSERT(!in_interrupt());
2049 LASSERT((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2050 peer->ibp_connecting > 0) ||
2051 (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2052 peer->ibp_accepting > 0));
2053
2054 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2055 conn->ibc_connvars = NULL;
2056
2057 if (status != 0) {
2058 /* failed to establish connection */
2059 kiblnd_peer_connect_failed(peer, active, status);
2060 kiblnd_finalise_conn(conn);
2061 return;
2062 }
2063
2064 /* connection established */
2065 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2066
2067 conn->ibc_last_send = jiffies;
2068 kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2069 kiblnd_peer_alive(peer);
2070
2071 /* Add conn to peer's list and nuke any dangling conns from a different
2072 * peer instance... */
2073 kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
2074 list_add(&conn->ibc_list, &peer->ibp_conns);
2075 if (active)
2076 peer->ibp_connecting--;
2077 else
2078 peer->ibp_accepting--;
2079
2080 if (peer->ibp_version == 0) {
2081 peer->ibp_version = conn->ibc_version;
2082 peer->ibp_incarnation = conn->ibc_incarnation;
2083 }
2084
2085 if (peer->ibp_version != conn->ibc_version ||
2086 peer->ibp_incarnation != conn->ibc_incarnation) {
2087 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2088 conn->ibc_incarnation);
2089 peer->ibp_version = conn->ibc_version;
2090 peer->ibp_incarnation = conn->ibc_incarnation;
2091 }
2092
2093 /* grab pending txs while I have the lock */
2094 list_add(&txs, &peer->ibp_tx_queue);
2095 list_del_init(&peer->ibp_tx_queue);
2096
2097 if (!kiblnd_peer_active(peer) || /* peer has been deleted */
2098 conn->ibc_comms_error != 0) { /* error has happened already */
2099 lnet_ni_t *ni = peer->ibp_ni;
2100
2101 /* start to shut down connection */
2102 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2103 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2104
2105 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2106
2107 return;
2108 }
2109
2110 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2111
2112 /* Schedule blocked txs */
2113 spin_lock(&conn->ibc_lock);
2114 list_for_each_entry_safe(tx, tmp, &txs, tx_list) {
2115 list_del(&tx->tx_list);
2116
2117 kiblnd_queue_tx_locked(tx, conn);
2118 }
2119 spin_unlock(&conn->ibc_lock);
2120
2121 kiblnd_check_sends(conn);
2122
2123 /* schedule blocked rxs */
2124 kiblnd_handle_early_rxs(conn);
2125 }
2126
2127 static void
kiblnd_reject(struct rdma_cm_id * cmid,kib_rej_t * rej)2128 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2129 {
2130 int rc;
2131
2132 rc = rdma_reject(cmid, rej, sizeof(*rej));
2133
2134 if (rc != 0)
2135 CWARN("Error %d sending reject\n", rc);
2136 }
2137
2138 static int
kiblnd_passive_connect(struct rdma_cm_id * cmid,void * priv,int priv_nob)2139 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2140 {
2141 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2142 kib_msg_t *reqmsg = priv;
2143 kib_msg_t *ackmsg;
2144 kib_dev_t *ibdev;
2145 kib_peer_t *peer;
2146 kib_peer_t *peer2;
2147 kib_conn_t *conn;
2148 lnet_ni_t *ni = NULL;
2149 kib_net_t *net = NULL;
2150 lnet_nid_t nid;
2151 struct rdma_conn_param cp;
2152 kib_rej_t rej;
2153 int version = IBLND_MSG_VERSION;
2154 unsigned long flags;
2155 int rc;
2156 struct sockaddr_in *peer_addr;
2157
2158 LASSERT(!in_interrupt());
2159
2160 /* cmid inherits 'context' from the corresponding listener id */
2161 ibdev = (kib_dev_t *)cmid->context;
2162 LASSERT(ibdev != NULL);
2163
2164 memset(&rej, 0, sizeof(rej));
2165 rej.ibr_magic = IBLND_MSG_MAGIC;
2166 rej.ibr_why = IBLND_REJECT_FATAL;
2167 rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2168
2169 peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2170 if (*kiblnd_tunables.kib_require_priv_port &&
2171 ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2172 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2173
2174 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2175 &ip, ntohs(peer_addr->sin_port));
2176 goto failed;
2177 }
2178
2179 if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2180 CERROR("Short connection request\n");
2181 goto failed;
2182 }
2183
2184 /* Future protocol version compatibility support! If the
2185 * o2iblnd-specific protocol changes, or when LNET unifies
2186 * protocols over all LNDs, the initial connection will
2187 * negotiate a protocol version. I trap this here to avoid
2188 * console errors; the reject tells the peer which protocol I
2189 * speak. */
2190 if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2191 reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2192 goto failed;
2193 if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2194 reqmsg->ibm_version != IBLND_MSG_VERSION &&
2195 reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2196 goto failed;
2197 if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2198 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2199 reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2200 goto failed;
2201
2202 rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2203 if (rc != 0) {
2204 CERROR("Can't parse connection request: %d\n", rc);
2205 goto failed;
2206 }
2207
2208 nid = reqmsg->ibm_srcnid;
2209 ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2210
2211 if (ni != NULL) {
2212 net = (kib_net_t *)ni->ni_data;
2213 rej.ibr_incarnation = net->ibn_incarnation;
2214 }
2215
2216 if (ni == NULL || /* no matching net */
2217 ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
