lnet/selftest/rpc.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/selftest/rpc.c
  37  *
  38  * Author: Isaac Huang <isaac@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include "selftest.h"
  44
  45
  46 typedef enum {
  47         SRPC_STATE_NONE,
  48         SRPC_STATE_NI_INIT,
  49         SRPC_STATE_EQ_INIT,
  50         SRPC_STATE_RUNNING,
  51         SRPC_STATE_STOPPING,
  52 } srpc_state_t;
  53
  54 struct smoketest_rpc {
  55         cfs_spinlock_t    rpc_glock;     /* global lock */
  56         srpc_service_t   *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  57         lnet_handle_eq_t  rpc_lnet_eq;   /* _the_ LNet event queue */
  58         srpc_state_t      rpc_state;
  59         srpc_counters_t   rpc_counters;
  60         __u64             rpc_matchbits; /* matchbits counter */
  61 } srpc_data;
  62
  63 /* forward ref's */
  64 int srpc_handle_rpc (swi_workitem_t *wi);
  65
  66 void srpc_get_counters (srpc_counters_t *cnt)
  67 {
  68         cfs_spin_lock(&srpc_data.rpc_glock);
  69         *cnt = srpc_data.rpc_counters;
  70         cfs_spin_unlock(&srpc_data.rpc_glock);
  71 }
  72
  73 void srpc_set_counters (const srpc_counters_t *cnt)
  74 {
  75         cfs_spin_lock(&srpc_data.rpc_glock);
  76         srpc_data.rpc_counters = *cnt;
  77         cfs_spin_unlock(&srpc_data.rpc_glock);
  78 }
  79
  80 void
  81 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
  82 {
  83         LASSERT (i >= 0 && i < bk->bk_niov);
  84
  85 #ifdef __KERNEL__
  86         bk->bk_iovs[i].kiov_offset = 0;
  87         bk->bk_iovs[i].kiov_page   = pg;
  88         bk->bk_iovs[i].kiov_len    = CFS_PAGE_SIZE;
  89 #else
  90         LASSERT (bk->bk_pages != NULL);
  91
  92         bk->bk_pages[i] = pg;
  93         bk->bk_iovs[i].iov_len  = CFS_PAGE_SIZE;
  94         bk->bk_iovs[i].iov_base = cfs_page_address(pg);
  95 #endif
  96         return;
  97 }
  98
  99 void
 100 srpc_free_bulk (srpc_bulk_t *bk)
 101 {
 102         int         i;
 103         cfs_page_t *pg;
 104
 105         LASSERT (bk != NULL);
 106 #ifndef __KERNEL__
 107         LASSERT (bk->bk_pages != NULL);
 108 #endif
 109
 110         for (i = 0; i < bk->bk_niov; i++) {
 111 #ifdef __KERNEL__
 112                 pg = bk->bk_iovs[i].kiov_page;
 113 #else
 114                 pg = bk->bk_pages[i];
 115 #endif
 116                 if (pg == NULL) break;
 117
 118                 cfs_free_page(pg);
 119         }
 120
 121 #ifndef __KERNEL__
 122         LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
 123 #endif
 124         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 125         return;
 126 }
 127
 128 srpc_bulk_t *
 129 srpc_alloc_bulk (int npages, int sink)
 130 {
 131         srpc_bulk_t  *bk;
 132         cfs_page_t  **pages;
 133         int           i;
 134
 135         LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
 136
 137         LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 138         if (bk == NULL) {
 139                 CERROR ("Can't allocate descriptor for %d pages\n", npages);
 140                 return NULL;
 141         }
 142
 143         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
 144         bk->bk_sink = sink;
 145         bk->bk_niov = npages;
 146         bk->bk_len  = npages * CFS_PAGE_SIZE;
 147 #ifndef __KERNEL__
 148         LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
 149         if (pages == NULL) {
 150                 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 151                 CERROR ("Can't allocate page array for %d pages\n", npages);
 152                 return NULL;
 153         }
 154
 155         memset(pages, 0, sizeof(cfs_page_t *) * npages);
 156         bk->bk_pages = pages;
 157 #else
 158         UNUSED (pages);
 159 #endif
 160
 161         for (i = 0; i < npages; i++) {
 162                 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
 163
 164                 if (pg == NULL) {
 165                         CERROR ("Can't allocate page %d of %d\n", i, npages);
 166                         srpc_free_bulk(bk);
 167                         return NULL;
 168                 }
 169
 170                 srpc_add_bulk_page(bk, pg, i);
 171         }
 172
 173         return bk;
 174 }
 175
 176 static inline __u64
 177 srpc_next_id (void)
 178 {
 179         __u64 id;
 180
 181         cfs_spin_lock(&srpc_data.rpc_glock);
 182         id = srpc_data.rpc_matchbits++;
 183         cfs_spin_unlock(&srpc_data.rpc_glock);
 184         return id;
 185 }
 186
 187 void
 188 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
 189                       srpc_service_t *sv, srpc_buffer_t *buffer)
 190 {
 191         memset(rpc, 0, sizeof(*rpc));
 192         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
 193                           sv->sv_id <= SRPC_FRAMEWORK_SERVICE_MAX_ID ?
