Whamcloud - gitweb
035aa47a2af4eeb8265238c3fed5db36c0508219
[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.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, 2015, 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 <libcfs/libcfs.h>
40 #include <linux/kernel.h>
41 #include <obd_class.h>
42 #include <lustre_net.h>
43 #include <lustre_sec.h>
44 #include "ptlrpc_internal.h"
45
46 lnet_handle_eq_t   ptlrpc_eq_h;
47
48 /*
49  *  Client's outgoing request callback
50  */
51 void request_out_callback(lnet_event_t *ev)
52 {
53         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
54         struct ptlrpc_request *req = cbid->cbid_arg;
55         bool                   wakeup = false;
56         ENTRY;
57
58         LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
59         LASSERT(ev->unlinked);
60
61         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
62
63         sptlrpc_request_out_callback(req);
64
65         spin_lock(&req->rq_lock);
66         req->rq_real_sent = cfs_time_current_sec();
67         req->rq_req_unlinked = 1;
68         /* reply_in_callback happened before request_out_callback? */
69         if (req->rq_reply_unlinked)
70                 wakeup = true;
71
72         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
73                 /* Failed send: make it seem like the reply timed out, just
74                  * like failing sends in client.c does currently...  */
75                 req->rq_net_err = 1;
76                 wakeup = true;
77         }
78
79         if (wakeup)
80                 ptlrpc_client_wake_req(req);
81
82         spin_unlock(&req->rq_lock);
83
84         ptlrpc_req_finished(req);
85         EXIT;
86 }
87
88 /*
89  * Client's incoming reply callback
90  */
91 void reply_in_callback(lnet_event_t *ev)
92 {
93         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
94         struct ptlrpc_request *req = cbid->cbid_arg;
95         ENTRY;
96
97         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
98
99         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
100         LASSERT (ev->md.start == req->rq_repbuf);
101         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
102         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
103            for adaptive timeouts' early reply. */
104         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
105
106         spin_lock(&req->rq_lock);
107
108         req->rq_receiving_reply = 0;
109         req->rq_early = 0;
110         if (ev->unlinked)
111                 req->rq_reply_unlinked = 1;
112
113         if (ev->status)
114                 goto out_wake;
115
116         if (ev->type == LNET_EVENT_UNLINK) {
117                 LASSERT(ev->unlinked);
118                 DEBUG_REQ(D_NET, req, "unlink");
119                 goto out_wake;
120         }
121
122         if (ev->mlength < ev->rlength ) {
123                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
124                        req->rq_replen, ev->rlength, ev->offset);
125                 req->rq_reply_truncated = 1;
126                 req->rq_replied = 1;
127                 req->rq_status = -EOVERFLOW;
128                 req->rq_nob_received = ev->rlength + ev->offset;
129                 goto out_wake;
130         }
131
132         if ((ev->offset == 0) &&
133             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
134                 /* Early reply */
135                 DEBUG_REQ(D_ADAPTTO, req,
136                           "Early reply received: mlen=%u offset=%d replen=%d "
137                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
138                           req->rq_replen, req->rq_replied, ev->unlinked);
139
140                 req->rq_early_count++; /* number received, client side */
141
142                 if (req->rq_replied)   /* already got the real reply */
143                         goto out_wake;
144
145                 req->rq_early = 1;
146                 req->rq_reply_off = ev->offset;
147                 req->rq_nob_received = ev->mlength;
148                 /* And we're still receiving */
149                 req->rq_receiving_reply = 1;
150         } else {
151                 /* Real reply */
152                 req->rq_rep_swab_mask = 0;
153                 req->rq_replied = 1;
154                 /* Got reply, no resend required */
155                 req->rq_resend = 0;
156                 req->rq_reply_off = ev->offset;
157                 req->rq_nob_received = ev->mlength;
158                 /* LNetMDUnlink can't be called under the LNET_LOCK,
159                    so we must unlink in ptlrpc_unregister_reply */
160                 DEBUG_REQ(D_INFO, req,
161                           "reply in flags=%x mlen=%u offset=%d replen=%d",
162                           lustre_msg_get_flags(req->rq_reqmsg),
163                           ev->mlength, ev->offset, req->rq_replen);
164         }
165
166         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
167
168 out_wake:
169         /* NB don't unlock till after wakeup; req can disappear under us
170          * since we don't have our own ref */
171         ptlrpc_client_wake_req(req);
172         spin_unlock(&req->rq_lock);
173         EXIT;
174 }
175
176 /*
177  * Client's bulk has been written/read
178  */
179 void client_bulk_callback (lnet_event_t *ev)
180 {
181         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
182         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
183         struct ptlrpc_request   *req;
184         ENTRY;
185
186         LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
187                  ev->type == LNET_EVENT_PUT) ||