2218 net->ibn_dev != ibdev) { /* wrong device */
2219 CERROR("Can't accept %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2220 libcfs_nid2str(nid),
2221 ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2222 ibdev->ibd_ifname, ibdev->ibd_nnets,
2223 &ibdev->ibd_ifip,
2224 libcfs_nid2str(reqmsg->ibm_dstnid));
2225
2226 goto failed;
2227 }
2228
2229 /* check time stamp as soon as possible */
2230 if (reqmsg->ibm_dststamp != 0 &&
2231 reqmsg->ibm_dststamp != net->ibn_incarnation) {
2232 CWARN("Stale connection request\n");
2233 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2234 goto failed;
2235 }
2236
2237 /* I can accept peer's version */
2238 version = reqmsg->ibm_version;
2239
2240 if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2241 CERROR("Unexpected connreq msg type: %x from %s\n",
2242 reqmsg->ibm_type, libcfs_nid2str(nid));
2243 goto failed;
2244 }
2245
2246 if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2247 IBLND_MSG_QUEUE_SIZE(version)) {
2248 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2249 libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2250 IBLND_MSG_QUEUE_SIZE(version));
2251
2252 if (version == IBLND_MSG_VERSION)
2253 rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2254
2255 goto failed;
2256 }
2257
2258 if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2259 IBLND_RDMA_FRAGS(version)) {
2260 CERROR("Can't accept %s(version %x): incompatible max_frags %d (%d wanted)\n",
2261 libcfs_nid2str(nid), version,
2262 reqmsg->ibm_u.connparams.ibcp_max_frags,
2263 IBLND_RDMA_FRAGS(version));
2264
2265 if (version == IBLND_MSG_VERSION)
2266 rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2267
2268 goto failed;
2269
2270 }
2271
2272 if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2273 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2274 libcfs_nid2str(nid),
2275 reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2276 IBLND_MSG_SIZE);
2277 goto failed;
2278 }
2279
2280 /* assume 'nid' is a new peer; create */
2281 rc = kiblnd_create_peer(ni, &peer, nid);
2282 if (rc != 0) {
2283 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2284 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2285 goto failed;
2286 }
2287
2288 write_lock_irqsave(g_lock, flags);
2289
2290 peer2 = kiblnd_find_peer_locked(nid);
2291 if (peer2 != NULL) {
2292 if (peer2->ibp_version == 0) {
2293 peer2->ibp_version = version;
2294 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2295 }
2296
2297 /* not the guy I've talked with */
2298 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2299 peer2->ibp_version != version) {
2300 kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2301 write_unlock_irqrestore(g_lock, flags);
2302
2303 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2304 libcfs_nid2str(nid), peer2->ibp_version, version);
2305
2306 kiblnd_peer_decref(peer);
2307 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2308 goto failed;
2309 }
2310
2311 /* tie-break connection race in favour of the higher NID */
2312 if (peer2->ibp_connecting != 0 &&
2313 nid < ni->ni_nid) {
2314 write_unlock_irqrestore(g_lock, flags);
2315
2316 CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2317
2318 kiblnd_peer_decref(peer);
2319 rej.ibr_why = IBLND_REJECT_CONN_RACE;
2320 goto failed;
2321 }
2322
2323 peer2->ibp_accepting++;
2324 kiblnd_peer_addref(peer2);
2325
2326 write_unlock_irqrestore(g_lock, flags);
2327 kiblnd_peer_decref(peer);
2328 peer = peer2;
2329 } else {
2330 /* Brand new peer */
2331 LASSERT(peer->ibp_accepting == 0);
2332 LASSERT(peer->ibp_version == 0 &&
2333 peer->ibp_incarnation == 0);
2334
2335 peer->ibp_accepting = 1;
2336 peer->ibp_version = version;
2337 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2338
2339 /* I have a ref on ni that prevents it being shutdown */
2340 LASSERT(net->ibn_shutdown == 0);
2341
2342 kiblnd_peer_addref(peer);
2343 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2344
2345 write_unlock_irqrestore(g_lock, flags);
2346 }
2347
2348 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2349 if (conn == NULL) {
2350 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2351 kiblnd_peer_decref(peer);
2352 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2353 goto failed;
2354 }
2355
2356 /* conn now "owns" cmid, so I return success from here on to ensure the
2357 * CM callback doesn't destroy cmid. */
2358
2359 conn->ibc_incarnation = reqmsg->ibm_srcstamp;
2360 conn->ibc_credits = IBLND_MSG_QUEUE_SIZE(version);
2361 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2362 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2363 <= IBLND_RX_MSGS(version));
2364
2365 ackmsg = &conn->ibc_connvars->cv_msg;
2366 memset(ackmsg, 0, sizeof(*ackmsg));
2367
2368 kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2369 sizeof(ackmsg->ibm_u.connparams));
2370 ackmsg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2371 ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2372 ackmsg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2373
2374 kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2375
2376 memset(&cp, 0, sizeof(cp));
2377 cp.private_data = ackmsg;
2378 cp.private_data_len = ackmsg->ibm_nob;
2379 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2380 cp.initiator_depth = 0;
2381 cp.flow_control = 1;
2382 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2383 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2384