 194                           CFS_WI_SCHED_SERIAL : CFS_WI_SCHED_ANY);
 195
 196         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 197
 198         rpc->srpc_service  = sv;
 199         rpc->srpc_reqstbuf = buffer;
 200         rpc->srpc_peer     = buffer->buf_peer;
 201         rpc->srpc_self     = buffer->buf_self;
 202         LNetInvalidateHandle(&rpc->srpc_replymdh);
 203 }
 204
 205 int
 206 srpc_add_service (srpc_service_t *sv)
 207 {
 208         int                id = sv->sv_id;
 209         int                i;
 210         srpc_server_rpc_t *rpc;
 211
 212         LASSERT (sv->sv_concur > 0);
 213         LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
 214
 215         cfs_spin_lock(&srpc_data.rpc_glock);
 216
 217         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
 218
 219         if (srpc_data.rpc_services[id] != NULL) {
 220                 cfs_spin_unlock(&srpc_data.rpc_glock);
 221                 return -EBUSY;
 222         }
 223
 224         srpc_data.rpc_services[id] = sv;
 225         cfs_spin_unlock(&srpc_data.rpc_glock);
 226
 227         sv->sv_nprune       = 0;
 228         sv->sv_nposted_msg  = 0;
 229         sv->sv_shuttingdown = 0;
 230         cfs_spin_lock_init(&sv->sv_lock);
 231         CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
 232         CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
 233         CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
 234         CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
 235
 236         sv->sv_ev.ev_data = sv;
 237         sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
 238
 239         for (i = 0; i < sv->sv_concur; i++) {
 240                 LIBCFS_ALLOC(rpc, sizeof(*rpc));
 241                 if (rpc == NULL) goto enomem;
 242
 243                 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 244         }
 245
 246         CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
 247                 id, sv->sv_name, sv->sv_concur);
 248         return 0;
 249
 250 enomem:
 251         while (!cfs_list_empty(&sv->sv_free_rpcq)) {
 252                 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
 253                                      srpc_server_rpc_t, srpc_list);
 254                 cfs_list_del(&rpc->srpc_list);
 255                 LIBCFS_FREE(rpc, sizeof(*rpc));
 256         }
 257
 258         cfs_spin_lock(&srpc_data.rpc_glock);
 259         srpc_data.rpc_services[id] = NULL;
 260         cfs_spin_unlock(&srpc_data.rpc_glock);
 261         return -ENOMEM;
 262 }
 263
 264 int
 265 srpc_remove_service (srpc_service_t *sv)
 266 {
 267         int id = sv->sv_id;
 268
 269         cfs_spin_lock(&srpc_data.rpc_glock);
 270
 271         if (srpc_data.rpc_services[id] != sv) {
 272                 cfs_spin_unlock(&srpc_data.rpc_glock);
 273                 return -ENOENT;
 274         }
 275
 276         srpc_data.rpc_services[id] = NULL;
 277         cfs_spin_unlock(&srpc_data.rpc_glock);
 278         return 0;
 279 }
 280
 281 int
 282 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
 283                        int len, int options, lnet_process_id_t peer,
 284                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 285 {
 286         int              rc;
 287         lnet_md_t        md;
 288         lnet_handle_me_t meh;
 289
 290         rc = LNetMEAttach(portal, peer, matchbits, 0,
 291                           LNET_UNLINK, LNET_INS_AFTER, &meh);
 292         if (rc != 0) {
 293                 CERROR ("LNetMEAttach failed: %d\n", rc);
 294                 LASSERT (rc == -ENOMEM);
 295                 return -ENOMEM;
 296         }
 297
 298         md.threshold = 1;
 299         md.user_ptr  = ev;
 300         md.start     = buf;
 301         md.length    = len;
 302         md.options   = options;
 303         md.eq_handle = srpc_data.rpc_lnet_eq;
 304
 305         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 306         if (rc != 0) {
 307                 CERROR ("LNetMDAttach failed: %d\n", rc);
 308                 LASSERT (rc == -ENOMEM);
 309
 310                 rc = LNetMEUnlink(meh);
 311                 LASSERT (rc == 0);
 312                 return -ENOMEM;
 313         }
 314
 315         CDEBUG (D_NET,
 316                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 317                 libcfs_id2str(peer), portal, matchbits);
 318         return 0;
 319 }
 320
 321 int
 322 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 323                       int options, lnet_process_id_t peer, lnet_nid_t self,
 324                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 325 {
 326         int       rc;
 327         lnet_md_t md;
 328
 329         md.user_ptr  = ev;
 330         md.start     = buf;
 331         md.length    = len;
 332         md.eq_handle = srpc_data.rpc_lnet_eq;
 333         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 334         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 335
 336         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 337         if (rc != 0) {
 338                 CERROR ("LNetMDBind failed: %d\n", rc);
 339                 LASSERT (rc == -ENOMEM);
 340                 return -ENOMEM;
 341         }
 342
 343         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 344          * they're only meaningful for MDs attached to an ME (i.e. passive
 345          * buffers... */
 346         if ((options & LNET_MD_OP_PUT) != 0) {
 347                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 348                              portal, matchbits, 0, 0);
 349         } else {
 350                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 351
 352                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 353         }
 354
 355         if (rc != 0) {
 356                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 357                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 358                         libcfs_id2str(peer), portal, matchbits, rc);
 359
 360                 /* The forthcoming unlink event will complete this operation
 361                  * with failure, so fall through and return success here.
