lnet/selftest/rpc.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   5  *   Author: Isaac Huang <isaac@clusterfs.com>
   6  *
   7  */
   8
   9 #define DEBUG_SUBSYSTEM S_LNET
  10
  11 #include "selftest.h"
  12
  13
  14 #define SRPC_PEER_HASH_SIZE       101  /* # peer lists */
  15
  16 typedef enum {
  17         SRPC_STATE_NONE,
  18         SRPC_STATE_NI_INIT,
  19         SRPC_STATE_EQ_INIT,
  20         SRPC_STATE_WI_INIT,
  21         SRPC_STATE_RUNNING,
  22         SRPC_STATE_STOPPING,
  23 } srpc_state_t;
  24
  25 struct smoketest_rpc {
  26         spinlock_t        rpc_glock;     /* global lock */
  27         srpc_service_t   *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  28         struct list_head *rpc_peers;     /* hash table of known peers */
  29         lnet_handle_eq_t  rpc_lnet_eq;   /* _the_ LNet event queue */
  30         srpc_state_t      rpc_state;
  31         srpc_counters_t   rpc_counters;
  32         __u64             rpc_matchbits; /* matchbits counter */
  33 } srpc_data;
  34
  35 static int srpc_peer_credits = 16;
  36 CFS_MODULE_PARM(srpc_peer_credits, "i", int, 0444,
  37                 "# in-flight RPCs per peer (16 by default)");
  38
  39 /* forward ref's */
  40 int srpc_handle_rpc (swi_workitem_t *wi);
  41
  42 void srpc_get_counters (srpc_counters_t *cnt)
  43 {
  44         spin_lock(&srpc_data.rpc_glock);
  45         *cnt = srpc_data.rpc_counters;
  46         spin_unlock(&srpc_data.rpc_glock);
  47 }
  48
  49 void srpc_set_counters (const srpc_counters_t *cnt)
  50 {
  51         spin_lock(&srpc_data.rpc_glock);
  52         srpc_data.rpc_counters = *cnt;
  53         spin_unlock(&srpc_data.rpc_glock);
  54 }
  55
  56 void
  57 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
  58 {
  59         LASSERT (i >= 0 && i < bk->bk_niov);
  60
  61 #ifdef __KERNEL__
  62         bk->bk_iovs[i].kiov_offset = 0;
  63         bk->bk_iovs[i].kiov_page   = pg;
  64         bk->bk_iovs[i].kiov_len    = CFS_PAGE_SIZE;
  65 #else
  66         LASSERT (bk->bk_pages != NULL);
  67
  68         bk->bk_pages[i] = pg;
  69         bk->bk_iovs[i].iov_len  = CFS_PAGE_SIZE;
  70         bk->bk_iovs[i].iov_base = cfs_page_address(pg);
  71 #endif
  72         return;
  73 }
  74
  75 void
  76 srpc_free_bulk (srpc_bulk_t *bk)
  77 {
  78         int         i;
  79         cfs_page_t *pg;
  80
  81         LASSERT (bk != NULL);
  82 #ifndef __KERNEL__
  83         LASSERT (bk->bk_pages != NULL);
  84 #endif
  85
  86         for (i = 0; i < bk->bk_niov; i++) {
  87 #ifdef __KERNEL__
  88                 pg = bk->bk_iovs[i].kiov_page;
  89 #else
  90                 pg = bk->bk_pages[i];
  91 #endif
  92                 if (pg == NULL) break;
  93
  94                 cfs_free_page(pg);
  95         }
  96
  97 #ifndef __KERNEL__
  98         LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
  99 #endif
 100         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 101         return;
 102 }
 103
 104 srpc_bulk_t *
 105 srpc_alloc_bulk (int npages, int sink)
 106 {
 107         srpc_bulk_t  *bk;
 108         cfs_page_t  **pages;
 109         int           i;
 110
 111         LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
 112
 113         LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 114         if (bk == NULL) {
 115                 CERROR ("Can't allocate descriptor for %d pages\n", npages);
 116                 return NULL;
 117         }
 118
 119         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
 120         bk->bk_sink = sink;
 121         bk->bk_niov = npages;
 122         bk->bk_len  = npages * CFS_PAGE_SIZE;
 123 #ifndef __KERNEL__
 124         LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
 125         if (pages == NULL) {
 126                 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 127                 CERROR ("Can't allocate page array for %d pages\n", npages);
 128                 return NULL;
 129         }
 130
 131         memset(pages, 0, sizeof(cfs_page_t *) * npages);
 132         bk->bk_pages = pages;
 133 #else
 134         UNUSED (pages);
 135 #endif
 136
 137         for (i = 0; i < npages; i++) {
 138                 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
 139
 140                 if (pg == NULL) {
 141                         CERROR ("Can't allocate page %d of %d\n", i, npages);
 142                         srpc_free_bulk(bk);
 143                         return NULL;
 144                 }
 145
 146                 srpc_add_bulk_page(bk, pg, i);
 147         }
 148
 149         return bk;
 150 }
 151
 152
 153 static inline struct list_head *
 154 srpc_nid2peerlist (lnet_nid_t nid)
 155 {
 156         unsigned int hash = ((unsigned int)nid) % SRPC_PEER_HASH_SIZE;
 157
 158         return &srpc_data.rpc_peers[hash];
 159 }
 160
 161 static inline srpc_peer_t *
 162 srpc_create_peer (lnet_nid_t nid)
 163 {
 164         srpc_peer_t *peer;
 165
 166         LASSERT (nid != LNET_NID_ANY);
 167
 168         LIBCFS_ALLOC(peer, sizeof(srpc_peer_t));
 169         if (peer == NULL) {
 170                 CERROR ("Failed to allocate peer structure for %s\n",
 171                         libcfs_nid2str(nid));
 172                 return NULL;
 173         }
 174
 175         memset(peer, 0, sizeof(srpc_peer_t));
 176         peer->stp_nid     = nid;
 177         peer->stp_credits = srpc_peer_credits;
 178
 179         spin_lock_init(&peer->stp_lock);
 180         CFS_INIT_LIST_HEAD(&peer->stp_rpcq);
 181         CFS_INIT_LIST_HEAD(&peer->stp_ctl_rpcq);
 182         return peer;
 183 }
 184
 185 srpc_peer_t *
 186 srpc_find_peer_locked (lnet_nid_t nid)
 187 {
 188         struct list_head *peer_list = srpc_nid2peerlist(nid);
 189         srpc_peer_t      *peer;
 190
 191         LASSERT (nid != LNET_NID_ANY);
 192
 193         list_for_each_entry (peer, peer_list, stp_list) {
 194                 if (peer->stp_nid == nid)
 195                         return peer;
 196         }
 197
 198         return NULL;
 199 }
 200
 201 static srpc_peer_t *
 202 srpc_nid2peer (lnet_nid_t nid)
 203 {
 204         srpc_peer_t *peer;
 205         srpc_peer_t *new_peer;
 206
 207         spin_lock(&srpc_data.rpc_glock);
 208         peer = srpc_find_peer_locked(nid);
 209         spin_unlock(&srpc_data.rpc_glock);
 210
 211         if (peer != NULL)
 212                 return peer;
 213
 214         new_peer = srpc_create_peer(nid);
 215
 216         spin_lock(&srpc_data.rpc_glock);
 217
 218         peer = srpc_find_peer_locked(nid);
 219         if (peer != NULL) {
 220                 spin_unlock(&srpc_data.rpc_glock);
 221                 if (new_peer != NULL)
 222                         LIBCFS_FREE(new_peer, sizeof(srpc_peer_t));
 223
 224                 return peer;
 225         }
 226
 227         if (new_peer == NULL) {
