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, 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 <linux/libcfs/libcfs.h>
  40 # ifdef __mips64__
  41 #  include <linux/kernel.h>
  42 # endif
  43
  44 #include <obd_class.h>
  45 #include <lustre_net.h>
  46 #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         req->rq_real_sent = cfs_time_current_sec();
  67
  68         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  69
  70                 /* Failed send: make it seem like the reply timed out, just
  71                  * like failing sends in client.c does currently...  */
  72
  73                 spin_lock(&req->rq_lock);
  74                 req->rq_net_err = 1;
  75                 spin_unlock(&req->rq_lock);
  76
  77                 ptlrpc_client_wake_req(req);
  78         }
  79
  80         ptlrpc_req_finished(req);
  81 }
  82
  83 /*
  84  * Client's incoming reply callback
  85  */
  86 void reply_in_callback(lnet_event_t *ev)
  87 {
  88         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  89         struct ptlrpc_request *req = cbid->cbid_arg;
  90
  91         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  92
  93         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
  94         LASSERT (ev->md.start == req->rq_repbuf);
  95         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
  96         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
  97            for adaptive timeouts' early reply. */
  98         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
  99
 100         spin_lock(&req->rq_lock);
 101
 102         req->rq_receiving_reply = 0;
 103         req->rq_early = 0;
 104         if (ev->unlinked)
 105                 req->rq_must_unlink = 0;
 106
 107         if (ev->status)
 108                 goto out_wake;
 109
 110         if (ev->type == LNET_EVENT_UNLINK) {
 111                 LASSERT(ev->unlinked);
 112                 DEBUG_REQ(D_NET, req, "unlink");
 113                 goto out_wake;
 114         }
 115
 116         if (ev->mlength < ev->rlength ) {
 117                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 118                        req->rq_replen, ev->rlength, ev->offset);
 119                 req->rq_reply_truncate = 1;
 120                 req->rq_replied = 1;
 121                 req->rq_status = -EOVERFLOW;
 122                 req->rq_nob_received = ev->rlength + ev->offset;
 123                 goto out_wake;
 124         }
 125
 126         if ((ev->offset == 0) &&
 127             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 128                 /* Early reply */
 129                 DEBUG_REQ(D_ADAPTTO, req,
 130                           "Early reply received: mlen=%u offset=%d replen=%d "
 131                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 132                           req->rq_replen, req->rq_replied, ev->unlinked);
 133
 134                 req->rq_early_count++; /* number received, client side */
 135
 136                 if (req->rq_replied)   /* already got the real reply */
 137                         goto out_wake;
 138
 139                 req->rq_early = 1;
 140                 req->rq_reply_off = ev->offset;
 141                 req->rq_nob_received = ev->mlength;
 142                 /* And we're still receiving */
 143                 req->rq_receiving_reply = 1;
 144         } else {
 145                 /* Real reply */
 146                 req->rq_rep_swab_mask = 0;
 147                 req->rq_replied = 1;
 148                 req->rq_reply_off = ev->offset;
 149                 req->rq_nob_received = ev->mlength;
 150                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 151                    so we must unlink in ptlrpc_unregister_reply */
 152                 DEBUG_REQ(D_INFO, req,
 153                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 154                           lustre_msg_get_flags(req->rq_reqmsg),
 155                           ev->mlength, ev->offset, req->rq_replen);
 156         }
 157
 158         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 159
 160 out_wake:
 161         /* NB don't unlock till after wakeup; req can disappear under us
 162          * since we don't have our own ref */
 163         ptlrpc_client_wake_req(req);
 164         spin_unlock(&req->rq_lock);
 165 }
 166
 167 /*
 168  * Client's bulk has been written/read
 169  */
 170 void client_bulk_callback (lnet_event_t *ev)
 171 {
 172         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 173         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 174         struct ptlrpc_request   *req;
 175
 176         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 177                   ev->type == LNET_EVENT_PUT) ||
 178                  (desc->bd_type == BULK_GET_SOURCE &&
 179                   ev->type == LNET_EVENT_GET) ||
 180                  ev->type == LNET_EVENT_UNLINK);
 181         LASSERT (ev->unlinked);
 182
 183         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 184                 ev->status = -EIO;
 185
