Whamcloud - gitweb
LU-8901 misc: update Intel copyright messages for 2016
[fs/lustre-release.git] / lustre / ptlrpc / events.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2016, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 #define DEBUG_SUBSYSTEM S_RPC
34
35 #include <libcfs/libcfs.h>
36 #include <linux/kernel.h>
37 #include <obd_class.h>
38 #include <lustre_net.h>
39 #include <lustre_sec.h>
40 #include "ptlrpc_internal.h"
41
42 lnet_handle_eq_t   ptlrpc_eq_h;
43
44 /*
45  *  Client's outgoing request callback
46  */
47 void request_out_callback(lnet_event_t *ev)
48 {
49         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
50         struct ptlrpc_request *req = cbid->cbid_arg;
51         bool                   wakeup = false;
52         ENTRY;
53
54         LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
55         LASSERT(ev->unlinked);
56
57         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
58
59         sptlrpc_request_out_callback(req);
60
61         spin_lock(&req->rq_lock);
62         req->rq_real_sent = cfs_time_current_sec();
63         req->rq_req_unlinked = 1;
64         /* reply_in_callback happened before request_out_callback? */
65         if (req->rq_reply_unlinked)
66                 wakeup = true;
67
68         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
69                 /* Failed send: make it seem like the reply timed out, just
70                  * like failing sends in client.c does currently...  */
71                 req->rq_net_err = 1;
72                 wakeup = true;
73         }
74
75         if (wakeup)
76                 ptlrpc_client_wake_req(req);
77
78         spin_unlock(&req->rq_lock);
79
80         ptlrpc_req_finished(req);
81         EXIT;
82 }
83
84 /*
85  * Client's incoming reply callback
86  */
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         ENTRY;
92
93         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
94
95         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
96         LASSERT (ev->md.start == req->rq_repbuf);
97         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
98         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
99            for adaptive timeouts' early reply. */
100         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
101
102         spin_lock(&req->rq_lock);
103
104         req->rq_receiving_reply = 0;
105         req->rq_early = 0;
106         if (ev->unlinked)
107                 req->rq_reply_unlinked = 1;
108
109         if (ev->status)
110                 goto out_wake;
111
112         if (ev->type == LNET_EVENT_UNLINK) {
113                 LASSERT(ev->unlinked);
114                 DEBUG_REQ(D_NET, req, "unlink");
115                 goto out_wake;
116         }
117
118         if (ev->mlength < ev->rlength ) {
119                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
120                        req->rq_replen, ev->rlength, ev->offset);
121                 req->rq_reply_truncated = 1;
122                 req->rq_replied = 1;
123                 req->rq_status = -EOVERFLOW;
124                 req->rq_nob_received = ev->rlength + ev->offset;
125                 goto out_wake;
126         }
127
128         if ((ev->offset == 0) &&
129             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
130                 /* Early reply */
131                 DEBUG_REQ(D_ADAPTTO, req,
132                           "Early reply received: mlen=%u offset=%d replen=%d "
133                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
134                           req->rq_replen, req->rq_replied, ev->unlinked);
135
136                 req->rq_early_count++; /* number received, client side */
137
138                 /* already got the real reply or buffers are already unlinked */
139                 if (req->rq_replied ||
140                     req->rq_reply_unlinked == 1)
141                         goto out_wake;
142
143                 req->rq_early = 1;
144                 req->rq_reply_off = ev->offset;
145                 req->rq_nob_received = ev->mlength;
146                 /* And we're still receiving */
147                 req->rq_receiving_reply = 1;
148         } else {
149                 /* Real reply */
150                 req->rq_rep_swab_mask = 0;
151                 req->rq_replied = 1;
152                 /* Got reply, no resend required */
153                 req->rq_resend = 0;
154                 req->rq_reply_off = ev->offset;
155                 req->rq_nob_received = ev->mlength;
156                 /* LNetMDUnlink can't be called under the LNET_LOCK,
157                    so we must unlink in ptlrpc_unregister_reply */
158                 DEBUG_REQ(D_INFO, req,
159                           "reply in flags=%x mlen=%u offset=%d replen=%d",
160                           lustre_msg_get_flags(req->rq_reqmsg),
161                           ev->mlength, ev->offset, req->rq_replen);
162         }
163
164         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
165
166 out_wake:
167         /* NB don't unlock till after wakeup; req can disappear under us
168          * since we don't have our own ref */
169         ptlrpc_client_wake_req(req);
170         spin_unlock(&req->rq_lock);
171         EXIT;
172 }
173
174 /*
175  * Client's bulk has been written/read
176  */
177 void client_bulk_callback (lnet_event_t *ev)
178 {
179         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
180         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
181         struct ptlrpc_request   *req;
182         ENTRY;
183
184         LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
185                  ev->type == LNET_EVENT_PUT) ||
