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) 2002, 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
37 #define DEBUG_SUBSYSTEM S_RPC
38
39 #include "../../include/linux/libcfs/libcfs.h"
40 # ifdef __mips64__
41 # include <linux/kernel.h>
42 # endif
43
44 #include "../include/obd_class.h"
45 #include "../include/lustre_net.h"
46 #include "../include/lustre_sec.h"
47 #include "ptlrpc_internal.h"
48
49 lnet_handle_eq_t ptlrpc_eq_h;
50
51 /*
52 * Client's outgoing request callback
53 */
request_out_callback(lnet_event_t * ev)54 void request_out_callback(lnet_event_t *ev)
55 {
56 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
57 struct ptlrpc_request *req = cbid->cbid_arg;
58
59 LASSERT(ev->type == LNET_EVENT_SEND ||
60 ev->type == LNET_EVENT_UNLINK);
61 LASSERT(ev->unlinked);
62
63 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
64
65 sptlrpc_request_out_callback(req);
66 spin_lock(&req->rq_lock);
67 req->rq_real_sent = ktime_get_real_seconds();
68 if (ev->unlinked)
69 req->rq_req_unlink = 0;
70
71 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
72
73 /* Failed send: make it seem like the reply timed out, just
74 * like failing sends in client.c does currently... */
75
76 req->rq_net_err = 1;
77 ptlrpc_client_wake_req(req);
78 }
79 spin_unlock(&req->rq_lock);
80
81 ptlrpc_req_finished(req);
82 }
83
84 /*
85 * Client's incoming reply callback
86 */
reply_in_callback(lnet_event_t * ev)87 void reply_in_callback(lnet_event_t *ev)
88 {
89 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
90 struct ptlrpc_request *req = cbid->cbid_arg;
91
92 DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
93
94 LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
95 LASSERT(ev->md.start == req->rq_repbuf);
96 LASSERT(ev->offset + ev->mlength <= req->rq_repbuf_len);
97 /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
98 for adaptive timeouts' early reply. */
99 LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
100
101 spin_lock(&req->rq_lock);
102
103 req->rq_receiving_reply = 0;
104 req->rq_early = 0;
105 if (ev->unlinked)
106 req->rq_reply_unlink = 0;
107
108 if (ev->status)
109 goto out_wake;
110
111 if (ev->type == LNET_EVENT_UNLINK) {
112 LASSERT(ev->unlinked);
113 DEBUG_REQ(D_NET, req, "unlink");
114 goto out_wake;
115 }
116
117 if (ev->mlength < ev->rlength) {
118 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
119 req->rq_replen, ev->rlength, ev->offset);
120 req->rq_reply_truncate = 1;
121 req->rq_replied = 1;
122 req->rq_status = -EOVERFLOW;
123 req->rq_nob_received = ev->rlength + ev->offset;
124 goto out_wake;
125 }
126
127 if ((ev->offset == 0) &&
128 ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
129 /* Early reply */
130 DEBUG_REQ(D_ADAPTTO, req,
131 "Early reply received: mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
132 ev->mlength, ev->offset,
133 req->rq_replen, req->rq_replied, ev->unlinked);
134
135 req->rq_early_count++; /* number received, client side */
136
137 if (req->rq_replied) /* already got the real reply */
138 goto out_wake;
139
140 req->rq_early = 1;
141 req->rq_reply_off = ev->offset;
142 req->rq_nob_received = ev->mlength;
143 /* And we're still receiving */
144 req->rq_receiving_reply = 1;
145 } else {
146 /* Real reply */
147 req->rq_rep_swab_mask = 0;
148 req->rq_replied = 1;
149 /* Got reply, no resend required */
150 req->rq_resend = 0;
151 req->rq_reply_off = ev->offset;
152 req->rq_nob_received = ev->mlength;
153 /* LNetMDUnlink can't be called under the LNET_LOCK,
154 so we must unlink in ptlrpc_unregister_reply */
155 DEBUG_REQ(D_INFO, req,
156 "reply in flags=%x mlen=%u offset=%d replen=%d",
157 lustre_msg_get_flags(req->rq_reqmsg),
158 ev->mlength, ev->offset, req->rq_replen);
159 }
160
161 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
162
163 out_wake:
164 /* NB don't unlock till after wakeup; req can disappear under us
165 * since we don't have our own ref */
166 ptlrpc_client_wake_req(req);
167 spin_unlock(&req->rq_lock);
168 }
169
170 /*
171 * Client's bulk has been written/read
172 */
client_bulk_callback(lnet_event_t * ev)173 void client_bulk_callback(lnet_event_t *ev)
174 {
