lustre/ptlrpc/events.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_RPC
  24
  25 #ifdef __KERNEL__
  26 #include <linux/module.h>
  27 #else
  28 #include <liblustre.h>
  29 #endif
  30 #include <linux/obd_class.h>
  31 #include <linux/lustre_net.h>
  32 #include "ptlrpc_internal.h"
  33
  34 struct ptlrpc_ni  ptlrpc_interfaces[NAL_MAX_NR];
  35 int               ptlrpc_ninterfaces;
  36
  37 /*
  38  *  Client's outgoing request callback
  39  */
  40 void request_out_callback(ptl_event_t *ev)
  41 {
  42         struct ptlrpc_cb_id   *cbid = ev->mem_desc.user_ptr;
  43         struct ptlrpc_request *req = cbid->cbid_arg;
  44         unsigned long          flags;
  45         ENTRY;
  46
  47         LASSERT (ev->type == PTL_EVENT_SEND_END ||
  48                  ev->type == PTL_EVENT_UNLINK);
  49         LASSERT (ev->unlinked);
  50
  51         DEBUG_REQ((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, req,
  52                   "type %d, status %d", ev->type, ev->ni_fail_type);
  53
  54         if (ev->type == PTL_EVENT_UNLINK ||
  55             ev->ni_fail_type != PTL_NI_OK) {
  56
  57                 /* Failed send: make it seem like the reply timed out, just
  58                  * like failing sends in client.c does currently...  */
  59
  60                 spin_lock_irqsave(&req->rq_lock, flags);
  61                 req->rq_timeout = 0;
  62                 spin_unlock_irqrestore(&req->rq_lock, flags);
  63
  64                 ptlrpc_wake_client_req(req);
  65         }
  66
  67         /* this balances the atomic_inc in ptl_send_rpc() */
  68         ptlrpc_req_finished(req);
  69         EXIT;
  70 }
  71
  72 /*
  73  * Client's incoming reply callback
  74  */
  75 void reply_in_callback(ptl_event_t *ev)
  76 {
  77         struct ptlrpc_cb_id   *cbid = ev->mem_desc.user_ptr;
  78         struct ptlrpc_request *req = cbid->cbid_arg;
  79         unsigned long flags;
  80         ENTRY;
  81
  82         LASSERT (ev->type == PTL_EVENT_PUT_END ||
  83                  ev->type == PTL_EVENT_UNLINK);
  84         LASSERT (ev->unlinked);
  85         LASSERT (ev->mem_desc.start == req->rq_repmsg);
  86         LASSERT (ev->offset == 0);
  87         LASSERT (ev->mlength <= req->rq_replen);
  88
  89         DEBUG_REQ((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, req,
  90                   "type %d, status %d", ev->type, ev->ni_fail_type);
  91
  92         spin_lock_irqsave (&req->rq_lock, flags);
  93
  94         LASSERT (req->rq_receiving_reply);
  95         req->rq_receiving_reply = 0;
  96
  97         if (ev->type == PTL_EVENT_PUT_END &&
  98             ev->ni_fail_type == PTL_NI_OK) {
  99                 req->rq_replied = 1;
 100                 req->rq_nob_received = ev->mlength;
 101         }
 102
 103         /* NB don't unlock till after wakeup; req can disappear under us
 104          * since we don't have our own ref */
 105         ptlrpc_wake_client_req(req);
 106
 107         spin_unlock_irqrestore (&req->rq_lock, flags);
 108         EXIT;
 109 }
 110
 111 /*
 112  * Client's bulk has been written/read
 113  */
 114 void client_bulk_callback (ptl_event_t *ev)
 115 {
 116         struct ptlrpc_cb_id     *cbid = ev->mem_desc.user_ptr;
 117         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 118         unsigned long            flags;
 119         ENTRY;
 120
 121         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 122                   ev->type == PTL_EVENT_PUT_END) ||
 123                  (desc->bd_type == BULK_GET_SOURCE &&
 124                   ev->type == PTL_EVENT_GET_END) ||
 125                  ev->type == PTL_EVENT_UNLINK);
 126         LASSERT (ev->unlinked);
 127
 128         CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR,
 129                "event type %d, status %d, desc %p\n",
 130                ev->type, ev->ni_fail_type, desc);
 131
 132         spin_lock_irqsave (&desc->bd_lock, flags);
 133
 134         LASSERT(desc->bd_network_rw);
 135         desc->bd_network_rw = 0;
 136
 137         if (ev->type != PTL_EVENT_UNLINK &&
 138             ev->ni_fail_type == PTL_NI_OK) {
 139                 desc->bd_success = 1;
 140                 desc->bd_nob_transferred = ev->mlength;
 141         }
 142
 143         /* NB don't unlock till after wakeup; desc can disappear under us
 144          * otherwise */