 228                 spin_unlock(&srpc_data.rpc_glock);
 229                 return NULL;
 230         }
 231
 232         list_add_tail(&new_peer->stp_list, srpc_nid2peerlist(nid));
 233         spin_unlock(&srpc_data.rpc_glock);
 234         return new_peer;
 235 }
 236
 237 static inline __u64
 238 srpc_next_id (void)
 239 {
 240         __u64 id;
 241
 242         spin_lock(&srpc_data.rpc_glock);
 243         id = srpc_data.rpc_matchbits++;
 244         spin_unlock(&srpc_data.rpc_glock);
 245         return id;
 246 }
 247
 248 void
 249 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
 250                       srpc_service_t *sv, srpc_buffer_t *buffer)
 251 {
 252         memset(rpc, 0, sizeof(*rpc));
 253         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc);
 254
 255         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 256
 257         rpc->srpc_service  = sv;
 258         rpc->srpc_reqstbuf = buffer;
 259         rpc->srpc_peer     = buffer->buf_peer;
 260         rpc->srpc_self     = buffer->buf_self;
 261         rpc->srpc_replymdh = LNET_INVALID_HANDLE;
 262 }
 263
 264 int
 265 srpc_add_service (srpc_service_t *sv)
 266 {
 267         int                id = sv->sv_id;
 268         int                i;
 269         srpc_server_rpc_t *rpc;
 270
 271         LASSERT (sv->sv_concur > 0);
 272         LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
 273
 274         spin_lock(&srpc_data.rpc_glock);
 275
 276         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
 277
 278         if (srpc_data.rpc_services[id] != NULL) {
 279                 spin_unlock(&srpc_data.rpc_glock);
 280                 return -EBUSY;
 281         }
 282
 283         srpc_data.rpc_services[id] = sv;
 284         spin_unlock(&srpc_data.rpc_glock);
 285
 286         sv->sv_nprune       = 0;
 287         sv->sv_nposted_msg  = 0;
 288         sv->sv_shuttingdown = 0;
 289         spin_lock_init(&sv->sv_lock);
 290         CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
 291         CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
 292         CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
 293         CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
 294
 295         sv->sv_ev.ev_data = sv;
 296         sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
 297
 298         for (i = 0; i < sv->sv_concur; i++) {
 299                 LIBCFS_ALLOC(rpc, sizeof(*rpc));
 300                 if (rpc == NULL) goto enomem;
 301
 302                 list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 303         }
 304
 305         CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
 306                 id, sv->sv_name, sv->sv_concur);
 307         return 0;
 308
 309 enomem:
 310         while (!list_empty(&sv->sv_free_rpcq)) {
 311                 rpc = list_entry(sv->sv_free_rpcq.next,
 312                                  srpc_server_rpc_t, srpc_list);
 313                 list_del(&rpc->srpc_list);
 314                 LIBCFS_FREE(rpc, sizeof(*rpc));
 315         }
 316
 317         spin_lock(&srpc_data.rpc_glock);
 318         srpc_data.rpc_services[id] = NULL;
 319         spin_unlock(&srpc_data.rpc_glock);
 320         return -ENOMEM;
 321 }
 322
 323 int
 324 srpc_remove_service (srpc_service_t *sv)
 325 {
 326         int id = sv->sv_id;
 327
 328         spin_lock(&srpc_data.rpc_glock);
 329
 330         if (srpc_data.rpc_services[id] != sv) {
 331                 spin_unlock(&srpc_data.rpc_glock);
 332                 return -ENOENT;
 333         }
 334
 335         srpc_data.rpc_services[id] = NULL;
 336         spin_unlock(&srpc_data.rpc_glock);
 337         return 0;
 338 }
 339
 340 int
 341 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
 342                        int len, int options, lnet_process_id_t peer,
 343                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 344 {
 345         int              rc;
 346         lnet_md_t        md;
 347         lnet_handle_me_t meh;
 348
 349         rc = LNetMEAttach(portal, peer, matchbits, 0,
 350                           LNET_UNLINK, LNET_INS_AFTER, &meh);
 351         if (rc != 0) {
 352                 CERROR ("LNetMEAttach failed: %d\n", rc);
 353                 LASSERT (rc == -ENOMEM);
 354                 return -ENOMEM;
 355         }
 356
 357         md.threshold = 1;
 358         md.user_ptr  = ev;
 359         md.start     = buf;
 360         md.length    = len;
 361         md.options   = options;
 362         md.eq_handle = srpc_data.rpc_lnet_eq;
 363
 364         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 365         if (rc != 0) {
 366                 CERROR ("LNetMDAttach failed: %d\n", rc);
 367                 LASSERT (rc == -ENOMEM);
 368
 369                 rc = LNetMEUnlink(meh);
 370                 LASSERT (rc == 0);
 371                 return -ENOMEM;
 372         }
 373
 374         CDEBUG (D_NET,
 375                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 376                 libcfs_id2str(peer), portal, matchbits);
 377         return 0;
 378 }
 379
 380 int
 381 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 382                       int options, lnet_process_id_t peer, lnet_nid_t self,
 383                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 384 {
 385         int       rc;
 386         lnet_md_t md;
 387
 388         md.user_ptr  = ev;
 389         md.start     = buf;
 390         md.length    = len;
 391         md.eq_handle = srpc_data.rpc_lnet_eq;
 392         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 393         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 394
 395         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 396         if (rc != 0) {
 397                 CERROR ("LNetMDBind failed: %d\n", rc);
 398                 LASSERT (rc == -ENOMEM);
 399                 return -ENOMEM;
 400         }
 401
 402         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 403          * they're only meaningful for MDs attached to an ME (i.e. passive
 404          * buffers... */
 405         if ((options & LNET_MD_OP_PUT) != 0) {
 406                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 407                              portal, matchbits, 0, 0);
 408         } else {
 409                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 410
 411                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 412         }
 413
 414         if (rc != 0) {
 415                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 416                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 417                         libcfs_id2str(peer), portal, matchbits, rc);
 418
 419                 /* The forthcoming unlink event will complete this operation
 420                  * with failure, so fall through and return success here.