188                 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
189                  ev->type == LNET_EVENT_GET) ||
190                 ev->type == LNET_EVENT_UNLINK);
191         LASSERT(ev->unlinked);
192
193         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
194                 ev->status = -EIO;
195
196         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
197                 ev->status = -EIO;
198
199         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
200                "event type %d, status %d, desc %p\n",
201                ev->type, ev->status, desc);
202
203         spin_lock(&desc->bd_lock);
204         req = desc->bd_req;
205         LASSERT(desc->bd_md_count > 0);
206         desc->bd_md_count--;
207
208         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
209                 desc->bd_nob_transferred += ev->mlength;
210                 desc->bd_sender = ev->sender;
211         } else {
212                 /* start reconnect and resend if network error hit */
213                 spin_lock(&req->rq_lock);
214                 req->rq_net_err = 1;
215                 spin_unlock(&req->rq_lock);
216         }
217
218         if (ev->status != 0)
219                 desc->bd_failure = 1;
220
221         /* NB don't unlock till after wakeup; desc can disappear under us
222          * otherwise */
223         if (desc->bd_md_count == 0)
224                 ptlrpc_client_wake_req(desc->bd_req);
225
226         spin_unlock(&desc->bd_lock);
227         EXIT;
228 }
229
230 /*
231  * We will have percpt request history list for ptlrpc service in upcoming
232  * patches because we don't want to be serialized by current per-service
233  * history operations. So we require history ID can (somehow) show arriving
234  * order w/o grabbing global lock, and user can sort them in userspace.
235  *
236  * This is how we generate history ID for ptlrpc_request:
237  * ----------------------------------------------------
238  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
239  * ----------------------------------------------------
240  * |  seconds  | usec / 16 |   sequence    | CPT id   |
241  * ----------------------------------------------------
242  *
243  * it might not be precise but should be good enough.
244  */
245
246 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
247
248 #define REQS_SEC_SHIFT          32
249 #define REQS_USEC_SHIFT         16
250 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
251
252 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
253                                    struct ptlrpc_request *req)
254 {
255         __u64   sec = req->rq_arrival_time.tv_sec;
256         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
257         __u64   new_seq;
258
259         /* set sequence ID for request and add it to history list,
260          * it must be called with hold svcpt::scp_lock */
261
262         new_seq = (sec << REQS_SEC_SHIFT) |
263                   (usec << REQS_USEC_SHIFT) |
264                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
265
266         if (new_seq > svcpt->scp_hist_seq) {
267                 /* This handles the initial case of scp_hist_seq == 0 or
268                  * we just jumped into a new time window */
269                 svcpt->scp_hist_seq = new_seq;
270         } else {
271                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
272                 /* NB: increase sequence number in current usec bucket,
273                  * however, it's possible that we used up all bits for
274                  * sequence and jumped into the next usec bucket (future time),
275                  * then we hope there will be less RPCs per bucket at some
276                  * point, and sequence will catch up again */
277                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
278                 new_seq = svcpt->scp_hist_seq;
279         }
280
281         req->rq_history_seq = new_seq;
282
283         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
284 }
285
286 /*
287  * Server's incoming request callback
288  */
289 void request_in_callback(lnet_event_t *ev)
290 {
291         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
292         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
293         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
294         struct ptlrpc_service             *service = svcpt->scp_service;
295         struct ptlrpc_request             *req;
296         ENTRY;
297
298         LASSERT (ev->type == LNET_EVENT_PUT ||
299                  ev->type == LNET_EVENT_UNLINK);
300         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
301         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
302                  rqbd->rqbd_buffer + service->srv_buf_size);
303
304         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
305                "event type %d, status %d, service %s\n",
306                ev->type, ev->status, service->srv_name);
307
308         if (ev->unlinked) {
309                 /* If this is the last request message to fit in the
310                  * request buffer we can use the request object embedded in
311                  * rqbd.  Note that if we failed to allocate a request,
312                  * we'd have to re-post the rqbd, which we can't do in this
313                  * context. */
314                 req = &rqbd->rqbd_req;
315                 memset(req, 0, sizeof (*req));
316         } else {
317                 LASSERT (ev->type == LNET_EVENT_PUT);
318                 if (ev->status != 0) {
319                         /* We moaned above already... */