 362                  */
 363                 rc = LNetMDUnlink(*mdh);
 364                 LASSERT (rc == 0);
 365         } else {
 366                 CDEBUG (D_NET,
 367                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 368                         libcfs_id2str(peer), portal, matchbits);
 369         }
 370         return 0;
 371 }
 372
 373 int
 374 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 375                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 376 {
 377         int rc;
 378         int portal;
 379
 380         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 381                 portal = SRPC_REQUEST_PORTAL;
 382         else
 383                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 384
 385         rc = srpc_post_active_rdma(portal, service, buf, len,
 386                                    LNET_MD_OP_PUT, peer,
 387                                    LNET_NID_ANY, mdh, ev);
 388         return rc;
 389 }
 390
 391 int
 392 srpc_post_passive_rqtbuf(int service, void *buf, int len,
 393                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 394 {
 395         int               rc;
 396         int               portal;
 397         lnet_process_id_t any = {0};
 398
 399         any.nid = LNET_NID_ANY;
 400         any.pid = LNET_PID_ANY;
 401
 402         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 403                 portal = SRPC_REQUEST_PORTAL;
 404         else
 405                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 406
 407         rc = srpc_post_passive_rdma(portal, service, buf, len,
 408                                     LNET_MD_OP_PUT, any, mdh, ev);
 409         return rc;
 410 }
 411
 412 int
 413 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 414 {
 415         srpc_msg_t *msg = &buf->buf_msg;
 416         int         rc;
 417
 418         LASSERT (!sv->sv_shuttingdown);
 419
 420         LNetInvalidateHandle(&buf->buf_mdh);
 421         cfs_list_add(&buf->buf_list, &sv->sv_posted_msgq);
 422         sv->sv_nposted_msg++;
 423         cfs_spin_unlock(&sv->sv_lock);
 424
 425         rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
 426                                       &buf->buf_mdh, &sv->sv_ev);
 427
 428         /* At this point, a RPC (new or delayed) may have arrived in
 429          * msg and its event handler has been called. So we must add
 430          * buf to sv_posted_msgq _before_ dropping sv_lock */
 431
 432         cfs_spin_lock(&sv->sv_lock);
 433
 434         if (rc == 0) {
 435                 if (sv->sv_shuttingdown) {
 436                         cfs_spin_unlock(&sv->sv_lock);
 437
 438                         /* srpc_shutdown_service might have tried to unlink me
 439                          * when my buf_mdh was still invalid */
 440                         LNetMDUnlink(buf->buf_mdh);
 441
 442                         cfs_spin_lock(&sv->sv_lock);
 443                 }
 444                 return 0;
 445         }
 446
 447         sv->sv_nposted_msg--;
 448         if (sv->sv_shuttingdown) return rc;
 449
 450         cfs_list_del(&buf->buf_list);
 451
 452         cfs_spin_unlock(&sv->sv_lock);
 453         LIBCFS_FREE(buf, sizeof(*buf));
 454         cfs_spin_lock(&sv->sv_lock);
 455         return rc;
 456 }
 457
 458 int
 459 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
 460 {
 461         int                rc;
 462         int                posted;
 463         srpc_buffer_t     *buf;
 464
 465         LASSERTF (nbuffer > 0,
 466                   "nbuffer must be positive: %d\n", nbuffer);
 467
 468         for (posted = 0; posted < nbuffer; posted++) {
 469                 LIBCFS_ALLOC(buf, sizeof(*buf));
 470                 if (buf == NULL) break;
 471
 472                 cfs_spin_lock(&sv->sv_lock);
 473                 rc = srpc_service_post_buffer(sv, buf);
 474                 cfs_spin_unlock(&sv->sv_lock);
 475
 476                 if (rc != 0) break;
 477         }
 478
 479         return posted;
 480 }
 481
 482 void
 483 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
 484 {
 485         LASSERTF (nbuffer > 0,
 486                   "nbuffer must be positive: %d\n", nbuffer);
 487
 488         cfs_spin_lock(&sv->sv_lock);
 489
 490         LASSERT (sv->sv_nprune >= 0);
 491         LASSERT (!sv->sv_shuttingdown);
 492
 493         sv->sv_nprune += nbuffer;
 494
 495         cfs_spin_unlock(&sv->sv_lock);
 496         return;
 497 }
 498
 499 /* returns 1 if sv has finished, otherwise 0 */
 500 int
 501 srpc_finish_service (srpc_service_t *sv)
 502 {
 503         srpc_server_rpc_t *rpc;
 504         srpc_buffer_t     *buf;
 505
 506         cfs_spin_lock(&sv->sv_lock);
 507
 508         LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
 509
 510         if (sv->sv_nposted_msg != 0 || !cfs_list_empty(&sv->sv_active_rpcq)) {
 511                 CDEBUG (D_NET,
 512                         "waiting for %d posted buffers to unlink and "
 513                         "in-flight RPCs to die.\n",
 514                         sv->sv_nposted_msg);
 515
 516                 if (!cfs_list_empty(&sv->sv_active_rpcq)) {
 517                         rpc = cfs_list_entry(sv->sv_active_rpcq.next,
 518                                              srpc_server_rpc_t, srpc_list);
 519                         CDEBUG (D_NETERROR,
 520                                 "Active RPC %p on shutdown: sv %s, peer %s, "
 521                                 "wi %s scheduled %d running %d, "
 522                                 "ev fired %d type %d status %d lnet %d\n",
 523                                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 524                                 swi_state2str(rpc->srpc_wi.swi_state),
 525                                 rpc->srpc_wi.swi_workitem.wi_scheduled,