 421                  */
 422                 rc = LNetMDUnlink(*mdh);
 423                 LASSERT (rc == 0);
 424         } else {
 425                 CDEBUG (D_NET,
 426                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 427                         libcfs_id2str(peer), portal, matchbits);
 428         }
 429         return 0;
 430 }
 431
 432 int
 433 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 434                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 435 {
 436         int rc;
 437         int portal;
 438
 439         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 440                 portal = SRPC_REQUEST_PORTAL;
 441         else
 442                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 443
 444         rc = srpc_post_active_rdma(portal, service, buf, len,
 445                                    LNET_MD_OP_PUT, peer,
 446                                    LNET_NID_ANY, mdh, ev);
 447         return rc;
 448 }
 449
 450 int
 451 srpc_post_passive_rqtbuf(int service, void *buf, int len,
 452                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 453 {
 454         int               rc;
 455         int               portal;
 456         lnet_process_id_t any = {.nid = LNET_NID_ANY,
 457                                  .pid = LNET_PID_ANY};
 458
 459         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 460                 portal = SRPC_REQUEST_PORTAL;
 461         else
 462                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 463
 464         rc = srpc_post_passive_rdma(portal, service, buf, len,
 465                                     LNET_MD_OP_PUT, any, mdh, ev);
 466         return rc;
 467 }
 468
 469 int
 470 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 471 {
 472         srpc_msg_t *msg = &buf->buf_msg;
 473         int         rc;
 474
 475         LASSERT (!sv->sv_shuttingdown);
 476
 477         buf->buf_mdh = LNET_INVALID_HANDLE;
 478         list_add(&buf->buf_list, &sv->sv_posted_msgq);
 479         sv->sv_nposted_msg++;
 480         spin_unlock(&sv->sv_lock);
 481
 482         rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
 483                                       &buf->buf_mdh, &sv->sv_ev);
 484
 485         /* At this point, a RPC (new or delayed) may have arrived in
 486          * msg and its event handler has been called. So we must add
 487          * buf to sv_posted_msgq _before_ dropping sv_lock */
 488
 489         spin_lock(&sv->sv_lock);
 490
 491         if (rc == 0) {
 492                 if (sv->sv_shuttingdown) {
 493                         spin_unlock(&sv->sv_lock);
 494
 495                         /* srpc_shutdown_service might have tried to unlink me
 496                          * when my buf_mdh was still invalid */
 497                         LNetMDUnlink(buf->buf_mdh);
 498
 499                         spin_lock(&sv->sv_lock);
 500                 }
 501                 return 0;
 502         }
 503
 504         sv->sv_nposted_msg--;
 505         if (sv->sv_shuttingdown) return rc;
 506
 507         list_del(&buf->buf_list);
 508
 509         spin_unlock(&sv->sv_lock);
 510         LIBCFS_FREE(buf, sizeof(*buf));
 511         spin_lock(&sv->sv_lock);
 512         return rc;
 513 }
 514
 515 int
 516 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
 517 {
 518         int                rc;
 519         int                posted;
 520         srpc_buffer_t     *buf;
 521
 522         LASSERTF (nbuffer > 0,
 523                   "nbuffer must be positive: %d\n", nbuffer);
 524
 525         for (posted = 0; posted < nbuffer; posted++) {
 526                 LIBCFS_ALLOC(buf, sizeof(*buf));
 527                 if (buf == NULL) break;
 528
 529                 spin_lock(&sv->sv_lock);
 530                 rc = srpc_service_post_buffer(sv, buf);
 531                 spin_unlock(&sv->sv_lock);
 532
 533                 if (rc != 0) break;
 534         }
 535
 536         return posted;
 537 }
 538
 539 void
 540 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
 541 {
 542         LASSERTF (nbuffer > 0,
 543                   "nbuffer must be positive: %d\n", nbuffer);
 544
 545         spin_lock(&sv->sv_lock);
 546
 547         LASSERT (sv->sv_nprune >= 0);
 548         LASSERT (!sv->sv_shuttingdown);
 549
 550         sv->sv_nprune += nbuffer;
 551
 552         spin_unlock(&sv->sv_lock);
 553         return;
 554 }
 555
 556 /* returns 1 if sv has finished, otherwise 0 */
 557 int
 558 srpc_finish_service (srpc_service_t *sv)
 559 {
 560         srpc_server_rpc_t *rpc;
 561         srpc_buffer_t     *buf;
 562
 563         spin_lock(&sv->sv_lock);
 564
 565         LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
 566
 567         if (sv->sv_nposted_msg != 0 || !list_empty(&sv->sv_active_rpcq)) {
 568                 CDEBUG (D_NET,
 569                         "waiting for %d posted buffers to unlink and "
 570                         "in-flight RPCs to die.\n",
 571                         sv->sv_nposted_msg);
 572
 573                 if (!list_empty(&sv->sv_active_rpcq)) {
 574                         rpc = list_entry(sv->sv_active_rpcq.next,
 575                                          srpc_server_rpc_t, srpc_list);
 576                         CDEBUG (D_NETERROR,
 577                                 "Active RPC on shutdown: sv %s, peer %s, "
 578                                 "wi %s scheduled %d running %d, "
 579                                 "ev fired %d type %d status %d lnet %d\n",
 580                                 sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 581                                 swi_state2str(rpc->srpc_wi.wi_state),
 582                                 rpc->srpc_wi.wi_scheduled,
 583                                 rpc->srpc_wi.wi_running,
 584                                 rpc->srpc_ev.ev_fired,
 585                                 rpc->srpc_ev.ev_type,
 586                                 rpc->srpc_ev.ev_status,
 587                                 rpc->srpc_ev.ev_lnet);
 588                 }
 589
 590                 spin_unlock(&sv->sv_lock);
 591                 return 0;
 592         }
 593
 594         spin_unlock(&sv->sv_lock); /* no lock needed from now on */
 595
 596         for (;;) {
 597                 struct list_head *q;
 598
 599                 if (!list_empty(&sv->sv_posted_msgq))
 600                         q = &sv->sv_posted_msgq;
 601                 else if (!list_empty(&sv->sv_blocked_msgq))
 602                         q = &sv->sv_blocked_msgq;
 603                 else
 604                         break;
 605
 606                 buf = list_entry(q->next, srpc_buffer_t, buf_list);
 607                 list_del(&buf->buf_list);
 608
 609                 LIBCFS_FREE(buf, sizeof(*buf));
 610         }
 611
 612         while (!list_empty(&sv->sv_free_rpcq)) {
 613                 rpc = list_entry(sv->sv_free_rpcq.next,
 614                                  srpc_server_rpc_t, srpc_list);
 615                 list_del(&rpc->srpc_list);
 616                 LIBCFS_FREE(rpc, sizeof(*rpc));
 617         }
 618
 619         return 1;
 620 }
 621
 622 /* called with sv->sv_lock held */
 623 void
 624 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 625 {