 145         ptlrpc_wake_client_req(desc->bd_req);
 146
 147         spin_unlock_irqrestore (&desc->bd_lock, flags);
 148         EXIT;
 149 }
 150
 151 /*
 152  * Server's incoming request callback
 153  */
 154 void request_in_callback(ptl_event_t *ev)
 155 {
 156         struct ptlrpc_cb_id               *cbid = ev->mem_desc.user_ptr;
 157         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 158         struct ptlrpc_srv_ni              *srv_ni = rqbd->rqbd_srv_ni;
 159         struct ptlrpc_service             *service = srv_ni->sni_service;
 160         struct ptlrpc_request             *req;
 161         long                               flags;
 162         ENTRY;
 163
 164         LASSERT (ev->type == PTL_EVENT_PUT_END ||
 165                  ev->type == PTL_EVENT_UNLINK);
 166         LASSERT ((char *)ev->mem_desc.start >= rqbd->rqbd_buffer);
 167         LASSERT ((char *)ev->mem_desc.start + ev->offset + ev->mlength <=
 168                  rqbd->rqbd_buffer + service->srv_buf_size);
 169
 170         CDEBUG((ev->ni_fail_type == PTL_OK) ? D_NET : D_ERROR,
 171                "event type %d, status %d, service %s\n",
 172                ev->type, ev->ni_fail_type, service->srv_name);
 173
 174         if (ev->unlinked) {
 175                 /* If this is the last request message to fit in the
 176                  * request buffer we can use the request object embedded in
 177                  * rqbd.  Note that if we failed to allocate a request,
 178                  * we'd have to re-post the rqbd, which we can't do in this
 179                  * context. */
 180                 req = &rqbd->rqbd_req;
 181                 memset(req, 0, sizeof (*req));
 182         } else {
 183                 LASSERT (ev->type == PTL_EVENT_PUT_END);
 184                 if (ev->ni_fail_type != PTL_NI_OK) {
 185                         /* We moaned above already... */
 186                         return;
 187                 }
 188                 OBD_ALLOC_GFP(req, sizeof(*req), GFP_ATOMIC);
 189                 if (req == NULL) {
 190                         CERROR("Can't allocate incoming request descriptor: "
 191                                "Dropping %s RPC from "LPX64"\n",
 192                                service->srv_name, ev->initiator.nid);
 193                         return;
 194                 }
 195         }
 196
 197         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 198          * flags are reset and scalars are zero.  We only set the message
 199          * size to non-zero if this was a successful receive. */
 200         req->rq_xid = ev->match_bits;
 201         req->rq_reqmsg = ev->mem_desc.start + ev->offset;
 202         if (ev->type == PTL_EVENT_PUT_END &&
 203             ev->ni_fail_type == PTL_NI_OK)
 204                 req->rq_reqlen = ev->mlength;
 205         do_gettimeofday(&req->rq_arrival_time);
 206         req->rq_peer.peer_nid = ev->initiator.nid;
 207         req->rq_peer.peer_ni = rqbd->rqbd_srv_ni->sni_ni;
 208         req->rq_rqbd = rqbd;
 209
 210         spin_lock_irqsave (&service->srv_lock, flags);
 211
 212         if (ev->unlinked) {
 213                 srv_ni->sni_nrqbd_receiving--;
 214                 if (ev->type != PTL_EVENT_UNLINK &&
 215                     srv_ni->sni_nrqbd_receiving == 0) {
 216                         /* This service is off-air on this interface because
 217                          * all its request buffers are busy.  Portals will
 218                          * start dropping incoming requests until more buffers
 219                          * get posted.  NB don't moan if it's because we're
 220                          * tearing down the service. */
 221                         CWARN("All %s %s request buffers busy\n",
 222                               service->srv_name, srv_ni->sni_ni->pni_name);
 223                 }
 224                 /* req takes over the network's ref on rqbd */
 225         } else {
 226                 /* req takes a ref on rqbd */
 227                 rqbd->rqbd_refcount++;
 228         }
 229
 230         list_add_tail(&req->rq_list, &service->srv_request_queue);
 231         service->srv_n_queued_reqs++;
 232
 233         /* NB everything can disappear under us once the request
 234          * has been queued and we unlock, so do the wake now... */
 235         wake_up(&service->srv_waitq);
 236