186                 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
187                  ev->type == LNET_EVENT_GET) ||
188                 ev->type == LNET_EVENT_UNLINK);
189         LASSERT(ev->unlinked);
190
191         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
192                 ev->status = -EIO;
193
194         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
195                 ev->status = -EIO;
196
197         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
198                "event type %d, status %d, desc %p\n",
199                ev->type, ev->status, desc);
200
201         spin_lock(&desc->bd_lock);
202         req = desc->bd_req;
203         LASSERT(desc->bd_md_count > 0);
204         desc->bd_md_count--;
205
206         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
207                 desc->bd_nob_transferred += ev->mlength;
208                 desc->bd_sender = ev->sender;
209         } else {
210                 /* start reconnect and resend if network error hit */
211                 spin_lock(&req->rq_lock);
212                 req->rq_net_err = 1;
213                 spin_unlock(&req->rq_lock);
214         }
215
216         if (ev->status != 0)
217                 desc->bd_failure = 1;
218
219         /* NB don't unlock till after wakeup; desc can disappear under us
220          * otherwise */
221         if (desc->bd_md_count == 0)
222                 ptlrpc_client_wake_req(desc->bd_req);
223
224         spin_unlock(&desc->bd_lock);
225         EXIT;
226 }
227
228 /*
229  * We will have percpt request history list for ptlrpc service in upcoming
230  * patches because we don't want to be serialized by current per-service
231  * history operations. So we require history ID can (somehow) show arriving
232  * order w/o grabbing global lock, and user can sort them in userspace.
233  *
234  * This is how we generate history ID for ptlrpc_request:
235  * ----------------------------------------------------
236  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
237  * ----------------------------------------------------
238  * |  seconds  | usec / 16 |   sequence    | CPT id   |
239  * ----------------------------------------------------
240  *
241  * it might not be precise but should be good enough.
242  */
243
244 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
245
246 #define REQS_SEC_SHIFT          32
247 #define REQS_USEC_SHIFT         16
248 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
249
250 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
251                                    struct ptlrpc_request *req)
252 {
253         __u64   sec = req->rq_arrival_time.tv_sec;
254         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
255         __u64   new_seq;
256
257         /* set sequence ID for request and add it to history list,
258          * it must be called with hold svcpt::scp_lock */
259
260         new_seq = (sec << REQS_SEC_SHIFT) |
261                   (usec << REQS_USEC_SHIFT) |
262                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
263
264         if (new_seq > svcpt->scp_hist_seq) {
265                 /* This handles the initial case of scp_hist_seq == 0 or
266                  * we just jumped into a new time window */
267                 svcpt->scp_hist_seq = new_seq;
268         } else {
269                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
270                 /* NB: increase sequence number in current usec bucket,
271                  * however, it's possible that we used up all bits for
272                  * sequence and jumped into the next usec bucket (future time),
273                  * then we hope there will be less RPCs per bucket at some
274                  * point, and sequence will catch up again */
275                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
276                 new_seq = svcpt->scp_hist_seq;
277         }
278
279         req->rq_history_seq = new_seq;
280
281         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
282 }
283
284 /*
285  * Server's incoming request callback
286  */
287 void request_in_callback(lnet_event_t *ev)
288 {
289         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
290         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
291         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
292         struct ptlrpc_service             *service = svcpt->scp_service;
293         struct ptlrpc_request             *req;
294         ENTRY;
295
296         LASSERT (ev->type == LNET_EVENT_PUT ||
297                  ev->type == LNET_EVENT_UNLINK);
298         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
299         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
300                  rqbd->rqbd_buffer + service->srv_buf_size);
301
302         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
303                "event type %d, status %d, service %s\n",
304                ev->type, ev->status, service->srv_name);
305
306         if (ev->unlinked) {
307                 /* If this is the last request message to fit in the
308                  * request buffer we can use the request object embedded in
309                  * rqbd.  Note that if we failed to allocate a request,
310                  * we'd have to re-post the rqbd, which we can't do in this
311                  * context. */
312                 req = &rqbd->rqbd_req;
313                 memset(req, 0, sizeof (*req));
314         } else {
315                 LASSERT (ev->type == LNET_EVENT_PUT);
316                 if (ev->status != 0) {
317                         /* We moaned above already... */
318                         return;
319                 }
320                 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