2385 CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2386
2387 rc = rdma_accept(cmid, &cp);
2388 if (rc != 0) {
2389 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2390 rej.ibr_version = version;
2391 rej.ibr_why = IBLND_REJECT_FATAL;
2392
2393 kiblnd_reject(cmid, &rej);
2394 kiblnd_connreq_done(conn, rc);
2395 kiblnd_conn_decref(conn);
2396 }
2397
2398 lnet_ni_decref(ni);
2399 return 0;
2400
2401 failed:
2402 if (ni != NULL)
2403 lnet_ni_decref(ni);
2404
2405 rej.ibr_version = version;
2406 rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2407 rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2408 kiblnd_reject(cmid, &rej);
2409
2410 return -ECONNREFUSED;
2411 }
2412
2413 static void
kiblnd_reconnect(kib_conn_t * conn,int version,__u64 incarnation,int why,kib_connparams_t * cp)2414 kiblnd_reconnect(kib_conn_t *conn, int version,
2415 __u64 incarnation, int why, kib_connparams_t *cp)
2416 {
2417 kib_peer_t *peer = conn->ibc_peer;
2418 char *reason;
2419 int retry = 0;
2420 unsigned long flags;
2421
2422 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2423 LASSERT(peer->ibp_connecting > 0); /* 'conn' at least */
2424
2425 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2426
2427 /* retry connection if it's still needed and no other connection
2428 * attempts (active or passive) are in progress
2429 * NB: reconnect is still needed even when ibp_tx_queue is
2430 * empty if ibp_version != version because reconnect may be
2431 * initiated by kiblnd_query() */
2432 if ((!list_empty(&peer->ibp_tx_queue) ||
2433 peer->ibp_version != version) &&
2434 peer->ibp_connecting == 1 &&
2435 peer->ibp_accepting == 0) {
2436 retry = 1;
2437 peer->ibp_connecting++;
2438
2439 peer->ibp_version = version;
2440 peer->ibp_incarnation = incarnation;
2441 }
2442
2443 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2444
2445 if (!retry)
2446 return;
2447
2448 switch (why) {
2449 default:
2450 reason = "Unknown";
2451 break;
2452
2453 case IBLND_REJECT_CONN_STALE:
2454 reason = "stale";
2455 break;
2456
2457 case IBLND_REJECT_CONN_RACE:
2458 reason = "conn race";
2459 break;
2460
2461 case IBLND_REJECT_CONN_UNCOMPAT:
2462 reason = "version negotiation";
2463 break;
2464 }
2465
2466 CNETERR("%s: retrying (%s), %x, %x, queue_dep: %d, max_frag: %d, msg_size: %d\n",
2467 libcfs_nid2str(peer->ibp_nid),
2468 reason, IBLND_MSG_VERSION, version,
2469 cp != NULL ? cp->ibcp_queue_depth : IBLND_MSG_QUEUE_SIZE(version),
2470 cp != NULL ? cp->ibcp_max_frags : IBLND_RDMA_FRAGS(version),
2471 cp != NULL ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
2472
2473 kiblnd_connect_peer(peer);
2474 }
2475
2476 static void
kiblnd_rejected(kib_conn_t * conn,int reason,void * priv,int priv_nob)2477 kiblnd_rejected(kib_conn_t *conn, int reason, void *priv, int priv_nob)
2478 {
2479 kib_peer_t *peer = conn->ibc_peer;
2480
2481 LASSERT(!in_interrupt());
2482 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2483
2484 switch (reason) {
2485 case IB_CM_REJ_STALE_CONN:
2486 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2487 IBLND_REJECT_CONN_STALE, NULL);
2488 break;
2489
2490 case IB_CM_REJ_INVALID_SERVICE_ID:
2491 CNETERR("%s rejected: no listener at %d\n",
2492 libcfs_nid2str(peer->ibp_nid),
2493 *kiblnd_tunables.kib_service);
2494 break;
2495
2496 case IB_CM_REJ_CONSUMER_DEFINED:
2497 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2498 kib_rej_t *rej = priv;
2499 kib_connparams_t *cp = NULL;
2500 int flip = 0;
2501 __u64 incarnation = -1;
2502
2503 /* NB. default incarnation is -1 because:
2504 * a) V1 will ignore dst incarnation in connreq.
2505 * b) V2 will provide incarnation while rejecting me,
2506 * -1 will be overwrote.
2507 *
2508 * if I try to connect to a V1 peer with V2 protocol,
2509 * it rejected me then upgrade to V2, I have no idea
2510 * about the upgrading and try to reconnect with V1,
2511 * in this case upgraded V2 can find out I'm trying to
2512 * talk to the old guy and reject me(incarnation is -1).
2513 */
2514
2515 if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2516 rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2517 __swab32s(&rej->ibr_magic);
2518 __swab16s(&rej->ibr_version);
2519 flip = 1;
2520 }
2521
2522 if (priv_nob >= sizeof(kib_rej_t) &&
2523 rej->ibr_version > IBLND_MSG_VERSION_1) {
2524 /* priv_nob is always 148 in current version
2525 * of OFED, so we still need to check version.
2526 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2527 cp = &rej->ibr_cp;
2528
2529 if (flip) {
2530 __swab64s(&rej->ibr_incarnation);
2531 __swab16s(&cp->ibcp_queue_depth);
2532 __swab16s(&cp->ibcp_max_frags);
2533 __swab32s(&cp->ibcp_max_msg_size);
2534 }
2535
2536 incarnation = rej->ibr_incarnation;
2537 }
2538
2539 if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2540 rej->ibr_magic != LNET_PROTO_MAGIC) {
2541 CERROR("%s rejected: consumer defined fatal error\n",
2542 libcfs_nid2str(peer->ibp_nid));
2543 break;
2544 }
2545
2546 if (rej->ibr_version != IBLND_MSG_VERSION &&
2547 rej->ibr_version != IBLND_MSG_VERSION_1) {
2548 CERROR("%s rejected: o2iblnd version %x error\n",
2549 libcfs_nid2str(peer->ibp_nid),
2550 rej->ibr_version);
2551 break;
2552 }
2553
2554 if (rej->ibr_why == IBLND_REJECT_FATAL &&
2555 rej->ibr_version == IBLND_MSG_VERSION_1) {
2556 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2557 libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2558
2559 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2560 rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2561 }