 237         spin_unlock_irqrestore(&service->srv_lock, flags);
 238         EXIT;
 239 }
 240
 241 /*
 242  *  Server's outgoing reply callback
 243  */
 244 void reply_out_callback(ptl_event_t *ev)
 245 {
 246         struct ptlrpc_cb_id       *cbid = ev->mem_desc.user_ptr;
 247         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 248         struct ptlrpc_srv_ni      *sni = rs->rs_srv_ni;
 249         struct ptlrpc_service     *svc = sni->sni_service;
 250         unsigned long              flags;
 251         ENTRY;
 252
 253         LASSERT (ev->type == PTL_EVENT_SEND_END ||
 254                  ev->type == PTL_EVENT_ACK ||
 255                  ev->type == PTL_EVENT_UNLINK);
 256
 257         if (!rs->rs_difficult) {
 258                 /* I'm totally responsible for freeing "easy" replies */
 259                 LASSERT (ev->unlinked);
 260                 lustre_free_reply_state (rs);
 261                 atomic_dec (&svc->srv_outstanding_replies);
 262                 EXIT;
 263                 return;
 264         }
 265
 266         LASSERT (rs->rs_on_net);
 267
 268         if (ev->unlinked) {
 269                 /* Last network callback */
 270                 spin_lock_irqsave (&svc->srv_lock, flags);
 271                 rs->rs_on_net = 0;
 272                 ptlrpc_schedule_difficult_reply (rs);
 273                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 274         }
 275
 276         EXIT;
 277 }
 278
 279 /*
 280  * Server's bulk completion callback
 281  */
 282 void server_bulk_callback (ptl_event_t *ev)
 283 {
 284         struct ptlrpc_cb_id     *cbid = ev->mem_desc.user_ptr;
 285         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 286         unsigned long            flags;
 287         ENTRY;
 288
 289         LASSERT (ev->type == PTL_EVENT_SEND_END ||
 290                  ev->type == PTL_EVENT_UNLINK ||
 291                  (desc->bd_type == BULK_PUT_SOURCE &&
 292                   ev->type == PTL_EVENT_ACK) ||
 293                  (desc->bd_type == BULK_GET_SINK &&
 294                   ev->type == PTL_EVENT_REPLY_END));
 295
 296         CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR,
 297                "event type %d, status %d, desc %p\n",
 298                ev->type, ev->ni_fail_type, desc);
 299
 300         spin_lock_irqsave (&desc->bd_lock, flags);
 301
 302         if ((ev->type == PTL_EVENT_ACK ||
 303              ev->type == PTL_EVENT_REPLY_END) &&
 304             ev->ni_fail_type == PTL_NI_OK) {
 305                 /* We heard back from the peer, so even if we get this
 306                  * before the SENT event (oh yes we can), we know we
 307                  * read/wrote the peer buffer and how much... */
 308                 desc->bd_success = 1;
 309                 desc->bd_nob_transferred = ev->mlength;
 310         }
 311
 312         if (ev->unlinked) {
 313                 /* This is the last callback no matter what... */
 314                 desc->bd_network_rw = 0;
 315                 wake_up(&desc->bd_waitq);
 316         }
 317
 318         spin_unlock_irqrestore (&desc->bd_lock, flags);
 319         EXIT;
 320 }
 321
 322 static int ptlrpc_master_callback(ptl_event_t *ev)
 323 {
 324         struct ptlrpc_cb_id *cbid = ev->mem_desc.user_ptr;
 325         void (*callback)(ptl_event_t *ev) = cbid->cbid_fn;
 326
 327         /* Honestly, it's best to find out early. */
 328         LASSERT (cbid->cbid_arg != (void *)0x5a5a5a5a5a5a5a5a);
 329         LASSERT (callback == request_out_callback ||
 330                  callback == reply_in_callback ||
 331                  callback == client_bulk_callback ||
 332                  callback == request_in_callback ||
 333                  callback == reply_out_callback ||
 334                  callback == server_bulk_callback);
 335
 336         callback (ev);
 337         return (0);
 338 }
 339
 340 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid, struct ptlrpc_peer *peer)
 341 {
 342         struct ptlrpc_ni   *pni;
 343         ptl_nid_t           peer_nid;
 344         ptl_handle_ni_t     peer_ni;
 345         int                 i;
 346         char                str[20];
 347         int                 rc = lustre_uuid_to_peer(uuid->uuid,
 348                                                      &peer_ni, &peer_nid);
 349         if (rc != 0)
 350                 RETURN (rc);
 351