 626         if (sv->sv_shuttingdown) goto free;
 627
 628         if (sv->sv_nprune == 0) {
 629                 if (srpc_service_post_buffer(sv, buf) != 0)
 630                         CWARN ("Failed to post %s buffer\n", sv->sv_name);
 631                 return;
 632         }
 633
 634         sv->sv_nprune--;
 635 free:
 636         spin_unlock(&sv->sv_lock);
 637         LIBCFS_FREE(buf, sizeof(*buf));
 638         spin_lock(&sv->sv_lock);
 639 }
 640
 641 void
 642 srpc_shutdown_service (srpc_service_t *sv)
 643 {
 644         srpc_server_rpc_t *rpc;
 645         srpc_buffer_t     *buf;
 646
 647         spin_lock(&sv->sv_lock);
 648
 649         CDEBUG (D_NET, "Shutting down service: id %d, name %s\n",
 650                 sv->sv_id, sv->sv_name);
 651
 652         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 653
 654         /* schedule in-flight RPCs to notice the shutdown */
 655         list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
 656                 swi_schedule_workitem(&rpc->srpc_wi);
 657         }
 658
 659         spin_unlock(&sv->sv_lock);
 660
 661         /* OK to traverse sv_posted_msgq without lock, since no one
 662          * touches sv_posted_msgq now */
 663         list_for_each_entry (buf, &sv->sv_posted_msgq, buf_list)
 664                 LNetMDUnlink(buf->buf_mdh);
 665
 666         return;
 667 }
 668
 669 int
 670 srpc_send_request (srpc_client_rpc_t *rpc)
 671 {
 672         srpc_event_t *ev = &rpc->crpc_reqstev;
 673         int           rc;
 674
 675         ev->ev_fired = 0;
 676         ev->ev_data  = rpc;
 677         ev->ev_type  = SRPC_REQUEST_SENT;
 678
 679         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 680                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 681                                      &rpc->crpc_reqstmdh, ev);
 682         if (rc != 0) {
 683                 LASSERT (rc == -ENOMEM);
 684                 ev->ev_fired = 1;  /* no more event expected */
 685         }
 686         return rc;
 687 }
 688
 689 int
 690 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 691 {
 692         srpc_event_t *ev = &rpc->crpc_replyev;
 693         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 694         int           rc;
 695
 696         ev->ev_fired = 0;
 697         ev->ev_data  = rpc;
 698         ev->ev_type  = SRPC_REPLY_RCVD;
 699
 700         *id = srpc_next_id();
 701
 702         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 703                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 704                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 705                                     &rpc->crpc_replymdh, ev);
 706         if (rc != 0) {
 707                 LASSERT (rc == -ENOMEM);
 708                 ev->ev_fired = 1;  /* no more event expected */
 709         }
 710         return rc;
 711 }
 712
 713 int
 714 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 715 {
 716         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 717         srpc_event_t *ev = &rpc->crpc_bulkev;
 718         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 719         int           rc;
 720         int           opt;
 721
 722         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 723
 724         if (bk->bk_niov == 0) return 0; /* nothing to do */
 725
 726         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 727 #ifdef __KERNEL__
 728         opt |= LNET_MD_KIOV;
 729 #else
 730         opt |= LNET_MD_IOVEC;
 731 #endif
 732
 733         ev->ev_fired = 0;
 734         ev->ev_data  = rpc;
 735         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 736
 737         *id = srpc_next_id();
 738
 739         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 740                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 741                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 742         if (rc != 0) {
 743                 LASSERT (rc == -ENOMEM);
 744                 ev->ev_fired = 1;  /* no more event expected */
 745         }
 746         return rc;
 747 }
 748
 749 int
 750 srpc_do_bulk (srpc_server_rpc_t *rpc)
 751 {
 752         srpc_event_t  *ev = &rpc->srpc_ev;
 753         srpc_bulk_t   *bk = rpc->srpc_bulk;
 754         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 755         int            rc;
 756         int            opt;
 757
 758         LASSERT (bk != NULL);
 759
 760         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 761 #ifdef __KERNEL__
 762         opt |= LNET_MD_KIOV;
 763 #else
 764         opt |= LNET_MD_IOVEC;
 765 #endif
 766
 767         ev->ev_fired = 0;
 768         ev->ev_data  = rpc;
 769         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 770
 771         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 772                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 773                                    rpc->srpc_peer, rpc->srpc_self,
 774                                    &bk->bk_mdh, ev);
 775         if (rc != 0)
 776                 ev->ev_fired = 1;  /* no more event expected */
 777         return rc;
 778 }
 779
 780 /* called with srpc_service_t::sv_lock held */
 781 inline void
 782 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
 783 {
 784         srpc_service_t *sv = rpc->srpc_service;
 785
 786         if (sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 787                 swi_schedule_workitem(&rpc->srpc_wi);
 788         else    /* framework RPCs are handled one by one */
 789                 swi_schedule_serial_workitem(&rpc->srpc_wi);
 790
 791         return;
 792 }
 793
 794 /* only called from srpc_handle_rpc */
 795 void
 796 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
 797 {
 798         srpc_service_t *sv = rpc->srpc_service;
 799         srpc_buffer_t  *buffer;
 800
 801         LASSERT (status != 0 || rpc->srpc_wi.wi_state == SWI_STATE_DONE);
 802
 803         rpc->srpc_status = status;
 804
 805         CDEBUG (status == 0 ? D_NET : D_NETERROR,
 806                 "Server RPC done: service %s, peer %s, status %s:%d\n",
 807                 sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 808                 swi_state2str(rpc->srpc_wi.wi_state), status);
 809
 810         if (status != 0) {
 811                 spin_lock(&srpc_data.rpc_glock);
 812                 srpc_data.rpc_counters.rpcs_dropped++;
 813                 spin_unlock(&srpc_data.rpc_glock);
 814         }
 815
 816         if (rpc->srpc_done != NULL)
 817                 (*rpc->srpc_done) (rpc);
 818         LASSERT (rpc->srpc_bulk == NULL);
 819
 820         spin_lock(&sv->sv_lock);
 821
 822         if (rpc->srpc_reqstbuf != NULL) {
 823                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 824                  * sv won't go away for sv_active_rpcq must not be empty */
 825                 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
 826                 rpc->srpc_reqstbuf = NULL;
 827         }
 828
 829         list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
 830
 831         /*
 832          * No one can schedule me now since:
 833          * - I'm not on sv_active_rpcq.
 834          * - all LNet events have been fired.