2562
2563 switch (rej->ibr_why) {
2564 case IBLND_REJECT_CONN_RACE:
2565 case IBLND_REJECT_CONN_STALE:
2566 case IBLND_REJECT_CONN_UNCOMPAT:
2567 kiblnd_reconnect(conn, rej->ibr_version,
2568 incarnation, rej->ibr_why, cp);
2569 break;
2570
2571 case IBLND_REJECT_MSG_QUEUE_SIZE:
2572 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2573 libcfs_nid2str(peer->ibp_nid),
2574 cp != NULL ? cp->ibcp_queue_depth :
2575 IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2576 IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2577 break;
2578
2579 case IBLND_REJECT_RDMA_FRAGS:
2580 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2581 libcfs_nid2str(peer->ibp_nid),
2582 cp != NULL ? cp->ibcp_max_frags :
2583 IBLND_RDMA_FRAGS(rej->ibr_version),
2584 IBLND_RDMA_FRAGS(conn->ibc_version));
2585 break;
2586
2587 case IBLND_REJECT_NO_RESOURCES:
2588 CERROR("%s rejected: o2iblnd no resources\n",
2589 libcfs_nid2str(peer->ibp_nid));
2590 break;
2591
2592 case IBLND_REJECT_FATAL:
2593 CERROR("%s rejected: o2iblnd fatal error\n",
2594 libcfs_nid2str(peer->ibp_nid));
2595 break;
2596
2597 default:
2598 CERROR("%s rejected: o2iblnd reason %d\n",
2599 libcfs_nid2str(peer->ibp_nid),
2600 rej->ibr_why);
2601 break;
2602 }
2603 break;
2604 }
2605 /* fall through */
2606 default:
2607 CNETERR("%s rejected: reason %d, size %d\n",
2608 libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2609 break;
2610 }
2611
2612 kiblnd_connreq_done(conn, -ECONNREFUSED);
2613 }
2614
2615 static void
kiblnd_check_connreply(kib_conn_t * conn,void * priv,int priv_nob)2616 kiblnd_check_connreply(kib_conn_t *conn, void *priv, int priv_nob)
2617 {
2618 kib_peer_t *peer = conn->ibc_peer;
2619 lnet_ni_t *ni = peer->ibp_ni;
2620 kib_net_t *net = ni->ni_data;
2621 kib_msg_t *msg = priv;
2622 int ver = conn->ibc_version;
2623 int rc = kiblnd_unpack_msg(msg, priv_nob);
2624 unsigned long flags;
2625
2626 LASSERT(net != NULL);
2627
2628 if (rc != 0) {
2629 CERROR("Can't unpack connack from %s: %d\n",
2630 libcfs_nid2str(peer->ibp_nid), rc);
2631 goto failed;
2632 }
2633
2634 if (msg->ibm_type != IBLND_MSG_CONNACK) {
2635 CERROR("Unexpected message %d from %s\n",
2636 msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2637 rc = -EPROTO;
2638 goto failed;
2639 }
2640
2641 if (ver != msg->ibm_version) {
2642 CERROR("%s replied version %x is different with requested version %x\n",
2643 libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2644 rc = -EPROTO;
2645 goto failed;
2646 }
2647
2648 if (msg->ibm_u.connparams.ibcp_queue_depth !=
2649 IBLND_MSG_QUEUE_SIZE(ver)) {
2650 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2651 libcfs_nid2str(peer->ibp_nid),
2652 msg->ibm_u.connparams.ibcp_queue_depth,
2653 IBLND_MSG_QUEUE_SIZE(ver));
2654 rc = -EPROTO;
2655 goto failed;
2656 }
2657
2658 if (msg->ibm_u.connparams.ibcp_max_frags !=
2659 IBLND_RDMA_FRAGS(ver)) {
2660 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2661 libcfs_nid2str(peer->ibp_nid),
2662 msg->ibm_u.connparams.ibcp_max_frags,
2663 IBLND_RDMA_FRAGS(ver));
2664 rc = -EPROTO;
2665 goto failed;
2666 }
2667
2668 if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2669 CERROR("%s max message size %d too big (%d max)\n",
2670 libcfs_nid2str(peer->ibp_nid),
2671 msg->ibm_u.connparams.ibcp_max_msg_size,
2672 IBLND_MSG_SIZE);
2673 rc = -EPROTO;
2674 goto failed;
2675 }
2676
2677 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2678 if (msg->ibm_dstnid == ni->ni_nid &&
2679 msg->ibm_dststamp == net->ibn_incarnation)
2680 rc = 0;
2681 else
2682 rc = -ESTALE;
2683 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2684
2685 if (rc != 0) {
2686 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2687 libcfs_nid2str(peer->ibp_nid), rc,
2688 msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2689 goto failed;
2690 }
2691
2692 conn->ibc_incarnation = msg->ibm_srcstamp;
2693 conn->ibc_credits =
2694 conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2695 LASSERT(conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2696 <= IBLND_RX_MSGS(ver));
2697
2698 kiblnd_connreq_done(conn, 0);
2699 return;
2700
2701 failed:
2702 /* NB My QP has already established itself, so I handle anything going
2703 * wrong here by setting ibc_comms_error.
2704 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2705 * immediately tears it down. */
2706
2707 LASSERT(rc != 0);
2708 conn->ibc_comms_error = rc;
2709 kiblnd_connreq_done(conn, 0);
2710 }
2711
2712 static int
kiblnd_active_connect(struct rdma_cm_id * cmid)2713 kiblnd_active_connect(struct rdma_cm_id *cmid)
2714 {
2715 kib_peer_t *peer = (kib_peer_t *)cmid->context;
2716 kib_conn_t *conn;
2717 kib_msg_t *msg;
2718 struct rdma_conn_param cp;
2719 int version;
2720 __u64 incarnation;
2721 unsigned long flags;
2722 int rc;
2723
2724 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2725
2726 incarnation = peer->ibp_incarnation;
2727 version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2728 peer->ibp_version;
2729
2730 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2731
2732 conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2733 if (conn == NULL) {
2734 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2735 kiblnd_peer_decref(peer); /* lose cmid's ref */
2736 return -ENOMEM;
2737 }
2738
2739 /* conn "owns" cmid now, so I return success from here on to ensure the
2740 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2741 * on peer */
2742
2743 msg = &conn->ibc_connvars->cv_msg;