175 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
176 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
177 struct ptlrpc_request *req;
178
179 LASSERT((desc->bd_type == BULK_PUT_SINK &&
180 ev->type == LNET_EVENT_PUT) ||
181 (desc->bd_type == BULK_GET_SOURCE &&
182 ev->type == LNET_EVENT_GET) ||
183 ev->type == LNET_EVENT_UNLINK);
184 LASSERT(ev->unlinked);
185
186 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
187 ev->status = -EIO;
188
189 if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
190 CFS_FAIL_ONCE))
191 ev->status = -EIO;
192
193 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
194 "event type %d, status %d, desc %p\n",
195 ev->type, ev->status, desc);
196
197 spin_lock(&desc->bd_lock);
198 req = desc->bd_req;
199 LASSERT(desc->bd_md_count > 0);
200 desc->bd_md_count--;
201
202 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
203 desc->bd_nob_transferred += ev->mlength;
204 desc->bd_sender = ev->sender;
205 } else {
206 /* start reconnect and resend if network error hit */
207 spin_lock(&req->rq_lock);
208 req->rq_net_err = 1;
209 spin_unlock(&req->rq_lock);
210 }
211
212 if (ev->status != 0)
213 desc->bd_failure = 1;
214
215 /* NB don't unlock till after wakeup; desc can disappear under us
216 * otherwise */
217 if (desc->bd_md_count == 0)
218 ptlrpc_client_wake_req(desc->bd_req);
219
220 spin_unlock(&desc->bd_lock);
221 }
222
223 /*
224 * We will have percpt request history list for ptlrpc service in upcoming
225 * patches because we don't want to be serialized by current per-service
226 * history operations. So we require history ID can (somehow) show arriving
227 * order w/o grabbing global lock, and user can sort them in userspace.
228 *
229 * This is how we generate history ID for ptlrpc_request:
230 * ----------------------------------------------------
231 * | 32 bits | 16 bits | (16 - X)bits | X bits |
232 * ----------------------------------------------------
233 * | seconds | usec / 16 | sequence | CPT id |
234 * ----------------------------------------------------
235 *
236 * it might not be precise but should be good enough.
237 */
238
239 #define REQS_CPT_BITS(svcpt) ((svcpt)->scp_service->srv_cpt_bits)
240
241 #define REQS_SEC_SHIFT 32
242 #define REQS_USEC_SHIFT 16
243 #define REQS_SEQ_SHIFT(svcpt) REQS_CPT_BITS(svcpt)
244
ptlrpc_req_add_history(struct ptlrpc_service_part * svcpt,struct ptlrpc_request * req)245 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
246 struct ptlrpc_request *req)
247 {
248 __u64 sec = req->rq_arrival_time.tv_sec;
249 __u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
250 __u64 new_seq;
251
252 /* set sequence ID for request and add it to history list,
253 * it must be called with hold svcpt::scp_lock */
254
255 new_seq = (sec << REQS_SEC_SHIFT) |
256 (usec << REQS_USEC_SHIFT) |
257 (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
258
259 if (new_seq > svcpt->scp_hist_seq) {
260 /* This handles the initial case of scp_hist_seq == 0 or
261 * we just jumped into a new time window */
262 svcpt->scp_hist_seq = new_seq;
263 } else {
264 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
265 /* NB: increase sequence number in current usec bucket,
266 * however, it's possible that we used up all bits for
267 * sequence and jumped into the next usec bucket (future time),
268 * then we hope there will be less RPCs per bucket at some
269 * point, and sequence will catch up again */
270 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
271 new_seq = svcpt->scp_hist_seq;
272 }
273
274 req->rq_history_seq = new_seq;
275
276 list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
277 }
278
279 /*
280 * Server's incoming request callback
281 */
request_in_callback(lnet_event_t * ev)282 void request_in_callback(lnet_event_t *ev)
283 {
284 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
285 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
286 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
287 struct ptlrpc_service *service = svcpt->scp_service;
288 struct ptlrpc_request *req;
289
290 LASSERT(ev->type == LNET_EVENT_PUT ||
291 ev->type == LNET_EVENT_UNLINK);
292 LASSERT((char *)ev->md.start >= rqbd->rqbd_buffer);