 835          * Cancel pending schedules and prevent future schedule attempts:
 836          */
 837         LASSERT (rpc->srpc_ev.ev_fired);
 838         swi_kill_workitem(&rpc->srpc_wi);
 839
 840         if (!sv->sv_shuttingdown && !list_empty(&sv->sv_blocked_msgq)) {
 841                 buffer = list_entry(sv->sv_blocked_msgq.next,
 842                                     srpc_buffer_t, buf_list);
 843                 list_del(&buffer->buf_list);
 844
 845                 srpc_init_server_rpc(rpc, sv, buffer);
 846                 list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
 847                 srpc_schedule_server_rpc(rpc);
 848         } else {
 849                 list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 850         }
 851
 852         spin_unlock(&sv->sv_lock);
 853         return;
 854 }
 855
 856 /* handles an incoming RPC */
 857 int
 858 srpc_handle_rpc (swi_workitem_t *wi)
 859 {
 860         srpc_server_rpc_t *rpc = wi->wi_data;
 861         srpc_service_t    *sv = rpc->srpc_service;
 862         srpc_event_t      *ev = &rpc->srpc_ev;
 863         int                rc = 0;
 864
 865         LASSERT (wi == &rpc->srpc_wi);
 866
 867         spin_lock(&sv->sv_lock);
 868
 869         if (sv->sv_shuttingdown) {
 870                 spin_unlock(&sv->sv_lock);
 871
 872                 if (rpc->srpc_bulk != NULL)
 873                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 874                 LNetMDUnlink(rpc->srpc_replymdh);
 875
 876                 if (ev->ev_fired) { /* no more event, OK to finish */
 877                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 878                         return 1;
 879                 }
 880                 return 0;
 881         }
 882
 883         spin_unlock(&sv->sv_lock);
 884
 885         switch (wi->wi_state) {
 886         default:
 887                 LBUG ();
 888         case SWI_STATE_NEWBORN: {
 889                 srpc_msg_t           *msg;
 890                 srpc_generic_reply_t *reply;
 891
 892                 msg = &rpc->srpc_reqstbuf->buf_msg;
 893                 reply = &rpc->srpc_replymsg.msg_body.reply;
 894
 895                 if (msg->msg_version != SRPC_MSG_VERSION &&
 896                     msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
 897                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
 898                                msg->msg_version, SRPC_MSG_VERSION,
 899                                libcfs_id2str(rpc->srpc_peer));
 900                         reply->status = EPROTO;
 901                 } else {
 902                         reply->status = 0;
 903                         rc = (*sv->sv_handler) (rpc);
 904                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
 905                 }
 906
 907                 if (rc != 0) {
 908                         srpc_server_rpc_done(rpc, rc);
 909                         return 1;
 910                 }
 911
 912                 wi->wi_state = SWI_STATE_BULK_STARTED;
 913
 914                 if (rpc->srpc_bulk != NULL) {
 915                         rc = srpc_do_bulk(rpc);
 916                         if (rc == 0)
 917                                 return 0; /* wait for bulk */
 918
 919                         LASSERT (ev->ev_fired);
 920                         ev->ev_status = rc;
 921                 }
 922         }
 923         case SWI_STATE_BULK_STARTED:
 924                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
 925
 926                 if (rpc->srpc_bulk != NULL) {
 927                         rc = ev->ev_status;
 928
 929                         if (sv->sv_bulk_ready != NULL)
 930                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
 931
 932                         if (rc != 0) {
 933                                 srpc_server_rpc_done(rpc, rc);
 934                                 return 1;
 935                         }
 936                 }
 937
 938                 wi->wi_state = SWI_STATE_REPLY_SUBMITTED;
 939                 rc = srpc_send_reply(rpc);
 940                 if (rc == 0)
 941                         return 0; /* wait for reply */
 942                 srpc_server_rpc_done(rpc, rc);
 943                 return 1;
 944
 945         case SWI_STATE_REPLY_SUBMITTED:
 946                 LASSERT (ev->ev_fired);
 947
 948                 wi->wi_state = SWI_STATE_DONE;
 949                 srpc_server_rpc_done(rpc, ev->ev_status);
 950                 return 1;
 951         }
 952
 953         return 0;
 954 }
 955
 956 void
 957 srpc_client_rpc_expired (void *data)
 958 {
 959         srpc_client_rpc_t *rpc = data;
 960
 961         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
 962                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
 963                rpc->crpc_timeout);
 964
 965         spin_lock(&rpc->crpc_lock);
 966
 967         rpc->crpc_timeout = 0;
 968         srpc_abort_rpc(rpc, -ETIMEDOUT);
 969
 970         spin_unlock(&rpc->crpc_lock);
 971
 972         spin_lock(&srpc_data.rpc_glock);
 973         srpc_data.rpc_counters.rpcs_expired++;
 974         spin_unlock(&srpc_data.rpc_glock);
 975         return;
 976 }
 977
 978 inline void
 979 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
 980 {
 981         stt_timer_t *timer = &rpc->crpc_timer;
 982
 983         if (rpc->crpc_timeout == 0) return;
 984
 985         CFS_INIT_LIST_HEAD(&timer->stt_list);
 986         timer->stt_data    = rpc;
 987         timer->stt_func    = srpc_client_rpc_expired;
 988         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
 989                                           cfs_time_current_sec());
 990         stt_add_timer(timer);
 991         return;
 992 }
 993
 994 /*
 995  * Called with rpc->crpc_lock held.
 996  *
 997  * Upon exit the RPC expiry timer is not queued and the handler is not
 998  * running on any CPU. */
 999 void
1000 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1001 {
1002         /* timer not planted or already exploded */
1003         if (rpc->crpc_timeout == 0) return;
1004
1005         /* timer sucessfully defused */
1006         if (stt_del_timer(&rpc->crpc_timer)) return;
1007
1008 #ifdef __KERNEL__
1009         /* timer detonated, wait for it to explode */
1010         while (rpc->crpc_timeout != 0) {
1011                 spin_unlock(&rpc->crpc_lock);
1012
1013                 cfs_schedule();
1014
1015                 spin_lock(&rpc->crpc_lock);
1016         }
1017 #else
1018         LBUG(); /* impossible in single-threaded runtime */
1019 #endif
1020         return;
1021 }
1022
1023 void
1024 srpc_check_sends (srpc_peer_t *peer, int credits)
1025 {
1026         struct list_head  *q;
1027         srpc_client_rpc_t *rpc;
1028
1029         LASSERT (credits >= 0);
1030         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1031
1032         spin_lock(&peer->stp_lock);
1033         peer->stp_credits += credits;
1034
1035         while (peer->stp_credits) {
1036                 if (!list_empty(&peer->stp_ctl_rpcq))
1037                         q = &peer->stp_ctl_rpcq;
1038                 else if (!list_empty(&peer->stp_rpcq))
1039                         q = &peer->stp_rpcq;
1040                 else
1041                         break;
1042
1043                 peer->stp_credits--;
1044
1045                 rpc = list_entry(q->next, srpc_client_rpc_t, crpc_privl);
1046                 list_del_init(&rpc->crpc_privl);
1047                 srpc_client_rpc_decref(rpc);  /* --ref for peer->*rpcq */
1048
1049                 swi_schedule_workitem(&rpc->crpc_wi);
1050         }
1051
1052         spin_unlock(&peer->stp_lock);
1053         return;
1054 }
1055
1056 void
1057 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1058 {
1059         swi_workitem_t *wi = &rpc->crpc_wi;
1060         srpc_peer_t    *peer = rpc->crpc_peer;
1061
1062         LASSERT (status != 0 || wi->wi_state == SWI_STATE_DONE);
1063
1064         spin_lock(&rpc->crpc_lock);
1065
1066         rpc->crpc_closed = 1;
1067         if (rpc->crpc_status == 0)
1068                 rpc->crpc_status = status;
1069
1070         srpc_del_client_rpc_timer(rpc);
1071
1072         CDEBUG ((status == 0) ? D_NET : D_NETERROR,
1073                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1074                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1075                 swi_state2str(wi->wi_state), rpc->crpc_aborted, status);
1076
1077         /*
1078          * No one can schedule me now since:
1079          * - RPC timer has been defused.
1080          * - all LNet events have been fired.
1081          * - crpc_closed has been set, preventing srpc_abort_rpc from
1082          *   scheduling me.