2744
2745 memset(msg, 0, sizeof(*msg));
2746 kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2747 msg->ibm_u.connparams.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2748 msg->ibm_u.connparams.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
2749 msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2750
2751 kiblnd_pack_msg(peer->ibp_ni, msg, version,
2752 0, peer->ibp_nid, incarnation);
2753
2754 memset(&cp, 0, sizeof(cp));
2755 cp.private_data = msg;
2756 cp.private_data_len = msg->ibm_nob;
2757 cp.responder_resources = 0; /* No atomic ops or RDMA reads */
2758 cp.initiator_depth = 0;
2759 cp.flow_control = 1;
2760 cp.retry_count = *kiblnd_tunables.kib_retry_count;
2761 cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2762
2763 LASSERT(cmid->context == (void *)conn);
2764 LASSERT(conn->ibc_cmid == cmid);
2765
2766 rc = rdma_connect(cmid, &cp);
2767 if (rc != 0) {
2768 CERROR("Can't connect to %s: %d\n",
2769 libcfs_nid2str(peer->ibp_nid), rc);
2770 kiblnd_connreq_done(conn, rc);
2771 kiblnd_conn_decref(conn);
2772 }
2773
2774 return 0;
2775 }
2776
2777 int
kiblnd_cm_callback(struct rdma_cm_id * cmid,struct rdma_cm_event * event)2778 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2779 {
2780 kib_peer_t *peer;
2781 kib_conn_t *conn;
2782 int rc;
2783
2784 switch (event->event) {
2785 default:
2786 CERROR("Unexpected event: %d, status: %d\n",
2787 event->event, event->status);
2788 LBUG();
2789
2790 case RDMA_CM_EVENT_CONNECT_REQUEST:
2791 /* destroy cmid on failure */
2792 rc = kiblnd_passive_connect(cmid,
2793 (void *)KIBLND_CONN_PARAM(event),
2794 KIBLND_CONN_PARAM_LEN(event));
2795 CDEBUG(D_NET, "connreq: %d\n", rc);
2796 return rc;
2797
2798 case RDMA_CM_EVENT_ADDR_ERROR:
2799 peer = (kib_peer_t *)cmid->context;
2800 CNETERR("%s: ADDR ERROR %d\n",
2801 libcfs_nid2str(peer->ibp_nid), event->status);
2802 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2803 kiblnd_peer_decref(peer);
2804 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2805
2806 case RDMA_CM_EVENT_ADDR_RESOLVED:
2807 peer = (kib_peer_t *)cmid->context;
2808
2809 CDEBUG(D_NET, "%s Addr resolved: %d\n",
2810 libcfs_nid2str(peer->ibp_nid), event->status);
2811
2812 if (event->status != 0) {
2813 CNETERR("Can't resolve address for %s: %d\n",
2814 libcfs_nid2str(peer->ibp_nid), event->status);
2815 rc = event->status;
2816 } else {
2817 rc = rdma_resolve_route(
2818 cmid, *kiblnd_tunables.kib_timeout * 1000);
2819 if (rc == 0)
2820 return 0;
2821 /* Can't initiate route resolution */
2822 CERROR("Can't resolve route for %s: %d\n",
2823 libcfs_nid2str(peer->ibp_nid), rc);
2824 }
2825 kiblnd_peer_connect_failed(peer, 1, rc);
2826 kiblnd_peer_decref(peer);
2827 return rc; /* rc != 0 destroys cmid */
2828
2829 case RDMA_CM_EVENT_ROUTE_ERROR:
2830 peer = (kib_peer_t *)cmid->context;
2831 CNETERR("%s: ROUTE ERROR %d\n",
2832 libcfs_nid2str(peer->ibp_nid), event->status);
2833 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2834 kiblnd_peer_decref(peer);
2835 return -EHOSTUNREACH; /* rc != 0 destroys cmid */
2836
2837 case RDMA_CM_EVENT_ROUTE_RESOLVED:
2838 peer = (kib_peer_t *)cmid->context;
2839 CDEBUG(D_NET, "%s Route resolved: %d\n",
2840 libcfs_nid2str(peer->ibp_nid), event->status);
2841
2842 if (event->status == 0)
2843 return kiblnd_active_connect(cmid);
2844
2845 CNETERR("Can't resolve route for %s: %d\n",
2846 libcfs_nid2str(peer->ibp_nid), event->status);
2847 kiblnd_peer_connect_failed(peer, 1, event->status);
2848 kiblnd_peer_decref(peer);
2849 return event->status; /* rc != 0 destroys cmid */
2850
2851 case RDMA_CM_EVENT_UNREACHABLE:
2852 conn = (kib_conn_t *)cmid->context;
2853 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2854 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2855 CNETERR("%s: UNREACHABLE %d\n",
2856 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2857 kiblnd_connreq_done(conn, -ENETDOWN);
2858 kiblnd_conn_decref(conn);
2859 return 0;
2860
2861 case RDMA_CM_EVENT_CONNECT_ERROR:
2862 conn = (kib_conn_t *)cmid->context;
2863 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2864 conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2865 CNETERR("%s: CONNECT ERROR %d\n",
2866 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2867 kiblnd_connreq_done(conn, -ENOTCONN);
2868 kiblnd_conn_decref(conn);
2869 return 0;
2870
2871 case RDMA_CM_EVENT_REJECTED:
2872 conn = (kib_conn_t *)cmid->context;
2873 switch (conn->ibc_state) {
2874 default:
2875 LBUG();
2876
2877 case IBLND_CONN_PASSIVE_WAIT:
2878 CERROR("%s: REJECTED %d\n",
2879 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2880 event->status);
2881 kiblnd_connreq_done(conn, -ECONNRESET);
2882 break;
2883
2884 case IBLND_CONN_ACTIVE_CONNECT:
2885 kiblnd_rejected(conn, event->status,
2886 (void *)KIBLND_CONN_PARAM(event),
2887 KIBLND_CONN_PARAM_LEN(event));
2888 break;
2889 }
2890 kiblnd_conn_decref(conn);
2891 return 0;
2892
2893 case RDMA_CM_EVENT_ESTABLISHED:
2894 conn = (kib_conn_t *)cmid->context;
2895 switch (conn->ibc_state) {
2896 default:
2897 LBUG();
2898
2899 case IBLND_CONN_PASSIVE_WAIT:
2900 CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2901 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2902 kiblnd_connreq_done(conn, 0);
2903 break;
2904
2905 case IBLND_CONN_ACTIVE_CONNECT:
2906 CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2907 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2908 kiblnd_check_connreply(conn,
2909 (void *)KIBLND_CONN_PARAM(event),