 186         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 187                 ev->status = -EIO;
 188
 189         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 190                "event type %d, status %d, desc %p\n",
 191                ev->type, ev->status, desc);
 192
 193         spin_lock(&desc->bd_lock);
 194         req = desc->bd_req;
 195         LASSERT(desc->bd_md_count > 0);
 196         desc->bd_md_count--;
 197
 198         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 199                 desc->bd_nob_transferred += ev->mlength;
 200                 desc->bd_sender = ev->sender;
 201         } else {
 202                 /* start reconnect and resend if network error hit */
 203                 spin_lock(&req->rq_lock);
 204                 req->rq_net_err = 1;
 205                 spin_unlock(&req->rq_lock);
 206         }
 207
 208         if (ev->status != 0)
 209                 desc->bd_failure = 1;
 210
 211         /* NB don't unlock till after wakeup; desc can disappear under us
 212          * otherwise */
 213         if (desc->bd_md_count == 0)
 214                 ptlrpc_client_wake_req(desc->bd_req);
 215
 216         spin_unlock(&desc->bd_lock);
 217 }
 218
 219 /*
 220  * We will have percpt request history list for ptlrpc service in upcoming
 221  * patches because we don't want to be serialized by current per-service
 222  * history operations. So we require history ID can (somehow) show arriving
 223  * order w/o grabbing global lock, and user can sort them in userspace.
 224  *
 225  * This is how we generate history ID for ptlrpc_request:
 226  * ----------------------------------------------------
 227  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 228  * ----------------------------------------------------
 229  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 230  * ----------------------------------------------------
 231  *
 232  * it might not be precise but should be good enough.
 233  */
 234
 235 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 236
 237 #define REQS_SEC_SHIFT          32
 238 #define REQS_USEC_SHIFT         16
 239 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 240
 241 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 242                                    struct ptlrpc_request *req)
 243 {
 244         __u64   sec = req->rq_arrival_time.tv_sec;
 245         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
 246         __u64   new_seq;
 247
 248         /* set sequence ID for request and add it to history list,
 249          * it must be called with hold svcpt::scp_lock */
 250
 251         new_seq = (sec << REQS_SEC_SHIFT) |
 252                   (usec << REQS_USEC_SHIFT) |
 253                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
 254
 255         if (new_seq > svcpt->scp_hist_seq) {
 256                 /* This handles the initial case of scp_hist_seq == 0 or
 257                  * we just jumped into a new time window */
 258                 svcpt->scp_hist_seq = new_seq;
 259         } else {
 260                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 261                 /* NB: increase sequence number in current usec bucket,
 262                  * however, it's possible that we used up all bits for
 263                  * sequence and jumped into the next usec bucket (future time),
 264                  * then we hope there will be less RPCs per bucket at some
 265                  * point, and sequence will catch up again */
 266                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
 267                 new_seq = svcpt->scp_hist_seq;
 268         }
 269
 270         req->rq_history_seq = new_seq;
 271
 272         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 273 }
 274
 275 /*
 276  * Server's incoming request callback
 277  */
 278 void request_in_callback(lnet_event_t *ev)
 279 {
 280         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 281         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 282         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 283         struct ptlrpc_service        *service = svcpt->scp_service;
 284         struct ptlrpc_request        *req;
 285
 286         LASSERT (ev->type == LNET_EVENT_PUT ||
 287                  ev->type == LNET_EVENT_UNLINK);
 288         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 289         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 290                  rqbd->rqbd_buffer + service->srv_buf_size);
 291
 292         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 293                "event type %d, status %d, service %s\n",
 294                ev->type, ev->status, service->srv_name);
 295
 296         if (ev->unlinked) {
 297                 /* If this is the last request message to fit in the
 298                  * request buffer we can use the request object embedded in
 299                  * rqbd.  Note that if we failed to allocate a request,
 300                  * we'd have to re-post the rqbd, which we can't do in this
 301                  * context. */
 302                 req = &rqbd->rqbd_req;