1083          * Cancel pending schedules and prevent future schedule attempts:
1084          */
1085         LASSERT (!srpc_event_pending(rpc));
1086         swi_kill_workitem(wi);
1087
1088         spin_unlock(&rpc->crpc_lock);
1089
1090         (*rpc->crpc_done) (rpc);
1091
1092         if (peer != NULL)
1093                 srpc_check_sends(peer, 1);
1094         return;
1095 }
1096
1097 /* sends an outgoing RPC */
1098 int
1099 srpc_send_rpc (swi_workitem_t *wi)
1100 {
1101         int                rc = 0;
1102         srpc_client_rpc_t *rpc = wi->wi_data;
1103         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1104         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1105
1106         LASSERT (rpc != NULL);
1107         LASSERT (wi == &rpc->crpc_wi);
1108
1109         spin_lock(&rpc->crpc_lock);
1110
1111         if (rpc->crpc_aborted) {
1112                 spin_unlock(&rpc->crpc_lock);
1113                 goto abort;
1114         }
1115
1116         spin_unlock(&rpc->crpc_lock);
1117
1118         switch (wi->wi_state) {
1119         default:
1120                 LBUG ();
1121         case SWI_STATE_NEWBORN:
1122                 LASSERT (!srpc_event_pending(rpc));
1123
1124                 rc = srpc_prepare_reply(rpc);
1125                 if (rc != 0) {
1126                         srpc_client_rpc_done(rpc, rc);
1127                         return 1;
1128                 }
1129
1130                 rc = srpc_prepare_bulk(rpc);
1131                 if (rc != 0) break;
1132
1133                 wi->wi_state = SWI_STATE_REQUEST_SUBMITTED;
1134                 rc = srpc_send_request(rpc);
1135                 break;
1136
1137         case SWI_STATE_REQUEST_SUBMITTED:
1138                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1139                  * order; however, they're processed in a strict order:
1140                  * rqt, rpy, and bulk. */
1141                 if (!rpc->crpc_reqstev.ev_fired) break;
1142
1143                 rc = rpc->crpc_reqstev.ev_status;
1144                 if (rc != 0) break;
1145
1146                 wi->wi_state = SWI_STATE_REQUEST_SENT;
1147                 /* perhaps more events, fall thru */
1148         case SWI_STATE_REQUEST_SENT: {
1149                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1150
1151                 if (!rpc->crpc_replyev.ev_fired) break;
1152
1153                 rc = rpc->crpc_replyev.ev_status;
1154                 if (rc != 0) break;
1155
1156                 if ((reply->msg_type != type &&
1157                      reply->msg_type != __swab32(type)) ||
1158                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1159                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1160                         CWARN ("Bad message from %s: type %u (%d expected),"
1161                                " magic %u (%d expected).\n",
1162                                libcfs_id2str(rpc->crpc_dest),
1163                                reply->msg_type, type,
1164                                reply->msg_magic, SRPC_MSG_MAGIC);
1165                         rc = -EBADMSG;
1166                         break;
1167                 }
1168
1169                 if (do_bulk && reply->msg_body.reply.status != 0) {
1170                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1171                                "case peer didn't initiate bulk transfer\n",
1172                                reply->msg_body.reply.status,
1173                                libcfs_id2str(rpc->crpc_dest));
1174                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1175                 }
1176
1177                 wi->wi_state = SWI_STATE_REPLY_RECEIVED;
1178         }
1179         case SWI_STATE_REPLY_RECEIVED:
1180                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1181
1182                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1183
1184                 /* Bulk buffer was unlinked due to remote error. Clear error
1185                  * since reply buffer still contains valid data.
1186                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1187                  * remote error. */
1188                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1189                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1190                         rc = 0;
1191
1192                 wi->wi_state = SWI_STATE_DONE;
1193                 srpc_client_rpc_done(rpc, rc);
1194                 return 1;
1195         }
1196
1197         if (rc != 0) {
1198                 spin_lock(&rpc->crpc_lock);
1199                 srpc_abort_rpc(rpc, rc);
1200                 spin_unlock(&rpc->crpc_lock);
1201         }
1202
1203 abort:
1204         if (rpc->crpc_aborted) {
1205                 LNetMDUnlink(rpc->crpc_reqstmdh);
1206                 LNetMDUnlink(rpc->crpc_replymdh);
1207                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1208
1209                 if (!srpc_event_pending(rpc)) {
1210                         srpc_client_rpc_done(rpc, -EINTR);
1211                         return 1;
1212                 }
1213         }
1214         return 0;
1215 }
1216
1217 srpc_client_rpc_t *
1218 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1219                         int nbulkiov, int bulklen,
1220                         void (*rpc_done)(srpc_client_rpc_t *),
1221                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1222 {
1223         srpc_client_rpc_t *rpc;
1224
1225         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1226                                    crpc_bulk.bk_iovs[nbulkiov]));
1227         if (rpc == NULL)
1228                 return NULL;
1229
1230         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1231                              bulklen, rpc_done, rpc_fini, priv);
1232         return rpc;
1233 }
1234
1235 /* called with rpc->crpc_lock held */
1236 static inline void
1237 srpc_queue_rpc (srpc_peer_t *peer, srpc_client_rpc_t *rpc)
1238 {
1239         int service = rpc->crpc_service;
1240
1241         LASSERT (peer->stp_nid == rpc->crpc_dest.nid);
1242         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1243
1244         rpc->crpc_peer = peer;
1245
1246         spin_lock(&peer->stp_lock);
1247
1248         /* Framework RPCs that alter session state shall take precedence
1249          * over test RPCs and framework query RPCs */
1250         if (service <= SRPC_FRAMEWORK_SERVICE_MAX_ID &&
1251             service != SRPC_SERVICE_DEBUG &&
1252             service != SRPC_SERVICE_QUERY_STAT)
1253                 list_add_tail(&rpc->crpc_privl, &peer->stp_ctl_rpcq);
1254         else
1255                 list_add_tail(&rpc->crpc_privl, &peer->stp_rpcq);
1256
1257         srpc_client_rpc_addref(rpc); /* ++ref for peer->*rpcq */
1258         spin_unlock(&peer->stp_lock);
1259         return;
1260 }
1261
1262 /* called with rpc->crpc_lock held */
1263 void
1264 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1265 {
1266         srpc_peer_t *peer = rpc->crpc_peer;
1267
1268         LASSERT (why != 0);
1269
1270         if (rpc->crpc_aborted || /* already aborted */
1271             rpc->crpc_closed)    /* callback imminent */
1272                 return;
1273
1274         CDEBUG (D_NET,
1275                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1276                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1277                 swi_state2str(rpc->crpc_wi.wi_state), why);
1278
1279         rpc->crpc_aborted = 1;
1280         rpc->crpc_status  = why;
1281
1282         if (peer != NULL) {
1283                 spin_lock(&peer->stp_lock);
1284
1285                 if (!list_empty(&rpc->crpc_privl)) { /* still queued */
1286                         list_del_init(&rpc->crpc_privl);
1287                         srpc_client_rpc_decref(rpc); /* --ref for peer->*rpcq */
1288                         rpc->crpc_peer = NULL;       /* no credit taken */
