drivers/staging/lustre/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 "../../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  */
  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                 /* Failed send: make it seem like the reply timed out, just
  73                  * like failing sends in client.c does currently...
  74                  */
  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  */
  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          */
 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_unlink = 0;
 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_truncate = 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 replied=%d unlinked=%d",
 133                           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                 if (req->rq_replied)   /* already got the real reply */
 139                         goto out_wake;
 140
 141                 req->rq_early = 1;
 142                 req->rq_reply_off = ev->offset;
 143                 req->rq_nob_received = ev->mlength;
 144                 /* And we're still receiving */
 145                 req->rq_receiving_reply = 1;
 146         } else {
 147                 /* Real reply */
 148                 req->rq_rep_swab_mask = 0;
 149                 req->rq_replied = 1;
 150                 /* Got reply, no resend required */
 151                 req->rq_resend = 0;
 152                 req->rq_reply_off = ev->offset;
 153                 req->rq_nob_received = ev->mlength;
 154                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 155                  * so we must unlink in ptlrpc_unregister_reply
 156                  */
 157                 DEBUG_REQ(D_INFO, req,
 158                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 159                           lustre_msg_get_flags(req->rq_reqmsg),
 160                           ev->mlength, ev->offset, req->rq_replen);
 161         }
 162
 163         req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
 164
 165 out_wake:
 166         /* NB don't unlock till after wakeup; req can disappear under us
 167          * since we don't have our own ref
 168          */
 169         ptlrpc_client_wake_req(req);
 170         spin_unlock(&req->rq_lock);
 171 }
 172
 173 /*
 174  * Client's bulk has been written/read
 175  */
 176 void client_bulk_callback(lnet_event_t *ev)
 177 {
 178         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 179         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 180         struct ptlrpc_request *req;
 181
 182         LASSERT((desc->bd_type == BULK_PUT_SINK &&
 183                  ev->type == LNET_EVENT_PUT) ||
 184                 (desc->bd_type == BULK_GET_SOURCE &&
 185                  ev->type == LNET_EVENT_GET) ||
 186                 ev->type == LNET_EVENT_UNLINK);
 187         LASSERT(ev->unlinked);
 188
 189         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 190                 ev->status = -EIO;
 191
 192         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
 193                                  CFS_FAIL_ONCE))
 194                 ev->status = -EIO;
 195
 196         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 197                "event type %d, status %d, desc %p\n",
 198                ev->type, ev->status, desc);
 199
 200         spin_lock(&desc->bd_lock);
 201         req = desc->bd_req;
 202         LASSERT(desc->bd_md_count > 0);
 203         desc->bd_md_count--;
 204
 205         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 206                 desc->bd_nob_transferred += ev->mlength;
 207                 desc->bd_sender = ev->sender;
 208         } else {
 209                 /* start reconnect and resend if network error hit */
 210                 spin_lock(&req->rq_lock);
 211                 req->rq_net_err = 1;
 212                 spin_unlock(&req->rq_lock);
 213         }
 214
 215         if (ev->status != 0)
 216                 desc->bd_failure = 1;
 217
 218         /* NB don't unlock till after wakeup; desc can disappear under us
 219          * otherwise
 220          */
 221         if (desc->bd_md_count == 0)
 222                 ptlrpc_client_wake_req(desc->bd_req);
 223
 224         spin_unlock(&desc->bd_lock);
 225 }
 226
 227 /*
 228  * We will have percpt request history list for ptlrpc service in upcoming
 229  * patches because we don't want to be serialized by current per-service
 230  * history operations. So we require history ID can (somehow) show arriving
 231  * order w/o grabbing global lock, and user can sort them in userspace.
 232  *
 233  * This is how we generate history ID for ptlrpc_request:
 234  * ----------------------------------------------------
 235  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 236  * ----------------------------------------------------
 237  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 238  * ----------------------------------------------------
 239  *
 240  * it might not be precise but should be good enough.