2910 KIBLND_CONN_PARAM_LEN(event));
2911 break;
2912 }
2913 /* net keeps its ref on conn! */
2914 return 0;
2915
2916 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2917 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2918 return 0;
2919 case RDMA_CM_EVENT_DISCONNECTED:
2920 conn = (kib_conn_t *)cmid->context;
2921 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2922 CERROR("%s DISCONNECTED\n",
2923 libcfs_nid2str(conn->ibc_peer->ibp_nid));
2924 kiblnd_connreq_done(conn, -ECONNRESET);
2925 } else {
2926 kiblnd_close_conn(conn, 0);
2927 }
2928 kiblnd_conn_decref(conn);
2929 cmid->context = NULL;
2930 return 0;
2931
2932 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2933 LCONSOLE_ERROR_MSG(0x131,
2934 "Received notification of device removal\n"
2935 "Please shutdown LNET to allow this to proceed\n");
2936 /* Can't remove network from underneath LNET for now, so I have
2937 * to ignore this */
2938 return 0;
2939
2940 case RDMA_CM_EVENT_ADDR_CHANGE:
2941 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2942 return 0;
2943 }
2944 }
2945
2946 static int
kiblnd_check_txs_locked(kib_conn_t * conn,struct list_head * txs)2947 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
2948 {
2949 kib_tx_t *tx;
2950 struct list_head *ttmp;
2951
2952 list_for_each(ttmp, txs) {
2953 tx = list_entry(ttmp, kib_tx_t, tx_list);
2954
2955 if (txs != &conn->ibc_active_txs) {
2956 LASSERT(tx->tx_queued);
2957 } else {
2958 LASSERT(!tx->tx_queued);
2959 LASSERT(tx->tx_waiting || tx->tx_sending != 0);
2960 }
2961
2962 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
2963 CERROR("Timed out tx: %s, %lu seconds\n",
2964 kiblnd_queue2str(conn, txs),
2965 cfs_duration_sec(jiffies - tx->tx_deadline));
2966 return 1;
2967 }
2968 }
2969
2970 return 0;
2971 }
2972
2973 static int
kiblnd_conn_timed_out_locked(kib_conn_t * conn)2974 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
2975 {
2976 return kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
2977 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
2978 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
2979 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
2980 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
2981 }
2982
2983 static void
kiblnd_check_conns(int idx)2984 kiblnd_check_conns(int idx)
2985 {
2986 LIST_HEAD(closes);
2987 LIST_HEAD(checksends);
2988 struct list_head *peers = &kiblnd_data.kib_peers[idx];
2989 struct list_head *ptmp;
2990 kib_peer_t *peer;
2991 kib_conn_t *conn;
2992 kib_conn_t *tmp;
2993 struct list_head *ctmp;
2994 unsigned long flags;
2995
2996 /* NB. We expect to have a look at all the peers and not find any
2997 * RDMAs to time out, so we just use a shared lock while we
2998 * take a look... */
2999 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3000
3001 list_for_each(ptmp, peers) {
3002 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3003
3004 list_for_each(ctmp, &peer->ibp_conns) {
3005 int timedout;
3006 int sendnoop;
3007
3008 conn = list_entry(ctmp, kib_conn_t, ibc_list);
3009
3010 LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3011
3012 spin_lock(&conn->ibc_lock);
3013
3014 sendnoop = kiblnd_need_noop(conn);
3015 timedout = kiblnd_conn_timed_out_locked(conn);
3016 if (!sendnoop && !timedout) {
3017 spin_unlock(&conn->ibc_lock);
3018 continue;
3019 }
3020
3021 if (timedout) {
3022 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3023 libcfs_nid2str(peer->ibp_nid),
3024 cfs_duration_sec(cfs_time_current() -
3025 peer->ibp_last_alive),
3026 conn->ibc_credits,
3027 conn->ibc_outstanding_credits,
3028 conn->ibc_reserved_credits);
3029 list_add(&conn->ibc_connd_list, &closes);
3030 } else {
3031 list_add(&conn->ibc_connd_list,
3032 &checksends);
3033 }
3034 /* +ref for 'closes' or 'checksends' */
3035 kiblnd_conn_addref(conn);
3036
3037 spin_unlock(&conn->ibc_lock);
3038 }
3039 }
3040
3041 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3042
3043 /* Handle timeout by closing the whole
3044 * connection. We can only be sure RDMA activity
3045 * has ceased once the QP has been modified. */
3046 list_for_each_entry_safe(conn, tmp, &closes, ibc_connd_list) {
3047 list_del(&conn->ibc_connd_list);
3048 kiblnd_close_conn(conn, -ETIMEDOUT);
3049 kiblnd_conn_decref(conn);
3050 }
3051
3052 /* In case we have enough credits to return via a
3053 * NOOP, but there were no non-blocking tx descs
3054 * free to do it last time... */
3055 while (!list_empty(&checksends)) {
3056 conn = list_entry(checksends.next,
3057 kib_conn_t, ibc_connd_list);
3058 list_del(&conn->ibc_connd_list);
3059 kiblnd_check_sends(conn);
3060 kiblnd_conn_decref(conn);
3061 }
3062 }
3063
3064 static void
kiblnd_disconnect_conn(kib_conn_t * conn)3065 kiblnd_disconnect_conn(kib_conn_t *conn)
3066 {
3067 LASSERT(!in_interrupt());
3068 LASSERT(current == kiblnd_data.kib_connd);
3069 LASSERT(conn->ibc_state == IBLND_CONN_CLOSING);
3070
3071 rdma_disconnect(conn->ibc_cmid);
3072 kiblnd_finalise_conn(conn);
3073
3074 kiblnd_peer_notify(conn->ibc_peer);
3075 }
3076
3077 int
kiblnd_connd(void * arg)3078 kiblnd_connd(void *arg)
3079 {
3080 wait_queue_t wait;
3081 unsigned long flags;
3082 kib_conn_t *conn;
3083 int timeout;
3084 int i;
3085 int dropped_lock;
3086 int peer_index = 0;
3087 unsigned long deadline = jiffies;
3088
3089 cfs_block_allsigs();
3090
3091 init_waitqueue_entry(&wait, current);
3092 kiblnd_data.kib_connd = current;
3093
3094 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3095
3096 while (!kiblnd_data.kib_shutdown) {