 526                                 rpc->srpc_wi.swi_workitem.wi_running,
 527                                 rpc->srpc_ev.ev_fired,
 528                                 rpc->srpc_ev.ev_type,
 529                                 rpc->srpc_ev.ev_status,
 530                                 rpc->srpc_ev.ev_lnet);
 531                 }
 532
 533                 cfs_spin_unlock(&sv->sv_lock);
 534                 return 0;
 535         }
 536
 537         cfs_spin_unlock(&sv->sv_lock); /* no lock needed from now on */
 538
 539         for (;;) {
 540                 cfs_list_t *q;
 541
 542                 if (!cfs_list_empty(&sv->sv_posted_msgq))
 543                         q = &sv->sv_posted_msgq;
 544                 else if (!cfs_list_empty(&sv->sv_blocked_msgq))
 545                         q = &sv->sv_blocked_msgq;
 546                 else
 547                         break;
 548
 549                 buf = cfs_list_entry(q->next, srpc_buffer_t, buf_list);
 550                 cfs_list_del(&buf->buf_list);
 551
 552                 LIBCFS_FREE(buf, sizeof(*buf));
 553         }
 554
 555         while (!cfs_list_empty(&sv->sv_free_rpcq)) {
 556                 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
 557                                      srpc_server_rpc_t, srpc_list);
 558                 cfs_list_del(&rpc->srpc_list);
 559                 LIBCFS_FREE(rpc, sizeof(*rpc));
 560         }
 561
 562         return 1;
 563 }
 564
 565 /* called with sv->sv_lock held */
 566 void
 567 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 568 {
 569         if (sv->sv_shuttingdown) goto free;
 570
 571         if (sv->sv_nprune == 0) {
 572                 if (srpc_service_post_buffer(sv, buf) != 0)
 573                         CWARN ("Failed to post %s buffer\n", sv->sv_name);
 574                 return;
 575         }
 576
 577         sv->sv_nprune--;
 578 free:
 579         cfs_spin_unlock(&sv->sv_lock);
 580         LIBCFS_FREE(buf, sizeof(*buf));
 581         cfs_spin_lock(&sv->sv_lock);
 582 }
 583
 584 /* called with srpc_service_t::sv_lock held */
 585 inline void
 586 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
 587 {
 588         swi_schedule_workitem(&rpc->srpc_wi);
 589 }
 590
 591 void
 592 srpc_abort_service (srpc_service_t *sv)
 593 {
 594         srpc_server_rpc_t *rpc;
 595
 596         cfs_spin_lock(&sv->sv_lock);
 597
 598         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 599                sv->sv_id, sv->sv_name);
 600
 601         /* schedule in-flight RPCs to notice the abort, NB:
 602          * racing with incoming RPCs; complete fix should make test
 603          * RPCs carry session ID in its headers */
 604         cfs_list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
 605                 rpc->srpc_aborted = 1;
 606                 srpc_schedule_server_rpc(rpc);
 607         }
 608
 609         cfs_spin_unlock(&sv->sv_lock);
 610         return;
 611 }
 612
 613 void
 614 srpc_shutdown_service (srpc_service_t *sv)
 615 {
 616         srpc_server_rpc_t *rpc;
 617         srpc_buffer_t     *buf;
 618
 619         cfs_spin_lock(&sv->sv_lock);
 620
 621         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 622                sv->sv_id, sv->sv_name);
 623
 624         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 625
 626         /* schedule in-flight RPCs to notice the shutdown */
 627         cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
 628                                        srpc_server_rpc_t, srpc_list) {
 629                 srpc_schedule_server_rpc(rpc);
 630         }
 631
 632         cfs_spin_unlock(&sv->sv_lock);
 633
 634         /* OK to traverse sv_posted_msgq without lock, since no one
 635          * touches sv_posted_msgq now */
 636         cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
 637                                        srpc_buffer_t, buf_list)
 638                 LNetMDUnlink(buf->buf_mdh);
 639
 640         return;
 641 }
 642
 643 int
 644 srpc_send_request (srpc_client_rpc_t *rpc)
 645 {
 646         srpc_event_t *ev = &rpc->crpc_reqstev;
 647         int           rc;
 648
 649         ev->ev_fired = 0;
 650         ev->ev_data  = rpc;
 651         ev->ev_type  = SRPC_REQUEST_SENT;
 652
 653         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 654                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 655                                      &rpc->crpc_reqstmdh, ev);
 656         if (rc != 0) {
 657                 LASSERT (rc == -ENOMEM);
 658                 ev->ev_fired = 1;  /* no more event expected */
 659         }
 660         return rc;
 661 }
 662
 663 int
 664 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 665 {
 666         srpc_event_t *ev = &rpc->crpc_replyev;
 667         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 668         int           rc;
 669
 670         ev->ev_fired = 0;
 671         ev->ev_data  = rpc;
 672         ev->ev_type  = SRPC_REPLY_RCVD;
 673
 674         *id = srpc_next_id();
 675
 676         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 677                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 678                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 679                                     &rpc->crpc_replymdh, ev);
 680         if (rc != 0) {
 681                 LASSERT (rc == -ENOMEM);
 682                 ev->ev_fired = 1;  /* no more event expected */
 683         }
 684         return rc;
 685 }
 686
 687 int
 688 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 689 {
 690         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 691         srpc_event_t *ev = &rpc->crpc_bulkev;
 692         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 693         int           rc;
 694         int           opt;
 695
 696         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 697