1289                 }
1290
1291                 spin_unlock(&peer->stp_lock);
1292         }
1293
1294         swi_schedule_workitem(&rpc->crpc_wi);
1295         return;
1296 }
1297
1298 /* called with rpc->crpc_lock held */
1299 void
1300 srpc_post_rpc (srpc_client_rpc_t *rpc)
1301 {
1302         srpc_peer_t *peer;
1303
1304         LASSERT (!rpc->crpc_aborted);
1305         LASSERT (rpc->crpc_peer == NULL);
1306         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1307         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1308
1309         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1310                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1311                 rpc->crpc_timeout);
1312
1313         srpc_add_client_rpc_timer(rpc);
1314
1315         peer = srpc_nid2peer(rpc->crpc_dest.nid);
1316         if (peer == NULL) {
1317                 srpc_abort_rpc(rpc, -ENOMEM);
1318                 return;
1319         }
1320
1321         srpc_queue_rpc(peer, rpc);
1322
1323         spin_unlock(&rpc->crpc_lock);
1324         srpc_check_sends(peer, 0);
1325         spin_lock(&rpc->crpc_lock);
1326         return;
1327 }
1328
1329
1330 int
1331 srpc_send_reply (srpc_server_rpc_t *rpc)
1332 {
1333         srpc_event_t   *ev = &rpc->srpc_ev;
1334         srpc_msg_t     *msg = &rpc->srpc_replymsg;
1335         srpc_buffer_t  *buffer = rpc->srpc_reqstbuf;
1336         srpc_service_t *sv = rpc->srpc_service;
1337         __u64           rpyid;
1338         int             rc;
1339
1340         LASSERT (buffer != NULL);
1341         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1342
1343         spin_lock(&sv->sv_lock);
1344
1345         if (!sv->sv_shuttingdown &&
1346             sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1347                 /* Repost buffer before replying since test client
1348                  * might send me another RPC once it gets the reply */
1349                 if (srpc_service_post_buffer(sv, buffer) != 0)
1350                         CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1351                 rpc->srpc_reqstbuf = NULL;
1352         }
1353
1354         spin_unlock(&sv->sv_lock);
1355
1356         ev->ev_fired = 0;
1357         ev->ev_data  = rpc;
1358         ev->ev_type  = SRPC_REPLY_SENT;
1359
1360         msg->msg_magic   = SRPC_MSG_MAGIC;
1361         msg->msg_version = SRPC_MSG_VERSION;
1362         msg->msg_type    = srpc_service2reply(sv->sv_id);
1363
1364         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1365                                    sizeof(*msg), LNET_MD_OP_PUT,
1366                                    rpc->srpc_peer, rpc->srpc_self,
1367                                    &rpc->srpc_replymdh, ev);
1368         if (rc != 0)
1369                 ev->ev_fired = 1;  /* no more event expected */
1370         return rc;
1371 }
1372
1373 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1374 void
1375 srpc_lnet_ev_handler (lnet_event_t *ev)
1376 {
1377         srpc_event_t      *rpcev = ev->md.user_ptr;
1378         srpc_client_rpc_t *crpc;
1379         srpc_server_rpc_t *srpc;
1380         srpc_buffer_t     *buffer;
1381         srpc_service_t    *sv;
1382         srpc_msg_t        *msg;
1383         srpc_msg_type_t    type;
1384
1385         LASSERT (!in_interrupt());
1386
1387         if (ev->status != 0) {
1388                 spin_lock(&srpc_data.rpc_glock);
1389                 srpc_data.rpc_counters.errors++;
1390                 spin_unlock(&srpc_data.rpc_glock);
1391         }
1392
1393         rpcev->ev_lnet = ev->type;
1394
1395         switch (rpcev->ev_type) {
1396         default:
1397                 LBUG ();
1398         case SRPC_REQUEST_SENT:
1399                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1400                         spin_lock(&srpc_data.rpc_glock);
1401                         srpc_data.rpc_counters.rpcs_sent++;
1402                         spin_unlock(&srpc_data.rpc_glock);
1403                 }
1404         case SRPC_REPLY_RCVD:
1405         case SRPC_BULK_REQ_RCVD:
1406                 crpc = rpcev->ev_data;
1407
1408                 LASSERT (rpcev == &crpc->crpc_reqstev ||
1409                          rpcev == &crpc->crpc_replyev ||
1410                          rpcev == &crpc->crpc_bulkev);
1411
1412                 spin_lock(&crpc->crpc_lock);
1413
1414                 LASSERT (rpcev->ev_fired == 0);
1415                 rpcev->ev_fired  = 1;
1416                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1417                                                 -EINTR : ev->status;
1418                 swi_schedule_workitem(&crpc->crpc_wi);
1419
1420                 spin_unlock(&crpc->crpc_lock);
1421                 break;
1422
1423         case SRPC_REQUEST_RCVD:
1424                 sv = rpcev->ev_data;
1425
1426                 LASSERT (rpcev == &sv->sv_ev);
1427
1428                 spin_lock(&sv->sv_lock);
1429
1430                 LASSERT (ev->unlinked);
1431                 LASSERT (ev->type == LNET_EVENT_PUT ||
1432                          ev->type == LNET_EVENT_UNLINK);
1433                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1434                          sv->sv_shuttingdown);
1435
1436                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1437                 buffer->buf_peer = ev->initiator;
1438                 buffer->buf_self = ev->target.nid;
1439
1440                 sv->sv_nposted_msg--;
1441                 LASSERT (sv->sv_nposted_msg >= 0);
1442
1443                 if (sv->sv_shuttingdown) {
1444                         /* Leave buffer on sv->sv_posted_msgq since
1445                          * srpc_finish_service needs to traverse it. */
1446                         spin_unlock(&sv->sv_lock);
1447                         break;
1448                 }
1449
1450                 list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1451                 msg = &buffer->buf_msg;
1452                 type = srpc_service2request(sv->sv_id);
1453
1454                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1455                     (msg->msg_type != type &&
1456                      msg->msg_type != __swab32(type)) ||
1457                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1458                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1459                         CERROR ("Dropping RPC (%s) from %s: "
1460                                 "status %d mlength %d type %u magic %u.\n",
1461                                 sv->sv_name, libcfs_id2str(ev->initiator),
1462                                 ev->status, ev->mlength,
1463                                 msg->msg_type, msg->msg_magic);
1464
1465                         /* NB might drop sv_lock in srpc_service_recycle_buffer,
1466                          * sv_nposted_msg++ as an implicit reference to prevent
1467                          * sv from disappearing under me */
1468                         sv->sv_nposted_msg++;
1469                         srpc_service_recycle_buffer(sv, buffer);
1470                         sv->sv_nposted_msg--;
1471                         spin_unlock(&sv->sv_lock);
1472
1473                         if (ev->status == 0) { /* status!=0 counted already */
1474                                 spin_lock(&srpc_data.rpc_glock);
1475                                 srpc_data.rpc_counters.errors++;
1476                                 spin_unlock(&srpc_data.rpc_glock);
1477                         }
1478                         break;
1479                 }
1480
1481                 if (!list_empty(&sv->sv_free_rpcq)) {
1482                         srpc = list_entry(sv->sv_free_rpcq.next,
1483                                           srpc_server_rpc_t, srpc_list);
1484                         list_del(&srpc->srpc_list);
1485
1486                         srpc_init_server_rpc(srpc, sv, buffer);