 303                 memset(req, 0, sizeof (*req));
 304         } else {
 305                 LASSERT (ev->type == LNET_EVENT_PUT);
 306                 if (ev->status != 0) {
 307                         /* We moaned above already... */
 308                         return;
 309                 }
 310                 OBD_ALLOC_GFP(req, sizeof(*req), ALLOC_ATOMIC_TRY);
 311                 if (req == NULL) {
 312                         CERROR("Can't allocate incoming request descriptor: "
 313                                "Dropping %s RPC from %s\n",
 314                                service->srv_name,
 315                                libcfs_id2str(ev->initiator));
 316                         return;
 317                 }
 318         }
 319
 320         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 321          * flags are reset and scalars are zero.  We only set the message
 322          * size to non-zero if this was a successful receive. */
 323         req->rq_xid = ev->match_bits;
 324         req->rq_reqbuf = ev->md.start + ev->offset;
 325         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 326                 req->rq_reqdata_len = ev->mlength;
 327         do_gettimeofday(&req->rq_arrival_time);
 328         req->rq_peer = ev->initiator;
 329         req->rq_self = ev->target.nid;
 330         req->rq_rqbd = rqbd;
 331         req->rq_phase = RQ_PHASE_NEW;
 332         spin_lock_init(&req->rq_lock);
 333         INIT_LIST_HEAD(&req->rq_timed_list);
 334         INIT_LIST_HEAD(&req->rq_exp_list);
 335         atomic_set(&req->rq_refcount, 1);
 336         if (ev->type == LNET_EVENT_PUT)
 337                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
 338                        req, req->rq_xid, ev->mlength);
 339
 340         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 341
 342         spin_lock(&svcpt->scp_lock);
 343
 344         ptlrpc_req_add_history(svcpt, req);
 345
 346         if (ev->unlinked) {
 347                 svcpt->scp_nrqbds_posted--;
 348                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 349                        svcpt->scp_nrqbds_posted);
 350
 351                 /* Normally, don't complain about 0 buffers posted; LNET won't
 352                  * drop incoming reqs since we set the portal lazy */
 353                 if (test_req_buffer_pressure &&
 354                     ev->type != LNET_EVENT_UNLINK &&
 355                     svcpt->scp_nrqbds_posted == 0)
 356                         CWARN("All %s request buffers busy\n",
 357                               service->srv_name);
 358
 359                 /* req takes over the network's ref on rqbd */
 360         } else {
 361                 /* req takes a ref on rqbd */
 362                 rqbd->rqbd_refcount++;
 363         }
 364
 365         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 366         svcpt->scp_nreqs_incoming++;
 367
 368         /* NB everything can disappear under us once the request
 369          * has been queued and we unlock, so do the wake now... */
 370         wake_up(&svcpt->scp_waitq);
 371
 372         spin_unlock(&svcpt->scp_lock);
 373 }
 374
 375 /*
 376  *  Server's outgoing reply callback
 377  */
 378 void reply_out_callback(lnet_event_t *ev)
 379 {
 380         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 381         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 382         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 383
 384         LASSERT (ev->type == LNET_EVENT_SEND ||
 385                  ev->type == LNET_EVENT_ACK ||
 386                  ev->type == LNET_EVENT_UNLINK);
 387
 388         if (!rs->rs_difficult) {
 389                 /* 'Easy' replies have no further processing so I drop the
 390                  * net's ref on 'rs' */
 391                 LASSERT (ev->unlinked);
 392                 ptlrpc_rs_decref(rs);
 393                 return;
 394         }
 395
 396         LASSERT (rs->rs_on_net);
 397
 398         if (ev->unlinked) {
 399                 /* Last network callback. The net's ref on 'rs' stays put
 400                  * until ptlrpc_handle_rs() is done with it */
 401                 spin_lock(&svcpt->scp_rep_lock);
 402                 spin_lock(&rs->rs_lock);
 403
 404                 rs->rs_on_net = 0;
 405                 if (!rs->rs_no_ack ||
 406                     rs->rs_transno <=
 407                     rs->rs_export->exp_obd->obd_last_committed)
 408                         ptlrpc_schedule_difficult_reply(rs);
 409
 410                 spin_unlock(&rs->rs_lock);
 411                 spin_unlock(&svcpt->scp_rep_lock);
 412         }
 413 }
 414
 415
 416 static void ptlrpc_master_callback(lnet_event_t *ev)
 417 {
 418         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 419         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 420
 421         /* Honestly, it's best to find out early. */
 422         LASSERT (cbid->cbid_arg != LP_POISON);
 423         LASSERT (callback == request_out_callback ||
 424                  callback == reply_in_callback ||
 425                  callback == client_bulk_callback ||