 698         if (bk->bk_niov == 0) return 0; /* nothing to do */
 699
 700         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 701 #ifdef __KERNEL__
 702         opt |= LNET_MD_KIOV;
 703 #else
 704         opt |= LNET_MD_IOVEC;
 705 #endif
 706
 707         ev->ev_fired = 0;
 708         ev->ev_data  = rpc;
 709         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 710
 711         *id = srpc_next_id();
 712
 713         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 714                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 715                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 716         if (rc != 0) {
 717                 LASSERT (rc == -ENOMEM);
 718                 ev->ev_fired = 1;  /* no more event expected */
 719         }
 720         return rc;
 721 }
 722
 723 int
 724 srpc_do_bulk (srpc_server_rpc_t *rpc)
 725 {
 726         srpc_event_t  *ev = &rpc->srpc_ev;
 727         srpc_bulk_t   *bk = rpc->srpc_bulk;
 728         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 729         int            rc;
 730         int            opt;
 731
 732         LASSERT (bk != NULL);
 733
 734         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 735 #ifdef __KERNEL__
 736         opt |= LNET_MD_KIOV;
 737 #else
 738         opt |= LNET_MD_IOVEC;
 739 #endif
 740
 741         ev->ev_fired = 0;
 742         ev->ev_data  = rpc;
 743         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 744
 745         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 746                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 747                                    rpc->srpc_peer, rpc->srpc_self,
 748                                    &bk->bk_mdh, ev);
 749         if (rc != 0)
 750                 ev->ev_fired = 1;  /* no more event expected */
 751         return rc;
 752 }
 753
 754 /* only called from srpc_handle_rpc */
 755 void
 756 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
 757 {
 758         srpc_service_t *sv = rpc->srpc_service;
 759         srpc_buffer_t  *buffer;
 760
 761         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 762
 763         rpc->srpc_status = status;
 764
 765         CDEBUG (status == 0 ? D_NET : D_NETERROR,
 766                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 767                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 768                 swi_state2str(rpc->srpc_wi.swi_state), status);
 769
 770         if (status != 0) {
 771                 cfs_spin_lock(&srpc_data.rpc_glock);
 772                 srpc_data.rpc_counters.rpcs_dropped++;
 773                 cfs_spin_unlock(&srpc_data.rpc_glock);
 774         }
 775
 776         if (rpc->srpc_done != NULL)
 777                 (*rpc->srpc_done) (rpc);
 778         LASSERT (rpc->srpc_bulk == NULL);
 779
 780         cfs_spin_lock(&sv->sv_lock);
 781
 782         if (rpc->srpc_reqstbuf != NULL) {
 783                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 784                  * sv won't go away for sv_active_rpcq must not be empty */
 785                 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
 786                 rpc->srpc_reqstbuf = NULL;
 787         }
 788
 789         cfs_list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
 790
 791         /*
 792          * No one can schedule me now since:
 793          * - I'm not on sv_active_rpcq.
 794          * - all LNet events have been fired.
 795          * Cancel pending schedules and prevent future schedule attempts:
 796          */
 797         LASSERT (rpc->srpc_ev.ev_fired);
 798         swi_kill_workitem(&rpc->srpc_wi);
 799
 800         if (!sv->sv_shuttingdown && !cfs_list_empty(&sv->sv_blocked_msgq)) {
 801                 buffer = cfs_list_entry(sv->sv_blocked_msgq.next,
 802                                     srpc_buffer_t, buf_list);
 803                 cfs_list_del(&buffer->buf_list);
 804
 805                 srpc_init_server_rpc(rpc, sv, buffer);
 806                 cfs_list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
 807                 srpc_schedule_server_rpc(rpc);
 808         } else {
 809                 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 810         }
 811
 812         cfs_spin_unlock(&sv->sv_lock);
 813         return;
 814 }
 815
 816 /* handles an incoming RPC */
 817 int
 818 srpc_handle_rpc (swi_workitem_t *wi)
 819 {
 820         srpc_server_rpc_t *rpc = wi->swi_workitem.wi_data;
 821         srpc_service_t    *sv = rpc->srpc_service;
 822         srpc_event_t      *ev = &rpc->srpc_ev;
 823         int                rc = 0;
 824
 825         LASSERT (wi == &rpc->srpc_wi);
 826
 827         cfs_spin_lock(&sv->sv_lock);
 828
 829         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 830                 cfs_spin_unlock(&sv->sv_lock);
 831
 832                 if (rpc->srpc_bulk != NULL)
 833                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 834                 LNetMDUnlink(rpc->srpc_replymdh);
 835
 836                 if (ev->ev_fired) { /* no more event, OK to finish */
 837                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 838                         return 1;
 839                 }
 840                 return 0;
 841         }
 842
 843         cfs_spin_unlock(&sv->sv_lock);
 844
 845         switch (wi->swi_state) {
 846         default:
 847                 LBUG ();
 848         case SWI_STATE_NEWBORN: {
 849                 srpc_msg_t           *msg;
 850                 srpc_generic_reply_t *reply;
 851
 852                 msg = &rpc->srpc_reqstbuf->buf_msg;
 853                 reply = &rpc->srpc_replymsg.msg_body.reply;
 854
 855                 if (msg->msg_magic == 0) {
 856                         /* moaned already in srpc_lnet_ev_handler */
 857                         rc = EBADMSG;