 352         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 353                 pni = &ptlrpc_interfaces[i];
 354
 355                 if (!memcmp(&peer_ni, &pni->pni_ni_h,
 356                             sizeof (peer_ni))) {
 357                         peer->peer_nid = peer_nid;
 358                         peer->peer_ni = pni;
 359                         return (0);
 360                 }
 361         }
 362
 363         PtlSnprintHandle(str, sizeof(str), peer_ni);
 364         CERROR("Can't find ptlrpc interface for "LPX64" ni %s\n",
 365                peer_nid, str);
 366         return (-ENOENT);
 367 }
 368
 369 void ptlrpc_ni_fini(struct ptlrpc_ni *pni)
 370 {
 371         wait_queue_head_t   waitq;
 372         struct l_wait_info  lwi;
 373         int                 rc;
 374         int                 retries;
 375
 376         /* Wait for the event queue to become idle since there may still be
 377          * messages in flight with pending events (i.e. the fire-and-forget
 378          * messages == client requests and "non-difficult" server
 379          * replies */
 380
 381         for (retries = 0;; retries++) {
 382                 rc = PtlEQFree(pni->pni_eq_h);
 383                 switch (rc) {
 384                 default:
 385                         LBUG();
 386
 387                 case PTL_OK:
 388                         kportal_put_ni (pni->pni_number);
 389                         return;
 390
 391                 case PTL_EQ_IN_USE:
 392                         if (retries != 0)
 393                                 CWARN("Event queue for %s still busy\n",
 394                                       pni->pni_name);
 395
 396                         /* Wait for a bit */
 397                         init_waitqueue_head(&waitq);
 398                         lwi = LWI_TIMEOUT(2*HZ, NULL, NULL);
 399                         l_wait_event(waitq, 0, &lwi);
 400                         break;
 401                 }
 402         }
 403         /* notreached */
 404 }
 405
 406 int ptlrpc_ni_init(int number, char *name, struct ptlrpc_ni *pni)
 407 {
 408         int              rc;
 409         char             str[20];
 410         ptl_handle_ni_t *nip = kportal_get_ni (number);
 411
 412         if (nip == NULL) {
 413                 CDEBUG (D_NET, "Network interface %s not loaded\n", name);
 414                 return (-ENOENT);
 415         }
 416
 417         PtlSnprintHandle(str, sizeof(str), *nip);
 418         CDEBUG (D_NET, "init %d %s: %s\n", number, name, str);
 419
 420         pni->pni_name = name;
 421         pni->pni_number = number;
 422         pni->pni_ni_h = *nip;
 423
 424         pni->pni_eq_h = PTL_INVALID_HANDLE;
 425
 426         rc = PtlEQAlloc(pni->pni_ni_h, PTLRPC_NUM_EQ, PTLRPC_EQ_CALLBACK,
 427                         &pni->pni_eq_h);
 428
 429         if (rc != PTL_OK)
 430                 GOTO (fail, rc = -ENOMEM);
 431
 432         return (0);
 433  fail:
 434         CERROR ("Failed to initialise network interface %s: %d\n",
 435                 name, rc);
 436
 437         /* OK to do complete teardown since we invalidated the handles above */
 438         ptlrpc_ni_fini (pni);
 439         return (rc);
 440 }
 441
 442 #ifndef __KERNEL__
 443 LIST_HEAD(liblustre_wait_callbacks);
 444 void *liblustre_services_callback;
 445
 446 void *
 447 liblustre_register_wait_callback (int (*fn)(void *arg), void *arg)
 448 {
 449         struct liblustre_wait_callback *llwc;
 450
 451         OBD_ALLOC(llwc, sizeof(*llwc));
 452         LASSERT (llwc != NULL);
 453
 454         llwc->llwc_fn = fn;
 455         llwc->llwc_arg = arg;
 456         list_add_tail(&llwc->llwc_list, &liblustre_wait_callbacks);
 457
 458         return (llwc);
 459 }
 460
 461 void
 462 liblustre_deregister_wait_callback (void *opaque)
 463 {
 464         struct liblustre_wait_callback *llwc = opaque;
 465
 466         list_del(&llwc->llwc_list);
 467         OBD_FREE(llwc, sizeof(*llwc));
 468 }
 469
 470 int
 471 liblustre_check_events (int timeout)
 472 {
 473         ptl_event_t ev;
 474         int         rc;
 475         int         i;
 476         ENTRY;
 477
 478         rc = PtlEQPoll(&ptlrpc_interfaces[0].pni_eq_h, 1, timeout * 1000,
 479                        &ev, &i);
 480         if (rc == PTL_EQ_EMPTY)
 481                 RETURN(0);
 482
 483         LASSERT (rc == PTL_EQ_DROPPED || rc == PTL_OK);
 484