3097
3098 dropped_lock = 0;
3099
3100 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3101 conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3102 kib_conn_t, ibc_list);
3103 list_del(&conn->ibc_list);
3104
3105 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3106 flags);
3107 dropped_lock = 1;
3108
3109 kiblnd_destroy_conn(conn);
3110
3111 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3112 }
3113
3114 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3115 conn = list_entry(kiblnd_data.kib_connd_conns.next,
3116 kib_conn_t, ibc_list);
3117 list_del(&conn->ibc_list);
3118
3119 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3120 flags);
3121 dropped_lock = 1;
3122
3123 kiblnd_disconnect_conn(conn);
3124 kiblnd_conn_decref(conn);
3125
3126 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3127 }
3128
3129 /* careful with the jiffy wrap... */
3130 timeout = (int)(deadline - jiffies);
3131 if (timeout <= 0) {
3132 const int n = 4;
3133 const int p = 1;
3134 int chunk = kiblnd_data.kib_peer_hash_size;
3135
3136 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3137 dropped_lock = 1;
3138
3139 /* Time to check for RDMA timeouts on a few more
3140 * peers: I do checks every 'p' seconds on a
3141 * proportion of the peer table and I need to check
3142 * every connection 'n' times within a timeout
3143 * interval, to ensure I detect a timeout on any
3144 * connection within (n+1)/n times the timeout
3145 * interval. */
3146
3147 if (*kiblnd_tunables.kib_timeout > n * p)
3148 chunk = (chunk * n * p) /
3149 *kiblnd_tunables.kib_timeout;
3150 if (chunk == 0)
3151 chunk = 1;
3152
3153 for (i = 0; i < chunk; i++) {
3154 kiblnd_check_conns(peer_index);
3155 peer_index = (peer_index + 1) %
3156 kiblnd_data.kib_peer_hash_size;
3157 }
3158
3159 deadline += p * HZ;
3160 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3161 }
3162
3163 if (dropped_lock)
3164 continue;
3165
3166 /* Nothing to do for 'timeout' */
3167 set_current_state(TASK_INTERRUPTIBLE);
3168 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3169 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3170
3171 schedule_timeout(timeout);
3172
3173 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3174 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3175 }
3176
3177 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3178
3179 kiblnd_thread_fini();
3180 return 0;
3181 }
3182
3183 void
kiblnd_qp_event(struct ib_event * event,void * arg)3184 kiblnd_qp_event(struct ib_event *event, void *arg)
3185 {
3186 kib_conn_t *conn = arg;
3187
3188 switch (event->event) {
3189 case IB_EVENT_COMM_EST:
3190 CDEBUG(D_NET, "%s established\n",
3191 libcfs_nid2str(conn->ibc_peer->ibp_nid));
3192 return;
3193
3194 default:
3195 CERROR("%s: Async QP event type %d\n",
3196 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3197 return;
3198 }
3199 }
3200
3201 static void
kiblnd_complete(struct ib_wc * wc)3202 kiblnd_complete(struct ib_wc *wc)
3203 {
3204 switch (kiblnd_wreqid2type(wc->wr_id)) {
3205 default:
3206 LBUG();
3207
3208 case IBLND_WID_RDMA:
3209 /* We only get RDMA completion notification if it fails. All
3210 * subsequent work items, including the final SEND will fail
3211 * too. However we can't print out any more info about the
3212 * failing RDMA because 'tx' might be back on the idle list or
3213 * even reused already if we didn't manage to post all our work
3214 * items */
3215 CNETERR("RDMA (tx: %p) failed: %d\n",
3216 kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3217 return;
3218
3219 case IBLND_WID_TX:
3220 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3221 return;
3222
3223 case IBLND_WID_RX:
3224 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3225 wc->byte_len);
3226 return;
3227 }
3228 }
3229
3230 void
kiblnd_cq_completion(struct ib_cq * cq,void * arg)3231 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3232 {
3233 /* NB I'm not allowed to schedule this conn once its refcount has
3234 * reached 0. Since fundamentally I'm racing with scheduler threads
3235 * consuming my CQ I could be called after all completions have
3236 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3237 * and this CQ is about to be destroyed so I NOOP. */
3238 kib_conn_t *conn = arg;
3239 struct kib_sched_info *sched = conn->ibc_sched;
3240 unsigned long flags;
3241
3242 LASSERT(cq == conn->ibc_cq);
3243
3244 spin_lock_irqsave(&sched->ibs_lock, flags);
3245
3246 conn->ibc_ready = 1;
3247
3248 if (!conn->ibc_scheduled &&
3249 (conn->ibc_nrx > 0 ||
3250 conn->ibc_nsends_posted > 0)) {
3251 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3252 conn->ibc_scheduled = 1;
3253 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3254
3255 if (waitqueue_active(&sched->ibs_waitq))
3256 wake_up(&sched->ibs_waitq);
3257 }
3258
3259 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3260 }
3261
3262 void
kiblnd_cq_event(struct ib_event * event,void * arg)3263 kiblnd_cq_event(struct ib_event *event, void *arg)
3264 {
3265 kib_conn_t *conn = arg;
3266
3267 CERROR("%s: async CQ event type %d\n",
3268 libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3269 }
3270
3271 int
kiblnd_scheduler(void * arg)3272 kiblnd_scheduler(void *arg)
3273 {
3274 long id = (long)arg;
3275 struct kib_sched_info *sched;
3276 kib_conn_t *conn;
3277 wait_queue_t wait;
3278 unsigned long flags;
3279 struct ib_wc wc;
3280 int did_something;
3281 int busy_loops = 0;
3282 int rc;
3283
3284 cfs_block_allsigs();
3285