321                 if (req == NULL) {
322                         CERROR("Can't allocate incoming request descriptor: "
323                                "Dropping %s RPC from %s\n",
324                                service->srv_name,
325                                libcfs_id2str(ev->initiator));
326                         return;
327                 }
328         }
329
330         ptlrpc_srv_req_init(req);
331         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
332          * flags are reset and scalars are zero.  We only set the message
333          * size to non-zero if this was a successful receive. */
334         req->rq_xid = ev->match_bits;
335         req->rq_reqbuf = ev->md.start + ev->offset;
336         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
337                 req->rq_reqdata_len = ev->mlength;
338         do_gettimeofday(&req->rq_arrival_time);
339         req->rq_peer = ev->initiator;
340         req->rq_self = ev->target.nid;
341         req->rq_rqbd = rqbd;
342         req->rq_phase = RQ_PHASE_NEW;
343         if (ev->type == LNET_EVENT_PUT)
344                 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
345                        req, req->rq_xid, ev->mlength);
346
347         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
348
349         spin_lock(&svcpt->scp_lock);
350
351         ptlrpc_req_add_history(svcpt, req);
352
353         if (ev->unlinked) {
354                 svcpt->scp_nrqbds_posted--;
355                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
356                        svcpt->scp_nrqbds_posted);
357
358                 /* Normally, don't complain about 0 buffers posted; LNET won't
359                  * drop incoming reqs since we set the portal lazy */
360                 if (test_req_buffer_pressure &&
361                     ev->type != LNET_EVENT_UNLINK &&
362                     svcpt->scp_nrqbds_posted == 0)
363                         CWARN("All %s request buffers busy\n",
364                               service->srv_name);
365
366                 /* req takes over the network's ref on rqbd */
367         } else {
368                 /* req takes a ref on rqbd */
369                 rqbd->rqbd_refcount++;
370         }
371
372         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
373         svcpt->scp_nreqs_incoming++;
374
375         /* NB everything can disappear under us once the request
376          * has been queued and we unlock, so do the wake now... */
377         wake_up(&svcpt->scp_waitq);
378
379         spin_unlock(&svcpt->scp_lock);
380         EXIT;
381 }
382
383 /*
384  *  Server's outgoing reply callback
385  */
386 void reply_out_callback(lnet_event_t *ev)
387 {
388         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
389         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
390         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
391         ENTRY;
392
393         LASSERT (ev->type == LNET_EVENT_SEND ||
394                  ev->type == LNET_EVENT_ACK ||
395                  ev->type == LNET_EVENT_UNLINK);
396
397         if (!rs->rs_difficult) {
398                 /* 'Easy' replies have no further processing so I drop the
399                  * net's ref on 'rs' */
400                 LASSERT (ev->unlinked);
401                 ptlrpc_rs_decref(rs);
402                 EXIT;
403                 return;
404         }
405
406         LASSERT (rs->rs_on_net);
407
408         if (ev->unlinked) {
409                 /* Last network callback. The net's ref on 'rs' stays put
410                  * until ptlrpc_handle_rs() is done with it */
411                 spin_lock(&svcpt->scp_rep_lock);
412                 spin_lock(&rs->rs_lock);
413
414                 rs->rs_on_net = 0;
415                 if (!rs->rs_no_ack ||
416                     rs->rs_transno <=
417                     rs->rs_export->exp_obd->obd_last_committed ||
418                     list_empty(&rs->rs_obd_list))
419                         ptlrpc_schedule_difficult_reply(rs);
420
421                 spin_unlock(&rs->rs_lock);
422                 spin_unlock(&svcpt->scp_rep_lock);
423         }
424         EXIT;
425 }
426
427 #ifdef HAVE_SERVER_SUPPORT
428 /*
429  * Server's bulk completion callback
430  */
431 void server_bulk_callback (lnet_event_t *ev)
432 {
433         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
434         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
435         ENTRY;
436
437         LASSERT(ev->type == LNET_EVENT_SEND ||
438                 ev->type == LNET_EVENT_UNLINK ||
439                 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
440                  ev->type == LNET_EVENT_ACK) ||
441                 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
442                  ev->type == LNET_EVENT_REPLY));
443
444         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
445                "event type %d, status %d, desc %p\n",
446                ev->type, ev->status, desc);
447
448         spin_lock(&desc->bd_lock);
449
450         LASSERT(desc->bd_md_count > 0);
451
452         if ((ev->type == LNET_EVENT_ACK ||
453              ev->type == LNET_EVENT_REPLY) &&
454             ev->status == 0) {
455                 /* We heard back from the peer, so even if we get this
456                  * before the SENT event (oh yes we can), we know we
457                  * read/wrote the peer buffer and how much... */
458                 desc->bd_nob_transferred += ev->mlength;
459                 desc->bd_sender = ev->sender;
460         }
461
462         if (ev->status != 0)
463                 desc->bd_failure = 1;
464
465         if (ev->unlinked) {