 241  */
 242
 243 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 244
 245 #define REQS_SEC_SHIFT          32
 246 #define REQS_USEC_SHIFT         16
 247 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 248
 249 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 250                                    struct ptlrpc_request *req)
 251 {
 252         __u64 sec = req->rq_arrival_time.tv_sec;
 253         __u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
 254         __u64 new_seq;
 255
 256         /* set sequence ID for request and add it to history list,
 257          * it must be called with hold svcpt::scp_lock
 258          */
 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                  */
 268                 svcpt->scp_hist_seq = new_seq;
 269         } else {
 270                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 271                 /* NB: increase sequence number in current usec bucket,
 272                  * however, it's possible that we used up all bits for
 273                  * sequence and jumped into the next usec bucket (future time),
 274                  * then we hope there will be less RPCs per bucket at some
 275                  * point, and sequence will catch up again
 276                  */
 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
 297         LASSERT(ev->type == LNET_EVENT_PUT ||
 298                 ev->type == LNET_EVENT_UNLINK);
 299         LASSERT((char *)ev->md.start >= rqbd->rqbd_buffer);
 300         LASSERT((char *)ev->md.start + ev->offset + ev->mlength <=
 301                 rqbd->rqbd_buffer + service->srv_buf_size);
 302
 303         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 304                "event type %d, status %d, service %s\n",
 305                ev->type, ev->status, service->srv_name);
 306
 307         if (ev->unlinked) {
 308                 /* If this is the last request message to fit in the
 309                  * request buffer we can use the request object embedded in
 310                  * rqbd.  Note that if we failed to allocate a request,
 311                  * we'd have to re-post the rqbd, which we can't do in this
 312                  * context.
 313                  */
 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) {
 324                         CERROR("Can't allocate incoming request descriptor: Dropping %s RPC from %s\n",
 325                                service->srv_name,
 326                                libcfs_id2str(ev->initiator));
 327                         return;
 328                 }
 329         }
 330
 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          */
 335         req->rq_xid = ev->match_bits;
 336         req->rq_reqbuf = ev->md.start + ev->offset;
 337         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 338                 req->rq_reqdata_len = ev->mlength;
 339         ktime_get_real_ts64(&req->rq_arrival_time);
 340         req->rq_peer = ev->initiator;
 341         req->rq_self = ev->target.nid;
 342         req->rq_rqbd = rqbd;
 343         req->rq_phase = RQ_PHASE_NEW;
 344         spin_lock_init(&req->rq_lock);
 345         INIT_LIST_HEAD(&req->rq_timed_list);
 346         INIT_LIST_HEAD(&req->rq_exp_list);
 347         atomic_set(&req->rq_refcount, 1);
 348         if (ev->type == LNET_EVENT_PUT)
 349                 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
 350                        req, req->rq_xid, ev->mlength);
 351
 352         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 353
 354         spin_lock(&svcpt->scp_lock);
 355
 356         ptlrpc_req_add_history(svcpt, req);
 357
 358         if (ev->unlinked) {
 359                 svcpt->scp_nrqbds_posted--;
 360                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 361                        svcpt->scp_nrqbds_posted);
 362
 363                 /* Normally, don't complain about 0 buffers posted; LNET won't
 364                  * drop incoming reqs since we set the portal lazy
 365                  */
 366                 if (test_req_buffer_pressure &&
 367                     ev->type != LNET_EVENT_UNLINK &&
 368                     svcpt->scp_nrqbds_posted == 0)
 369                         CWARN("All %s request buffers busy\n",
 370                               service->srv_name);
 371
 372                 /* req takes over the network's ref on rqbd */
 373         } else {
 374                 /* req takes a ref on rqbd */
 375                 rqbd->rqbd_refcount++;
 376         }
 377
 378         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 379         svcpt->scp_nreqs_incoming++;
 380
 381         /* NB everything can disappear under us once the request
 382          * has been queued and we unlock, so do the wake now...
 383          */
 384         wake_up(&svcpt->scp_waitq);
 385
 386         spin_unlock(&svcpt->scp_lock);
 387 }
 388
 389 /*
 390  *  Server's outgoing reply callback
 391  */
 392 void reply_out_callback(lnet_event_t *ev)
 393 {
 394         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 395         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 396         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 397
 398         LASSERT(ev->type == LNET_EVENT_SEND ||
 399                 ev->type == LNET_EVENT_ACK ||
 400                 ev->type == LNET_EVENT_UNLINK);
 401
 402         if (!rs->rs_difficult) {
 403                 /* 'Easy' replies have no further processing so I drop the
 404                  * net's ref on 'rs'
 405                  */
 406                 LASSERT(ev->unlinked);
 407                 ptlrpc_rs_decref(rs);
 408                 return;
 409         }
 410
 411         LASSERT(rs->rs_on_net);
 412
 413         if (ev->unlinked) {
 414                 /* Last network callback. The net's ref on 'rs' stays put