 858                 } else if (msg->msg_version != SRPC_MSG_VERSION &&
 859                            msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
 860                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
 861                                msg->msg_version, SRPC_MSG_VERSION,
 862                                libcfs_id2str(rpc->srpc_peer));
 863                         reply->status = EPROTO;
 864                 } else {
 865                         reply->status = 0;
 866                         rc = (*sv->sv_handler) (rpc);
 867                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
 868                 }
 869
 870                 if (rc != 0) {
 871                         srpc_server_rpc_done(rpc, rc);
 872                         return 1;
 873                 }
 874
 875                 wi->swi_state = SWI_STATE_BULK_STARTED;
 876
 877                 if (rpc->srpc_bulk != NULL) {
 878                         rc = srpc_do_bulk(rpc);
 879                         if (rc == 0)
 880                                 return 0; /* wait for bulk */
 881
 882                         LASSERT (ev->ev_fired);
 883                         ev->ev_status = rc;
 884                 }
 885         }
 886         case SWI_STATE_BULK_STARTED:
 887                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
 888
 889                 if (rpc->srpc_bulk != NULL) {
 890                         rc = ev->ev_status;
 891
 892                         if (sv->sv_bulk_ready != NULL)
 893                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
 894
 895                         if (rc != 0) {
 896                                 srpc_server_rpc_done(rpc, rc);
 897                                 return 1;
 898                         }
 899                 }
 900
 901                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
 902                 rc = srpc_send_reply(rpc);
 903                 if (rc == 0)
 904                         return 0; /* wait for reply */
 905                 srpc_server_rpc_done(rpc, rc);
 906                 return 1;
 907
 908         case SWI_STATE_REPLY_SUBMITTED:
 909                 if (!ev->ev_fired) {
 910                         CERROR("RPC %p: bulk %p, service %d\n",
 911                                rpc, rpc->srpc_bulk, rpc->srpc_service->sv_id);
 912                         CERROR("Event: status %d, type %d, lnet %d\n",
 913                                ev->ev_status, ev->ev_type, ev->ev_lnet);
 914                         LASSERT (ev->ev_fired);
 915                 }
 916
 917                 wi->swi_state = SWI_STATE_DONE;
 918                 srpc_server_rpc_done(rpc, ev->ev_status);
 919                 return 1;
 920         }
 921
 922         return 0;
 923 }
 924
 925 void
 926 srpc_client_rpc_expired (void *data)
 927 {
 928         srpc_client_rpc_t *rpc = data;
 929
 930         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
 931                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
 932                rpc->crpc_timeout);
 933
 934         cfs_spin_lock(&rpc->crpc_lock);
 935
 936         rpc->crpc_timeout = 0;
 937         srpc_abort_rpc(rpc, -ETIMEDOUT);
 938
 939         cfs_spin_unlock(&rpc->crpc_lock);
 940
 941         cfs_spin_lock(&srpc_data.rpc_glock);
 942         srpc_data.rpc_counters.rpcs_expired++;
 943         cfs_spin_unlock(&srpc_data.rpc_glock);
 944         return;
 945 }
 946
 947 inline void
 948 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
 949 {
 950         stt_timer_t *timer = &rpc->crpc_timer;
 951
 952         if (rpc->crpc_timeout == 0) return;
 953
 954         CFS_INIT_LIST_HEAD(&timer->stt_list);
 955         timer->stt_data    = rpc;
 956         timer->stt_func    = srpc_client_rpc_expired;
 957         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
 958                                           cfs_time_current_sec());
 959         stt_add_timer(timer);
 960         return;
 961 }
 962
 963 /*
 964  * Called with rpc->crpc_lock held.
 965  *
 966  * Upon exit the RPC expiry timer is not queued and the handler is not
 967  * running on any CPU. */
 968 void
 969 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
 970 {
 971         /* timer not planted or already exploded */
 972         if (rpc->crpc_timeout == 0) return;
 973
 974         /* timer sucessfully defused */
 975         if (stt_del_timer(&rpc->crpc_timer)) return;
 976
 977 #ifdef __KERNEL__
 978         /* timer detonated, wait for it to explode */
 979         while (rpc->crpc_timeout != 0) {
 980                 cfs_spin_unlock(&rpc->crpc_lock);
 981
 982                 cfs_schedule();
 983
 984                 cfs_spin_lock(&rpc->crpc_lock);
 985         }
 986 #else
 987         LBUG(); /* impossible in single-threaded runtime */
 988 #endif
 989         return;
 990 }
 991
 992 void
 993 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
 994 {
 995         swi_workitem_t *wi = &rpc->crpc_wi;
 996
 997         LASSERT (status != 0 || wi->swi_state == SWI_STATE_DONE);
 998
 999         cfs_spin_lock(&rpc->crpc_lock);
1000
1001         rpc->crpc_closed = 1;
1002         if (rpc->crpc_status == 0)
1003                 rpc->crpc_status = status;
1004
1005         srpc_del_client_rpc_timer(rpc);
1006
1007         CDEBUG ((status == 0) ? D_NET : D_NETERROR,
1008                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1009                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1010                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1011
1012         /*
1013          * No one can schedule me now since:
1014          * - RPC timer has been defused.
1015          * - all LNet events have been fired.
1016          * - crpc_closed has been set, preventing srpc_abort_rpc from
1017          *   scheduling me.