466                 desc->bd_md_count--;
467                 /* This is the last callback no matter what... */
468                 if (desc->bd_md_count == 0)
469                         wake_up(&desc->bd_waitq);
470         }
471
472         spin_unlock(&desc->bd_lock);
473         EXIT;
474 }
475 #endif
476
477 static void ptlrpc_master_callback(lnet_event_t *ev)
478 {
479         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
480         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
481
482         /* Honestly, it's best to find out early. */
483         LASSERT (cbid->cbid_arg != LP_POISON);
484         LASSERT (callback == request_out_callback ||
485                  callback == reply_in_callback ||
486                  callback == client_bulk_callback ||
487                  callback == request_in_callback ||
488                  callback == reply_out_callback
489 #ifdef HAVE_SERVER_SUPPORT
490                  || callback == server_bulk_callback
491 #endif
492                  );
493
494         callback (ev);
495 }
496
497 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
498                          lnet_process_id_t *peer, lnet_nid_t *self)
499 {
500         int               best_dist = 0;
501         __u32             best_order = 0;
502         int               count = 0;
503         int               rc = -ENOENT;
504         int               dist;
505         __u32             order;
506         lnet_nid_t        dst_nid;
507         lnet_nid_t        src_nid;
508
509         peer->pid = LNET_PID_LUSTRE;
510
511         /* Choose the matching UUID that's closest */
512         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
513                 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
514                     LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
515                         continue;
516
517                 dist = LNetDist(dst_nid, &src_nid, &order);
518                 if (dist < 0)
519                         continue;
520
521                 if (dist == 0) {                /* local! use loopback LND */
522                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
523                         rc = 0;
524                         break;
525                 }
526
527                 if (rc < 0 ||
528                     dist < best_dist ||
529                     (dist == best_dist && order < best_order)) {
530                         best_dist = dist;
531                         best_order = order;
532
533                         peer->nid = dst_nid;
534                         *self = src_nid;
535                         rc = 0;
536                 }
537         }
538
539         CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
540         return rc;
541 }
542
543 void ptlrpc_ni_fini(void)
544 {
545         wait_queue_head_t         waitq;
546         struct l_wait_info  lwi;
547         int                 rc;
548         int                 retries;
549
550         /* Wait for the event queue to become idle since there may still be
551          * messages in flight with pending events (i.e. the fire-and-forget
552          * messages == client requests and "non-difficult" server
553          * replies */
554
555         for (retries = 0;; retries++) {
556                 rc = LNetEQFree(ptlrpc_eq_h);
557                 switch (rc) {
558                 default:
559                         LBUG();
560
561                 case 0:
562                         LNetNIFini();
563                         return;
564
565                 case -EBUSY:
566                         if (retries != 0)
567                                 CWARN("Event queue still busy\n");
568
569                         /* Wait for a bit */
570                         init_waitqueue_head(&waitq);
571                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
572                         l_wait_event(waitq, 0, &lwi);
573                         break;
574                 }
575         }
576         /* notreached */
577 }
578
579 lnet_pid_t ptl_get_pid(void)
580 {
581         return LNET_PID_LUSTRE;
582 }
583
584 int ptlrpc_ni_init(void)
585 {
586         int              rc;
587         lnet_pid_t       pid;
588
589         pid = ptl_get_pid();
590         CDEBUG(D_NET, "My pid is: %x\n", pid);
591
592         /* We're not passing any limits yet... */
593         rc = LNetNIInit(pid);
594         if (rc < 0) {
595                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
596                 return rc;
597         }
598
599         /* CAVEAT EMPTOR: how we process portals events is _radically_
600          * different depending on... */
601         /* kernel LNet calls our master callback when there are new event,
602          * because we are guaranteed to get every event via callback,
603          * so we just set EQ size to 0 to avoid overhread of serializing
604          * enqueue/dequeue operations in LNet. */
605         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
606         if (rc == 0)
607                 return 0;
608
609         CERROR ("Failed to allocate event queue: %d\n", rc);
610         LNetNIFini();
611
612         return rc;
613 }
614
615
616 int ptlrpc_init_portals(void)
617 {
618         int   rc = ptlrpc_ni_init();
619
620         if (rc != 0) {
621                 CERROR("network initialisation failed\n");
622                 return rc;
623         }
624         rc = ptlrpcd_addref();
625         if (rc == 0)
626                 return 0;
627
628         CERROR("rpcd initialisation failed\n");
629         ptlrpc_ni_fini();
630         return rc;
631 }
632
633 void ptlrpc_exit_portals(void)
634 {
635         ptlrpcd_decref();
636         ptlrpc_ni_fini();
637 }