3286 init_waitqueue_entry(&wait, current);
3287
3288 sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3289
3290 rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3291 if (rc != 0) {
3292 CWARN("Failed to bind on CPT %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3293 sched->ibs_cpt);
3294 }
3295
3296 spin_lock_irqsave(&sched->ibs_lock, flags);
3297
3298 while (!kiblnd_data.kib_shutdown) {
3299 if (busy_loops++ >= IBLND_RESCHED) {
3300 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3301
3302 cond_resched();
3303 busy_loops = 0;
3304
3305 spin_lock_irqsave(&sched->ibs_lock, flags);
3306 }
3307
3308 did_something = 0;
3309
3310 if (!list_empty(&sched->ibs_conns)) {
3311 conn = list_entry(sched->ibs_conns.next,
3312 kib_conn_t, ibc_sched_list);
3313 /* take over kib_sched_conns' ref on conn... */
3314 LASSERT(conn->ibc_scheduled);
3315 list_del(&conn->ibc_sched_list);
3316 conn->ibc_ready = 0;
3317
3318 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3319
3320 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3321 if (rc == 0) {
3322 rc = ib_req_notify_cq(conn->ibc_cq,
3323 IB_CQ_NEXT_COMP);
3324 if (rc < 0) {
3325 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3326 libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3327 kiblnd_close_conn(conn, -EIO);
3328 kiblnd_conn_decref(conn);
3329 spin_lock_irqsave(&sched->ibs_lock,
3330 flags);
3331 continue;
3332 }
3333
3334 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3335 }
3336
3337 if (rc < 0) {
3338 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3339 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3340 rc);
3341 kiblnd_close_conn(conn, -EIO);
3342 kiblnd_conn_decref(conn);
3343 spin_lock_irqsave(&sched->ibs_lock, flags);
3344 continue;
3345 }
3346
3347 spin_lock_irqsave(&sched->ibs_lock, flags);
3348
3349 if (rc != 0 || conn->ibc_ready) {
3350 /* There may be another completion waiting; get
3351 * another scheduler to check while I handle
3352 * this one... */
3353 /* +1 ref for sched_conns */
3354 kiblnd_conn_addref(conn);
3355 list_add_tail(&conn->ibc_sched_list,
3356 &sched->ibs_conns);
3357 if (waitqueue_active(&sched->ibs_waitq))
3358 wake_up(&sched->ibs_waitq);
3359 } else {
3360 conn->ibc_scheduled = 0;
3361 }
3362
3363 if (rc != 0) {
3364 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3365 kiblnd_complete(&wc);
3366
3367 spin_lock_irqsave(&sched->ibs_lock, flags);
3368 }
3369
3370 kiblnd_conn_decref(conn); /* ...drop my ref from above */
3371 did_something = 1;
3372 }
3373
3374 if (did_something)
3375 continue;
3376
3377 set_current_state(TASK_INTERRUPTIBLE);
3378 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3379 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3380
3381 schedule();
3382 busy_loops = 0;
3383
3384 remove_wait_queue(&sched->ibs_waitq, &wait);
3385 spin_lock_irqsave(&sched->ibs_lock, flags);
3386 }
3387
3388 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3389
3390 kiblnd_thread_fini();
3391 return 0;
3392 }
3393
3394 int
kiblnd_failover_thread(void * arg)3395 kiblnd_failover_thread(void *arg)
3396 {
3397 rwlock_t *glock = &kiblnd_data.kib_global_lock;
3398 kib_dev_t *dev;
3399 wait_queue_t wait;
3400 unsigned long flags;
3401 int rc;
3402
3403 LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3404
3405 cfs_block_allsigs();
3406
3407 init_waitqueue_entry(&wait, current);
3408 write_lock_irqsave(glock, flags);
3409
3410 while (!kiblnd_data.kib_shutdown) {
3411 int do_failover = 0;
3412 int long_sleep;
3413
3414 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3415 ibd_fail_list) {
3416 if (time_before(cfs_time_current(),
3417 dev->ibd_next_failover))
3418 continue;
3419 do_failover = 1;
3420 break;
3421 }
3422
3423 if (do_failover) {
3424 list_del_init(&dev->ibd_fail_list);
3425 dev->ibd_failover = 1;
3426 write_unlock_irqrestore(glock, flags);
3427
3428 rc = kiblnd_dev_failover(dev);
3429
3430 write_lock_irqsave(glock, flags);
3431
3432 LASSERT(dev->ibd_failover);
3433 dev->ibd_failover = 0;
3434 if (rc >= 0) { /* Device is OK or failover succeed */
3435 dev->ibd_next_failover = cfs_time_shift(3);
3436 continue;
3437 }
3438
3439 /* failed to failover, retry later */
3440 dev->ibd_next_failover =
3441 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3442 if (kiblnd_dev_can_failover(dev)) {
3443 list_add_tail(&dev->ibd_fail_list,
3444 &kiblnd_data.kib_failed_devs);
3445 }
3446
3447 continue;
3448 }
3449
3450 /* long sleep if no more pending failover */
3451 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3452
3453 set_current_state(TASK_INTERRUPTIBLE);
3454 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3455 write_unlock_irqrestore(glock, flags);
3456
3457 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3458 cfs_time_seconds(1));
3459 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3460 write_lock_irqsave(glock, flags);
3461
3462 if (!long_sleep || rc != 0)
3463 continue;
3464
3465 /* have a long sleep, routine check all active devices,
3466 * we need checking like this because if there is not active
3467 * connection on the dev and no SEND from local, we may listen
3468 * on wrong HCA for ever while there is a bonding failover */
3469 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3470 if (kiblnd_dev_can_failover(dev)) {
3471 list_add_tail(&dev->ibd_fail_list,
3472 &kiblnd_data.kib_failed_devs);
3473 }
3474 }
3475 }
3476
3477 write_unlock_irqrestore(glock, flags);
3478
3479 kiblnd_thread_fini();
3480 return 0;
3481 }
3482