293 LASSERT((char *)ev->md.start + ev->offset + ev->mlength <=
294 rqbd->rqbd_buffer + service->srv_buf_size);
295
296 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
297 "event type %d, status %d, service %s\n",
298 ev->type, ev->status, service->srv_name);
299
300 if (ev->unlinked) {
301 /* If this is the last request message to fit in the
302 * request buffer we can use the request object embedded in
303 * rqbd. Note that if we failed to allocate a request,
304 * we'd have to re-post the rqbd, which we can't do in this
305 * context. */
306 req = &rqbd->rqbd_req;
307 memset(req, 0, sizeof(*req));
308 } else {
309 LASSERT(ev->type == LNET_EVENT_PUT);
310 if (ev->status != 0) {
311 /* We moaned above already... */
312 return;
313 }
314 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
315 if (req == NULL) {
316 CERROR("Can't allocate incoming request descriptor: Dropping %s RPC from %s\n",
317 service->srv_name,
318 libcfs_id2str(ev->initiator));
319 return;
320 }
321 }
322
323 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
324 * flags are reset and scalars are zero. We only set the message
325 * size to non-zero if this was a successful receive. */
326 req->rq_xid = ev->match_bits;
327 req->rq_reqbuf = ev->md.start + ev->offset;
328 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
329 req->rq_reqdata_len = ev->mlength;
330 ktime_get_real_ts64(&req->rq_arrival_time);
331 req->rq_peer = ev->initiator;
332 req->rq_self = ev->target.nid;
333 req->rq_rqbd = rqbd;
334 req->rq_phase = RQ_PHASE_NEW;
335 spin_lock_init(&req->rq_lock);
336 INIT_LIST_HEAD(&req->rq_timed_list);
337 INIT_LIST_HEAD(&req->rq_exp_list);
338 atomic_set(&req->rq_refcount, 1);
339 if (ev->type == LNET_EVENT_PUT)
340 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
341 req, req->rq_xid, ev->mlength);
342
343 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
344
345 spin_lock(&svcpt->scp_lock);
346
347 ptlrpc_req_add_history(svcpt, req);
348
349 if (ev->unlinked) {
350 svcpt->scp_nrqbds_posted--;
351 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
352 svcpt->scp_nrqbds_posted);
353
354 /* Normally, don't complain about 0 buffers posted; LNET won't
355 * drop incoming reqs since we set the portal lazy */
356 if (test_req_buffer_pressure &&
357 ev->type != LNET_EVENT_UNLINK &&
358 svcpt->scp_nrqbds_posted == 0)
359 CWARN("All %s request buffers busy\n",
360 service->srv_name);
361
362 /* req takes over the network's ref on rqbd */
363 } else {
364 /* req takes a ref on rqbd */
365 rqbd->rqbd_refcount++;
366 }
367
368 list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
369 svcpt->scp_nreqs_incoming++;
370
371 /* NB everything can disappear under us once the request
372 * has been queued and we unlock, so do the wake now... */
373 wake_up(&svcpt->scp_waitq);
374
375 spin_unlock(&svcpt->scp_lock);
376 }
377
378 /*
379 * Server's outgoing reply callback
380 */
reply_out_callback(lnet_event_t * ev)381 void reply_out_callback(lnet_event_t *ev)
382 {
383 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
384 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
385 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
386
387 LASSERT(ev->type == LNET_EVENT_SEND ||
388 ev->type == LNET_EVENT_ACK ||
389 ev->type == LNET_EVENT_UNLINK);
390
391 if (!rs->rs_difficult) {
392 /* 'Easy' replies have no further processing so I drop the
393 * net's ref on 'rs' */
394 LASSERT(ev->unlinked);
395 ptlrpc_rs_decref(rs);
396 return;
397 }
398
399 LASSERT(rs->rs_on_net);
400
401 if (ev->unlinked) {
402 /* Last network callback. The net's ref on 'rs' stays put
403 * until ptlrpc_handle_rs() is done with it */
404 spin_lock(&svcpt->scp_rep_lock);
405 spin_lock(&rs->rs_lock);
406
407 rs->rs_on_net = 0;
408 if (!rs->rs_no_ack ||
409 rs->rs_transno <=
410 rs->rs_export->exp_obd->obd_last_committed)
411 ptlrpc_schedule_difficult_reply(rs);
412
413 spin_unlock(&rs->rs_lock);
414 spin_unlock(&svcpt->scp_rep_lock);
415 }
416 }
417
ptlrpc_master_callback(lnet_event_t * ev)418 static void ptlrpc_master_callback(lnet_event_t *ev)
419 {
420 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