1487                         list_add_tail(&srpc->srpc_list, &sv->sv_active_rpcq);
1488                         srpc_schedule_server_rpc(srpc);
1489                 } else {
1490                         list_add_tail(&buffer->buf_list, &sv->sv_blocked_msgq);
1491                 }
1492
1493                 spin_unlock(&sv->sv_lock);
1494
1495                 spin_lock(&srpc_data.rpc_glock);
1496                 srpc_data.rpc_counters.rpcs_rcvd++;
1497                 spin_unlock(&srpc_data.rpc_glock);
1498                 break;
1499
1500         case SRPC_BULK_GET_RPLD:
1501                 LASSERT (ev->type == LNET_EVENT_SEND ||
1502                          ev->type == LNET_EVENT_REPLY ||
1503                          ev->type == LNET_EVENT_UNLINK);
1504
1505                 if (ev->type == LNET_EVENT_SEND &&
1506                     ev->status == 0 && !ev->unlinked)
1507                         break; /* wait for the final LNET_EVENT_REPLY */
1508
1509         case SRPC_BULK_PUT_SENT:
1510                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1511                         spin_lock(&srpc_data.rpc_glock);
1512
1513                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1514                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1515                         else
1516                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1517
1518                         spin_unlock(&srpc_data.rpc_glock);
1519                 }
1520         case SRPC_REPLY_SENT:
1521                 srpc = rpcev->ev_data;
1522                 sv = srpc->srpc_service;
1523
1524                 LASSERT (rpcev == &srpc->srpc_ev);
1525
1526                 spin_lock(&sv->sv_lock);
1527                 rpcev->ev_fired  = 1;
1528                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1529                                                 -EINTR : ev->status;
1530                 srpc_schedule_server_rpc(srpc);
1531                 spin_unlock(&sv->sv_lock);
1532                 break;
1533         }
1534
1535         return;
1536 }
1537
1538 #ifndef __KERNEL__
1539
1540 int
1541 srpc_check_event (int timeout)
1542 {
1543         lnet_event_t ev;
1544         int          rc;
1545         int          i;
1546
1547         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1548                         timeout * 1000, &ev, &i);
1549         if (rc == 0) return 0;
1550
1551         LASSERT (rc == -EOVERFLOW || rc == 1);
1552
1553         /* We can't affort to miss any events... */
1554         if (rc == -EOVERFLOW) {
1555                 CERROR ("Dropped an event!!!\n");
1556                 abort();
1557         }
1558
1559         srpc_lnet_ev_handler(&ev);
1560         return 1;
1561 }
1562
1563 #endif
1564
1565 int
1566 srpc_startup (void)
1567 {
1568         int i;
1569         int rc;
1570
1571 #ifndef __KERNEL__
1572         char *s;
1573
1574         s = getenv("SRPC_PEER_CREDITS");
1575         srpc_peer_credits = (s != NULL) ? atoi(s) : srpc_peer_credits;
1576 #endif
1577
1578         if (srpc_peer_credits <= 0) {
1579                 CERROR("Peer credits must be positive: %d\n", srpc_peer_credits);
1580                 return -EINVAL;
1581         }
1582
1583         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1584         spin_lock_init(&srpc_data.rpc_glock);
1585
1586         /* 1 second pause to avoid timestamp reuse */
1587         cfs_pause(cfs_time_seconds(1));
1588         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1589
1590         srpc_data.rpc_state = SRPC_STATE_NONE;
1591
1592         LIBCFS_ALLOC(srpc_data.rpc_peers,
1593                      sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1594         if (srpc_data.rpc_peers == NULL) {
1595                 CERROR ("Failed to alloc peer hash.\n");
1596                 return -ENOMEM;
1597         }
1598
1599         for (i = 0; i < SRPC_PEER_HASH_SIZE; i++)
1600                 CFS_INIT_LIST_HEAD(&srpc_data.rpc_peers[i]);
1601
1602 #ifdef __KERNEL__
1603         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1604 #else
1605         if (the_lnet.ln_server_mode_flag)
1606                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1607         else
1608                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1609 #endif
1610         if (rc < 0) {
1611                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1612                 LIBCFS_FREE(srpc_data.rpc_peers,
1613                             sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1614                 return rc;
1615         }
1616
1617         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1618
1619         srpc_data.rpc_lnet_eq = LNET_EQ_NONE;
1620 #ifdef __KERNEL__
1621         rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1622 #else
1623         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1624 #endif
1625         if (rc != 0) {
1626                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1627                 goto bail;
1628         }
1629
1630         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1631         LASSERT (rc == 0);
1632
1633         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1634
1635         rc = swi_startup();
1636         if (rc != 0)
1637                 goto bail;
1638
1639         srpc_data.rpc_state = SRPC_STATE_WI_INIT;
1640
1641         rc = stt_startup();
1642
1643 bail:
1644         if (rc != 0)
1645                 srpc_shutdown();
1646         else
1647                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1648
1649         return rc;
1650 }
1651
1652 void
1653 srpc_shutdown (void)
1654 {
1655         int i;
1656         int rc;
1657         int state;
1658
1659         state = srpc_data.rpc_state;
1660         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1661
1662         switch (state) {
1663         default:
1664                 LBUG ();
1665         case SRPC_STATE_RUNNING:
1666                 spin_lock(&srpc_data.rpc_glock);
1667
1668                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1669                         srpc_service_t *sv = srpc_data.rpc_services[i];
1670
1671                         LASSERTF (sv == NULL,
1672                                   "service not empty: id %d, name %s\n",
1673                                   i, sv->sv_name);
1674                 }
1675
1676                 spin_unlock(&srpc_data.rpc_glock);
1677
1678                 stt_shutdown();
1679
1680         case SRPC_STATE_WI_INIT:
1681                 swi_shutdown();
1682
1683         case SRPC_STATE_EQ_INIT:
1684                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1685                 LASSERT (rc == 0);
1686                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1687                 LASSERT (rc == 0); /* the EQ should have no user by now */
1688
1689         case SRPC_STATE_NI_INIT:
1690                 LNetNIFini();
1691                 break;
1692         }
1693
1694         /* srpc_peer_t's are kept in hash until shutdown */
1695         for (i = 0; i < SRPC_PEER_HASH_SIZE; i++) {
1696                 srpc_peer_t *peer;
1697
1698                 while (!list_empty(&srpc_data.rpc_peers[i])) {
1699                         peer = list_entry(srpc_data.rpc_peers[i].next,
1700                                           srpc_peer_t, stp_list);
1701                         list_del(&peer->stp_list);
1702
1703                         LASSERT (list_empty(&peer->stp_rpcq));
1704                         LASSERT (list_empty(&peer->stp_ctl_rpcq));
1705                         LASSERT (peer->stp_credits == srpc_peer_credits);
1706
1707                         LIBCFS_FREE(peer, sizeof(srpc_peer_t));
1708                 }
1709         }
1710
1711         LIBCFS_FREE(srpc_data.rpc_peers,
1712                     sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1713         return;
1714 }