320                         return;
321                 }
322                 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
323                 if (req == NULL) {
324                         CERROR("Can't allocate incoming request descriptor: "
325                                "Dropping %s RPC from %s\n",
326                                service->srv_name,
327                                libcfs_id2str(ev->initiator));
328                         return;
329                 }
330         }
331
332         ptlrpc_srv_req_init(req);
333         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
334          * flags are reset and scalars are zero.  We only set the message
335          * size to non-zero if this was a successful receive. */
336         req->rq_xid = ev->match_bits;
337         req->rq_reqbuf = ev->md.start + ev->offset;
338         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
339                 req->rq_reqdata_len = ev->mlength;
340         do_gettimeofday(&req->rq_arrival_time);
341         req->rq_peer = ev->initiator;
342         req->rq_self = ev->target.nid;
343         req->rq_rqbd = rqbd;
344         req->rq_phase = RQ_PHASE_NEW;
345         if (ev->type == LNET_EVENT_PUT)
346                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
347                        req, req->rq_xid, ev->mlength);
348
349         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
350
351         spin_lock(&svcpt->scp_lock);
352
353         ptlrpc_req_add_history(svcpt, req);
354
355         if (ev->unlinked) {
356                 svcpt->scp_nrqbds_posted--;
357                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
358                        svcpt->scp_nrqbds_posted);
359
360                 /* Normally, don't complain about 0 buffers posted; LNET won't
361                  * drop incoming reqs since we set the portal lazy */
362                 if (test_req_buffer_pressure &&
363                     ev->type != LNET_EVENT_UNLINK &&
364                     svcpt->scp_nrqbds_posted == 0)
365                         CWARN("All %s request buffers busy\n",
366                               service->srv_name);
367
368                 /* req takes over the network's ref on rqbd */
369         } else {
370                 /* req takes a ref on rqbd */
371                 rqbd->rqbd_refcount++;
372         }
373
374         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
375         svcpt->scp_nreqs_incoming++;
376
377         /* NB everything can disappear under us once the request
378          * has been queued and we unlock, so do the wake now... */
379         wake_up(&svcpt->scp_waitq);
380
381         spin_unlock(&svcpt->scp_lock);
382         EXIT;
383 }
384
385 /*
386  *  Server's outgoing reply callback
387  */
388 void reply_out_callback(lnet_event_t *ev)
389 {
390         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
391         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
392         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
393         ENTRY;
394
395         LASSERT (ev->type == LNET_EVENT_SEND ||
396                  ev->type == LNET_EVENT_ACK ||
397                  ev->type == LNET_EVENT_UNLINK);
398
399         if (!rs->rs_difficult) {
400                 /* 'Easy' replies have no further processing so I drop the
401                  * net's ref on 'rs' */
402                 LASSERT (ev->unlinked);
403                 ptlrpc_rs_decref(rs);
404                 EXIT;
405                 return;
406         }
407
408         LASSERT (rs->rs_on_net);
409
410         if (ev->unlinked) {
411                 /* Last network callback. The net's ref on 'rs' stays put
412                  * until ptlrpc_handle_rs() is done with it */
413                 spin_lock(&svcpt->scp_rep_lock);
414                 spin_lock(&rs->rs_lock);
415
416                 rs->rs_on_net = 0;
417                 if (!rs->rs_no_ack ||
418                     rs->rs_transno <=
419                     rs->rs_export->exp_obd->obd_last_committed)
420                         ptlrpc_schedule_difficult_reply(rs);
421
422                 spin_unlock(&rs->rs_lock);
423                 spin_unlock(&svcpt->scp_rep_lock);
424         }
425         EXIT;
426 }
427
428 #ifdef HAVE_SERVER_SUPPORT
429 /*
430  * Server's bulk completion callback
431  */
432 void server_bulk_callback (lnet_event_t *ev)
433 {
434         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
435         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
436         ENTRY;
437
438         LASSERT(ev->type == LNET_EVENT_SEND ||
439                 ev->type == LNET_EVENT_UNLINK ||
440                 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
441                  ev->type == LNET_EVENT_ACK) ||
442                 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
443                  ev->type == LNET_EVENT_REPLY));
444
445         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
446                "event type %d, status %d, desc %p\n",
447                ev->type, ev->status, desc);
448
449         spin_lock(&desc->bd_lock);
450
451         LASSERT(desc->bd_md_count > 0);
452
453         if ((ev->type == LNET_EVENT_ACK ||
454              ev->type == LNET_EVENT_REPLY) &&
455             ev->status == 0) {
456                 /* We heard back from the peer, so even if we get this
457                  * before the SENT event (oh yes we can), we know we
458                  * read/wrote the peer buffer and how much... */
459                 desc->bd_nob_transferred += ev->mlength;
460                 desc->bd_sender = ev->sender;