1018          * Cancel pending schedules and prevent future schedule attempts:
1019          */
1020         LASSERT (!srpc_event_pending(rpc));
1021         swi_kill_workitem(wi);
1022
1023         cfs_spin_unlock(&rpc->crpc_lock);
1024
1025         (*rpc->crpc_done) (rpc);
1026         return;
1027 }
1028
1029 /* sends an outgoing RPC */
1030 int
1031 srpc_send_rpc (swi_workitem_t *wi)
1032 {
1033         int                rc = 0;
1034         srpc_client_rpc_t *rpc = wi->swi_workitem.wi_data;
1035         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1036         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1037
1038         LASSERT (rpc != NULL);
1039         LASSERT (wi == &rpc->crpc_wi);
1040
1041         cfs_spin_lock(&rpc->crpc_lock);
1042
1043         if (rpc->crpc_aborted) {
1044                 cfs_spin_unlock(&rpc->crpc_lock);
1045                 goto abort;
1046         }
1047
1048         cfs_spin_unlock(&rpc->crpc_lock);
1049
1050         switch (wi->swi_state) {
1051         default:
1052                 LBUG ();
1053         case SWI_STATE_NEWBORN:
1054                 LASSERT (!srpc_event_pending(rpc));
1055
1056                 rc = srpc_prepare_reply(rpc);
1057                 if (rc != 0) {
1058                         srpc_client_rpc_done(rpc, rc);
1059                         return 1;
1060                 }
1061
1062                 rc = srpc_prepare_bulk(rpc);
1063                 if (rc != 0) break;
1064
1065                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1066                 rc = srpc_send_request(rpc);
1067                 break;
1068
1069         case SWI_STATE_REQUEST_SUBMITTED:
1070                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1071                  * order; however, they're processed in a strict order:
1072                  * rqt, rpy, and bulk. */
1073                 if (!rpc->crpc_reqstev.ev_fired) break;
1074
1075                 rc = rpc->crpc_reqstev.ev_status;
1076                 if (rc != 0) break;
1077
1078                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1079                 /* perhaps more events, fall thru */
1080         case SWI_STATE_REQUEST_SENT: {
1081                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1082
1083                 if (!rpc->crpc_replyev.ev_fired) break;
1084
1085                 rc = rpc->crpc_replyev.ev_status;
1086                 if (rc != 0) break;
1087
1088                 if ((reply->msg_type != type &&
1089                      reply->msg_type != __swab32(type)) ||
1090                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1091                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1092                         CWARN ("Bad message from %s: type %u (%d expected),"
1093                                " magic %u (%d expected).\n",
1094                                libcfs_id2str(rpc->crpc_dest),
1095                                reply->msg_type, type,
1096                                reply->msg_magic, SRPC_MSG_MAGIC);
1097                         rc = -EBADMSG;
1098                         break;
1099                 }
1100
1101                 if (do_bulk && reply->msg_body.reply.status != 0) {
1102                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1103                                "case peer didn't initiate bulk transfer\n",
1104                                reply->msg_body.reply.status,
1105                                libcfs_id2str(rpc->crpc_dest));
1106                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1107                 }
1108
1109                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1110         }
1111         case SWI_STATE_REPLY_RECEIVED:
1112                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1113
1114                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1115
1116                 /* Bulk buffer was unlinked due to remote error. Clear error
1117                  * since reply buffer still contains valid data.
1118                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1119                  * remote error. */
1120                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1121                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1122                         rc = 0;
1123
1124                 wi->swi_state = SWI_STATE_DONE;
1125                 srpc_client_rpc_done(rpc, rc);
1126                 return 1;
1127         }
1128
1129         if (rc != 0) {
1130                 cfs_spin_lock(&rpc->crpc_lock);
1131                 srpc_abort_rpc(rpc, rc);
1132                 cfs_spin_unlock(&rpc->crpc_lock);
1133         }
1134
1135 abort:
1136         if (rpc->crpc_aborted) {
1137                 LNetMDUnlink(rpc->crpc_reqstmdh);
1138                 LNetMDUnlink(rpc->crpc_replymdh);
1139                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1140
1141                 if (!srpc_event_pending(rpc)) {
1142                         srpc_client_rpc_done(rpc, -EINTR);
1143                         return 1;
1144                 }
1145         }
1146         return 0;
1147 }
1148
1149 srpc_client_rpc_t *
1150 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1151                         int nbulkiov, int bulklen,
1152                         void (*rpc_done)(srpc_client_rpc_t *),
1153                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1154 {
1155         srpc_client_rpc_t *rpc;
1156
1157         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1158                                    crpc_bulk.bk_iovs[nbulkiov]));
1159         if (rpc == NULL)
1160                 return NULL;
1161
1162         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1163                              bulklen, rpc_done, rpc_fini, priv);
1164         return rpc;
1165 }
1166
1167 /* called with rpc->crpc_lock held */
1168 void
1169 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1170 {
1171         LASSERT (why != 0);
1172
1173         if (rpc->crpc_aborted || /* already aborted */
1174             rpc->crpc_closed)    /* callback imminent */
1175                 return;
1176
1177         CDEBUG (D_NET,
1178                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1179                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1180                 swi_state2str(rpc->crpc_wi.swi_state), why);
1181
1182         rpc->crpc_aborted = 1;
1183         rpc->crpc_status  = why;
1184         swi_schedule_workitem(&rpc->crpc_wi);
1185         return;
1186 }
1187
1188 /* called with rpc->crpc_lock held */
1189 void
1190 srpc_post_rpc (srpc_client_rpc_t *rpc)
1191 {
1192         LASSERT (!rpc->crpc_aborted);
1193         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1194         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1195
1196         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1197                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1198                 rpc->crpc_timeout);
1199
1200         srpc_add_client_rpc_timer(rpc);
1201         swi_schedule_workitem(&rpc->crpc_wi);
1202         return;
1203 }
1204
1205
1206 int