421 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
422
423 /* Honestly, it's best to find out early. */
424 LASSERT(cbid->cbid_arg != LP_POISON);
425 LASSERT(callback == request_out_callback ||
426 callback == reply_in_callback ||
427 callback == client_bulk_callback ||
428 callback == request_in_callback ||
429 callback == reply_out_callback);
430
431 callback(ev);
432 }
433
ptlrpc_uuid_to_peer(struct obd_uuid * uuid,lnet_process_id_t * peer,lnet_nid_t * self)434 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
435 lnet_process_id_t *peer, lnet_nid_t *self)
436 {
437 int best_dist = 0;
438 __u32 best_order = 0;
439 int count = 0;
440 int rc = -ENOENT;
441 int portals_compatibility;
442 int dist;
443 __u32 order;
444 lnet_nid_t dst_nid;
445 lnet_nid_t src_nid;
446
447 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
448
449 peer->pid = LUSTRE_SRV_LNET_PID;
450
451 /* Choose the matching UUID that's closest */
452 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
453 dist = LNetDist(dst_nid, &src_nid, &order);
454 if (dist < 0)
455 continue;
456
457 if (dist == 0) { /* local! use loopback LND */
458 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
459 rc = 0;
460 break;
461 }
462
463 if (rc < 0 ||
464 dist < best_dist ||
465 (dist == best_dist && order < best_order)) {
466 best_dist = dist;
467 best_order = order;
468
469 if (portals_compatibility > 1) {
470 /* Strong portals compatibility: Zero the nid's
471 * NET, so if I'm reading new config logs, or
472 * getting configured by (new) lconf I can
473 * still talk to old servers. */
474 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
475 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
476 }
477 peer->nid = dst_nid;
478 *self = src_nid;
479 rc = 0;
480 }
481 }
482
483 CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
484 return rc;
485 }
486
ptlrpc_ni_fini(void)487 static void ptlrpc_ni_fini(void)
488 {
489 wait_queue_head_t waitq;
490 struct l_wait_info lwi;
491 int rc;
492 int retries;
493
494 /* Wait for the event queue to become idle since there may still be
495 * messages in flight with pending events (i.e. the fire-and-forget
496 * messages == client requests and "non-difficult" server
497 * replies */
498
499 for (retries = 0;; retries++) {
500 rc = LNetEQFree(ptlrpc_eq_h);
501 switch (rc) {
502 default:
503 LBUG();
504
505 case 0:
506 LNetNIFini();
507 return;
508
509 case -EBUSY:
510 if (retries != 0)
511 CWARN("Event queue still busy\n");
512
513 /* Wait for a bit */
514 init_waitqueue_head(&waitq);
515 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
516 l_wait_event(waitq, 0, &lwi);
517 break;
518 }
519 }
520 /* notreached */
521 }
522
ptl_get_pid(void)523 static lnet_pid_t ptl_get_pid(void)
524 {
525 lnet_pid_t pid;
526
527 pid = LUSTRE_SRV_LNET_PID;
528 return pid;
529 }
530
ptlrpc_ni_init(void)531 static int ptlrpc_ni_init(void)
532 {
533 int rc;
534 lnet_pid_t pid;
535
536 pid = ptl_get_pid();
537 CDEBUG(D_NET, "My pid is: %x\n", pid);
538
539 /* We're not passing any limits yet... */
540 rc = LNetNIInit(pid);
541 if (rc < 0) {
542 CDEBUG(D_NET, "Can't init network interface: %d\n", rc);
543 return -ENOENT;
544 }
545
546 /* CAVEAT EMPTOR: how we process portals events is _radically_
547 * different depending on... */
548 /* kernel LNet calls our master callback when there are new event,
549 * because we are guaranteed to get every event via callback,
550 * so we just set EQ size to 0 to avoid overhead of serializing
551 * enqueue/dequeue operations in LNet. */
552 rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
553 if (rc == 0)
554 return 0;
555
556 CERROR("Failed to allocate event queue: %d\n", rc);
557 LNetNIFini();
558
559 return -ENOMEM;
560 }
561
ptlrpc_init_portals(void)562 int ptlrpc_init_portals(void)
563 {
564 int rc = ptlrpc_ni_init();
565
566 if (rc != 0) {
567 CERROR("network initialisation failed\n");
568 return -EIO;
569 }
570 rc = ptlrpcd_addref();
571 if (rc == 0)
572 return 0;
573
574 CERROR("rpcd initialisation failed\n");
575 ptlrpc_ni_fini();
576 return rc;
577 }
578
ptlrpc_exit_portals(void)579 void ptlrpc_exit_portals(void)
580 {
581 ptlrpcd_decref();
582 ptlrpc_ni_fini();
583 }
584