461         }
462
463         if (ev->status != 0)
464                 desc->bd_failure = 1;
465
466         if (ev->unlinked) {
467                 desc->bd_md_count--;
468                 /* This is the last callback no matter what... */
469                 if (desc->bd_md_count == 0)
470                         wake_up(&desc->bd_waitq);
471         }
472
473         spin_unlock(&desc->bd_lock);
474         EXIT;
475 }
476 #endif
477
478 static void ptlrpc_master_callback(lnet_event_t *ev)
479 {
480         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
481         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
482
483         /* Honestly, it's best to find out early. */
484         LASSERT (cbid->cbid_arg != LP_POISON);
485         LASSERT (callback == request_out_callback ||
486                  callback == reply_in_callback ||
487                  callback == client_bulk_callback ||
488                  callback == request_in_callback ||
489                  callback == reply_out_callback
490 #ifdef HAVE_SERVER_SUPPORT
491                  || callback == server_bulk_callback
492 #endif
493                  );
494
495         callback (ev);
496 }
497
498 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
499                          lnet_process_id_t *peer, lnet_nid_t *self)
500 {
501         int               best_dist = 0;
502         __u32             best_order = 0;
503         int               count = 0;
504         int               rc = -ENOENT;
505         int               dist;
506         __u32             order;
507         lnet_nid_t        dst_nid;
508         lnet_nid_t        src_nid;
509
510         peer->pid = LNET_PID_LUSTRE;
511
512         /* Choose the matching UUID that's closest */
513         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
514                 dist = LNetDist(dst_nid, &src_nid, &order);
515                 if (dist < 0)
516                         continue;
517
518                 if (dist == 0) {                /* local! use loopback LND */
519                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
520                         rc = 0;
521                         break;
522                 }
523
524                 if (rc < 0 ||
525                     dist < best_dist ||
526                     (dist == best_dist && order < best_order)) {
527                         best_dist = dist;
528                         best_order = order;
529
530                         peer->nid = dst_nid;
531                         *self = src_nid;
532                         rc = 0;
533                 }
534         }
535
536         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
537         return rc;
538 }
539
540 void ptlrpc_ni_fini(void)
541 {
542         wait_queue_head_t         waitq;
543         struct l_wait_info  lwi;
544         int                 rc;
545         int                 retries;
546
547         /* Wait for the event queue to become idle since there may still be
548          * messages in flight with pending events (i.e. the fire-and-forget
549          * messages == client requests and "non-difficult" server
550          * replies */
551
552         for (retries = 0;; retries++) {
553                 rc = LNetEQFree(ptlrpc_eq_h);
554                 switch (rc) {
555                 default:
556                         LBUG();
557
558                 case 0:
559                         LNetNIFini();
560                         return;
561
562                 case -EBUSY:
563                         if (retries != 0)
564                                 CWARN("Event queue still busy\n");
565
566                         /* Wait for a bit */
567                         init_waitqueue_head(&waitq);
568                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
569                         l_wait_event(waitq, 0, &lwi);
570                         break;
571                 }
572         }
573         /* notreached */
574 }
575
576 lnet_pid_t ptl_get_pid(void)
577 {
578         return LNET_PID_LUSTRE;
579 }
580
581 int ptlrpc_ni_init(void)
582 {
583         int              rc;
584         lnet_pid_t       pid;
585
586         pid = ptl_get_pid();
587         CDEBUG(D_NET, "My pid is: %x\n", pid);
588
589         /* We're not passing any limits yet... */
590         rc = LNetNIInit(pid);
591         if (rc < 0) {
592                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
593                 return rc;
594         }
595
596         /* CAVEAT EMPTOR: how we process portals events is _radically_
597          * different depending on... */
598         /* kernel LNet calls our master callback when there are new event,
599          * because we are guaranteed to get every event via callback,
600          * so we just set EQ size to 0 to avoid overhread of serializing
601          * enqueue/dequeue operations in LNet. */
602         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
603         if (rc == 0)
604                 return 0;
605
606         CERROR ("Failed to allocate event queue: %d\n", rc);
607         LNetNIFini();
608
609         return rc;
610 }
611
612
613 int ptlrpc_init_portals(void)
614 {
615         int   rc = ptlrpc_ni_init();
616
617         if (rc != 0) {
618                 CERROR("network initialisation failed\n");
619                 return rc;
620         }
621         rc = ptlrpcd_addref();
622         if (rc == 0)
623                 return 0;
624
625         CERROR("rpcd initialisation failed\n");
626         ptlrpc_ni_fini();
627         return rc;
628 }
629
630 void ptlrpc_exit_portals(void)
631 {
632         ptlrpcd_decref();
633         ptlrpc_ni_fini();
634 }