 415                  * until ptlrpc_handle_rs() is done with it
 416                  */
 417                 spin_lock(&svcpt->scp_rep_lock);
 418                 spin_lock(&rs->rs_lock);
 419
 420                 rs->rs_on_net = 0;
 421                 if (!rs->rs_no_ack ||
 422                     rs->rs_transno <=
 423                     rs->rs_export->exp_obd->obd_last_committed)
 424                         ptlrpc_schedule_difficult_reply(rs);
 425
 426                 spin_unlock(&rs->rs_lock);
 427                 spin_unlock(&svcpt->scp_rep_lock);
 428         }
 429 }
 430
 431 static void ptlrpc_master_callback(lnet_event_t *ev)
 432 {
 433         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 434         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 435
 436         /* Honestly, it's best to find out early. */
 437         LASSERT(cbid->cbid_arg != LP_POISON);
 438         LASSERT(callback == request_out_callback ||
 439                 callback == reply_in_callback ||
 440                 callback == client_bulk_callback ||
 441                 callback == request_in_callback ||
 442                 callback == reply_out_callback);
 443
 444         callback(ev);
 445 }
 446
 447 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
 448                         lnet_process_id_t *peer, lnet_nid_t *self)
 449 {
 450         int best_dist = 0;
 451         __u32 best_order = 0;
 452         int count = 0;
 453         int rc = -ENOENT;
 454         int dist;
 455         __u32 order;
 456         lnet_nid_t dst_nid;
 457         lnet_nid_t src_nid;
 458
 459         peer->pid = LNET_PID_LUSTRE;
 460
 461         /* Choose the matching UUID that's closest */
 462         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 463                 dist = LNetDist(dst_nid, &src_nid, &order);
 464                 if (dist < 0)
 465                         continue;
 466
 467                 if (dist == 0) {                /* local! use loopback LND */
 468                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 469                         rc = 0;
 470                         break;
 471                 }
 472
 473                 if (rc < 0 ||
 474                     dist < best_dist ||
 475                     (dist == best_dist && order < best_order)) {
 476                         best_dist = dist;
 477                         best_order = order;
 478
 479                         peer->nid = dst_nid;
 480                         *self = src_nid;
 481                         rc = 0;
 482                 }
 483         }
 484
 485         CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 486         return rc;
 487 }
 488
 489 static void ptlrpc_ni_fini(void)
 490 {
 491         wait_queue_head_t waitq;
 492         struct l_wait_info lwi;
 493         int rc;
 494         int retries;
 495
 496         /* Wait for the event queue to become idle since there may still be
 497          * messages in flight with pending events (i.e. the fire-and-forget
 498          * messages == client requests and "non-difficult" server
 499          * replies
 500          */
 501
 502         for (retries = 0;; retries++) {
 503                 rc = LNetEQFree(ptlrpc_eq_h);
 504                 switch (rc) {
 505                 default:
 506                         LBUG();
 507
 508                 case 0:
 509                         LNetNIFini();
 510                         return;
 511
 512                 case -EBUSY:
 513                         if (retries != 0)
 514                                 CWARN("Event queue still busy\n");
 515
 516                         /* Wait for a bit */
 517                         init_waitqueue_head(&waitq);
 518                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 519                         l_wait_event(waitq, 0, &lwi);
 520                         break;
 521                 }
 522         }
 523         /* notreached */
 524 }
 525
 526 static lnet_pid_t ptl_get_pid(void)
 527 {
 528         lnet_pid_t pid;
 529
 530         pid = LNET_PID_LUSTRE;
 531         return pid;
 532 }
 533
 534 static int ptlrpc_ni_init(void)
 535 {
 536         int rc;
 537         lnet_pid_t pid;
 538
 539         pid = ptl_get_pid();
 540         CDEBUG(D_NET, "My pid is: %x\n", pid);
 541
 542         /* We're not passing any limits yet... */
 543         rc = LNetNIInit(pid);
 544         if (rc < 0) {
 545                 CDEBUG(D_NET, "Can't init network interface: %d\n", rc);
 546                 return -ENOENT;
 547         }
 548
 549         /* CAVEAT EMPTOR: how we process portals events is _radically_
 550          * different depending on...
 551          */
 552         /* kernel LNet calls our master callback when there are new event,
 553          * because we are guaranteed to get every event via callback,
 554          * so we just set EQ size to 0 to avoid overhead of serializing
 555          * enqueue/dequeue operations in LNet.
 556          */
 557         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 558         if (rc == 0)
 559                 return 0;
 560
 561         CERROR("Failed to allocate event queue: %d\n", rc);
 562         LNetNIFini();
 563
 564         return -ENOMEM;
 565 }
 566
 567 int ptlrpc_init_portals(void)
 568 {
 569         int rc = ptlrpc_ni_init();
 570
 571         if (rc != 0) {
 572                 CERROR("network initialisation failed\n");
 573                 return -EIO;
 574         }
 575         rc = ptlrpcd_addref();
 576         if (rc == 0)
 577                 return 0;
 578
 579         CERROR("rpcd initialisation failed\n");
 580         ptlrpc_ni_fini();
 581         return rc;
 582 }
 583
 584 void ptlrpc_exit_portals(void)
 585 {
 586         ptlrpcd_decref();
 587         ptlrpc_ni_fini();
 588 }