 485         /* liblustre: no asynch callback so we can't affort to miss any
 486          * events... */
 487         if (rc == PTL_EQ_DROPPED) {
 488                 CERROR ("Dropped an event!!!\n");
 489                 abort();
 490         }
 491
 492         ptlrpc_master_callback (&ev);
 493         RETURN(1);
 494 }
 495
 496 int liblustre_waiting = 0;
 497
 498 int
 499 liblustre_wait_event (int timeout)
 500 {
 501         struct list_head               *tmp;
 502         struct liblustre_wait_callback *llwc;
 503         int                             found_something = 0;
 504
 505         /* single threaded recursion check... */
 506         liblustre_waiting = 1;
 507
 508         for (;;) {
 509                 /* Deal with all pending events */
 510                 while (liblustre_check_events(0))
 511                         found_something = 1;
 512
 513                 /* Give all registered callbacks a bite at the cherry */
 514                 list_for_each(tmp, &liblustre_wait_callbacks) {
 515                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 516                                           llwc_list);
 517
 518                         if (llwc->llwc_fn(llwc->llwc_arg))
 519                                 found_something = 1;
 520                 }
 521
 522                 if (found_something || timeout == 0)
 523                         break;
 524
 525                 /* Nothing so far, but I'm allowed to block... */
 526                 found_something = liblustre_check_events(timeout);
 527                 if (!found_something)           /* still nothing */
 528                         break;                  /* I timed out */
 529         }
 530
 531         liblustre_waiting = 0;
 532
 533         return found_something;
 534 }
 535
 536 static int cray_portals_callback(ptl_event_t *ev)
 537 {
 538         /* We get a callback from the client Cray portals implementation
 539          * whenever anyone calls PtlEQPoll(), and an event queue with a
 540          * callback handler has outstanding events.
 541          *
 542          * If it's not liblustre calling PtlEQPoll(), this lets us know we
 543          * have outstanding events which we handle with
 544          * liblustre_wait_event().
 545          *
 546          * Otherwise, we're already eagerly consuming events and we'd
 547          * handle events out of order if we recursed. */
 548         if (liblustre_waiting)
 549                 return;
 550
 551         liblustre_wait_event(0);
 552 }
 553 #endif /* __KERNEL__ */
 554
 555 int ptlrpc_init_portals(void)
 556 {
 557         /* Add new portals network interfaces here.
 558          * Order is irrelevent! */
 559         static struct {
 560                 int   number;
 561                 char *name;
 562         } ptl_nis[] = {
 563                 {QSWNAL,  "qswnal"},
 564                 {SOCKNAL, "socknal"},
 565                 {GMNAL,   "gmnal"},
 566                 {IBNAL,   "ibnal"},
 567                 {TCPNAL,  "tcpnal"},
 568                 {SCIMACNAL, "scimacnal"}};
 569         int   rc;
 570         int   i;
 571
 572         LASSERT(ptlrpc_ninterfaces == 0);
 573
 574         for (i = 0; i < sizeof (ptl_nis) / sizeof (ptl_nis[0]); i++) {
 575                 LASSERT(ptlrpc_ninterfaces < (sizeof(ptlrpc_interfaces) /
 576                                               sizeof(ptlrpc_interfaces[0])));
 577
 578                 rc = ptlrpc_ni_init(ptl_nis[i].number, ptl_nis[i].name,
 579                                     &ptlrpc_interfaces[ptlrpc_ninterfaces]);
 580                 if (rc == 0)
 581                         ptlrpc_ninterfaces++;
 582         }
 583
 584         if (ptlrpc_ninterfaces == 0) {
 585                 CERROR("network initialisation failed: is a NAL module "
 586                        "loaded?\n");
 587                 return -EIO;
 588         }
 589 #ifndef __KERNEL__
 590         liblustre_services_callback =
 591                 liblustre_register_wait_callback(&liblustre_check_services, NULL);
 592 #endif
 593         return 0;
 594 }
 595
 596 void ptlrpc_exit_portals(void)
 597 {
 598 #ifndef __KERNEL__
 599         liblustre_deregister_wait_callback(liblustre_services_callback);
 600 #endif
 601         while (ptlrpc_ninterfaces > 0)
 602                 ptlrpc_ni_fini (&ptlrpc_interfaces[--ptlrpc_ninterfaces]);
 603 }