 426                  callback == request_in_callback ||
 427                  callback == reply_out_callback
 428                  );
 429
 430         callback (ev);
 431 }
 432
 433 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 434                          lnet_process_id_t *peer, lnet_nid_t *self)
 435 {
 436         int            best_dist = 0;
 437         __u32        best_order = 0;
 438         int            count = 0;
 439         int            rc = -ENOENT;
 440         int            portals_compatibility;
 441         int            dist;
 442         __u32        order;
 443         lnet_nid_t      dst_nid;
 444         lnet_nid_t      src_nid;
 445
 446         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 447
 448         peer->pid = LUSTRE_SRV_LNET_PID;
 449
 450         /* Choose the matching UUID that's closest */
 451         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 452                 dist = LNetDist(dst_nid, &src_nid, &order);
 453                 if (dist < 0)
 454                         continue;
 455
 456                 if (dist == 0) {                /* local! use loopback LND */
 457                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 458                         rc = 0;
 459                         break;
 460                 }
 461
 462                 if (rc < 0 ||
 463                     dist < best_dist ||
 464                     (dist == best_dist && order < best_order)) {
 465                         best_dist = dist;
 466                         best_order = order;
 467
 468                         if (portals_compatibility > 1) {
 469                                 /* Strong portals compatibility: Zero the nid's
 470                                  * NET, so if I'm reading new config logs, or
 471                                  * getting configured by (new) lconf I can
 472                                  * still talk to old servers. */
 473                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 474                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 475                         }
 476                         peer->nid = dst_nid;
 477                         *self = src_nid;
 478                         rc = 0;
 479                 }
 480         }
 481
 482         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 483         return rc;
 484 }
 485
 486 void ptlrpc_ni_fini(void)
 487 {
 488         wait_queue_head_t        waitq;
 489         struct l_wait_info  lwi;
 490         int              rc;
 491         int              retries;
 492
 493         /* Wait for the event queue to become idle since there may still be
 494          * messages in flight with pending events (i.e. the fire-and-forget
 495          * messages == client requests and "non-difficult" server
 496          * replies */
 497
 498         for (retries = 0;; retries++) {
 499                 rc = LNetEQFree(ptlrpc_eq_h);
 500                 switch (rc) {
 501                 default:
 502                         LBUG();
 503
 504                 case 0:
 505                         LNetNIFini();
 506                         return;
 507
 508                 case -EBUSY:
 509                         if (retries != 0)
 510                                 CWARN("Event queue still busy\n");
 511
 512                         /* Wait for a bit */
 513                         init_waitqueue_head(&waitq);
 514                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 515                         l_wait_event(waitq, 0, &lwi);
 516                         break;
 517                 }
 518         }
 519         /* notreached */
 520 }
 521
 522 lnet_pid_t ptl_get_pid(void)
 523 {
 524         lnet_pid_t      pid;
 525
 526         pid = LUSTRE_SRV_LNET_PID;
 527         return pid;
 528 }
 529
 530 int ptlrpc_ni_init(void)
 531 {
 532         int           rc;
 533         lnet_pid_t       pid;
 534
 535         pid = ptl_get_pid();
 536         CDEBUG(D_NET, "My pid is: %x\n", pid);
 537
 538         /* We're not passing any limits yet... */
 539         rc = LNetNIInit(pid);
 540         if (rc < 0) {
 541                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 542                 return (-ENOENT);
 543         }
 544
 545         /* CAVEAT EMPTOR: how we process portals events is _radically_
 546          * different depending on... */
 547         /* kernel LNet calls our master callback when there are new event,
 548          * because we are guaranteed to get every event via callback,
 549          * so we just set EQ size to 0 to avoid overhread of serializing
 550          * enqueue/dequeue operations in LNet. */
 551         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 552         if (rc == 0)
 553                 return 0;
 554
 555         CERROR ("Failed to allocate event queue: %d\n", rc);
 556         LNetNIFini();
 557
 558         return (-ENOMEM);
 559 }
 560
 561
 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
 579 void ptlrpc_exit_portals(void)
 580 {
 581         ptlrpcd_decref();
 582         ptlrpc_ni_fini();
 583 }