1207 srpc_send_reply (srpc_server_rpc_t *rpc)
1208 {
1209         srpc_event_t   *ev = &rpc->srpc_ev;
1210         srpc_msg_t     *msg = &rpc->srpc_replymsg;
1211         srpc_buffer_t  *buffer = rpc->srpc_reqstbuf;
1212         srpc_service_t *sv = rpc->srpc_service;
1213         __u64           rpyid;
1214         int             rc;
1215
1216         LASSERT (buffer != NULL);
1217         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1218
1219         cfs_spin_lock(&sv->sv_lock);
1220
1221         if (!sv->sv_shuttingdown &&
1222             sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1223                 /* Repost buffer before replying since test client
1224                  * might send me another RPC once it gets the reply */
1225                 if (srpc_service_post_buffer(sv, buffer) != 0)
1226                         CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1227                 rpc->srpc_reqstbuf = NULL;
1228         }
1229
1230         cfs_spin_unlock(&sv->sv_lock);
1231
1232         ev->ev_fired = 0;
1233         ev->ev_data  = rpc;
1234         ev->ev_type  = SRPC_REPLY_SENT;
1235
1236         msg->msg_magic   = SRPC_MSG_MAGIC;
1237         msg->msg_version = SRPC_MSG_VERSION;
1238         msg->msg_type    = srpc_service2reply(sv->sv_id);
1239
1240         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1241                                    sizeof(*msg), LNET_MD_OP_PUT,
1242                                    rpc->srpc_peer, rpc->srpc_self,
1243                                    &rpc->srpc_replymdh, ev);
1244         if (rc != 0)
1245                 ev->ev_fired = 1;  /* no more event expected */
1246         return rc;
1247 }
1248
1249 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1250 void
1251 srpc_lnet_ev_handler (lnet_event_t *ev)
1252 {
1253         srpc_event_t      *rpcev = ev->md.user_ptr;
1254         srpc_client_rpc_t *crpc;
1255         srpc_server_rpc_t *srpc;
1256         srpc_buffer_t     *buffer;
1257         srpc_service_t    *sv;
1258         srpc_msg_t        *msg;
1259         srpc_msg_type_t    type;
1260
1261         LASSERT (!cfs_in_interrupt());
1262
1263         if (ev->status != 0) {
1264                 cfs_spin_lock(&srpc_data.rpc_glock);
1265                 srpc_data.rpc_counters.errors++;
1266                 cfs_spin_unlock(&srpc_data.rpc_glock);
1267         }
1268
1269         rpcev->ev_lnet = ev->type;
1270
1271         switch (rpcev->ev_type) {
1272         default:
1273                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1274                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1275                 LBUG ();
1276         case SRPC_REQUEST_SENT:
1277                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1278                         cfs_spin_lock(&srpc_data.rpc_glock);
1279                         srpc_data.rpc_counters.rpcs_sent++;
1280                         cfs_spin_unlock(&srpc_data.rpc_glock);
1281                 }
1282         case SRPC_REPLY_RCVD:
1283         case SRPC_BULK_REQ_RCVD:
1284                 crpc = rpcev->ev_data;
1285
1286                 if (rpcev != &crpc->crpc_reqstev &&
1287                     rpcev != &crpc->crpc_replyev &&
1288                     rpcev != &crpc->crpc_bulkev) {
1289                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1290                                rpcev, crpc, &crpc->crpc_reqstev,
1291                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1292                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1293                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1294                         LBUG ();
1295                 }
1296
1297                 cfs_spin_lock(&crpc->crpc_lock);
1298
1299                 LASSERT (rpcev->ev_fired == 0);
1300                 rpcev->ev_fired  = 1;
1301                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1302                                                 -EINTR : ev->status;
1303                 swi_schedule_workitem(&crpc->crpc_wi);
1304
1305                 cfs_spin_unlock(&crpc->crpc_lock);
1306                 break;
1307
1308         case SRPC_REQUEST_RCVD:
1309                 sv = rpcev->ev_data;
1310
1311                 LASSERT (rpcev == &sv->sv_ev);
1312
1313                 cfs_spin_lock(&sv->sv_lock);
1314
1315                 LASSERT (ev->unlinked);
1316                 LASSERT (ev->type == LNET_EVENT_PUT ||
1317                          ev->type == LNET_EVENT_UNLINK);
1318                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1319                          sv->sv_shuttingdown);
1320
1321                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1322                 buffer->buf_peer = ev->initiator;
1323                 buffer->buf_self = ev->target.nid;
1324
1325                 sv->sv_nposted_msg--;
1326                 LASSERT (sv->sv_nposted_msg >= 0);
1327
1328                 if (sv->sv_shuttingdown) {
1329                         /* Leave buffer on sv->sv_posted_msgq since
1330                          * srpc_finish_service needs to traverse it. */
1331                         cfs_spin_unlock(&sv->sv_lock);
1332                         break;
1333                 }
1334
1335                 cfs_list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1336                 msg = &buffer->buf_msg;
1337                 type = srpc_service2request(sv->sv_id);
1338
1339                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1340                     (msg->msg_type != type &&
1341                      msg->msg_type != __swab32(type)) ||
1342                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1343                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1344                         CERROR ("Dropping RPC (%s) from %s: "
1345                                 "status %d mlength %d type %u magic %u.\n",
1346                                 sv->sv_name, libcfs_id2str(ev->initiator),
1347                                 ev->status, ev->mlength,
1348                                 msg->msg_type, msg->msg_magic);
1349
1350                         /* NB can't call srpc_service_recycle_buffer here since
1351                          * it may call LNetM[DE]Attach. The invalid magic tells
1352                          * srpc_handle_rpc to drop this RPC */
1353                         msg->msg_magic = 0;
1354                 }
1355
1356                 if (!cfs_list_empty(&sv->sv_free_rpcq)) {
1357                         srpc = cfs_list_entry(sv->sv_free_rpcq.next,
1358                                               srpc_server_rpc_t, srpc_list);
1359                         cfs_list_del(&srpc->srpc_list);
1360
1361                         srpc_init_server_rpc(srpc, sv, buffer);
1362                         cfs_list_add_tail(&srpc->srpc_list,
1363                                           &sv->sv_active_rpcq);
1364                         srpc_schedule_server_rpc(srpc);
1365                 } else {
1366                         cfs_list_add_tail(&buffer->buf_list,
1367                                           &sv->sv_blocked_msgq);
1368                 }
1369
1370                 cfs_spin_unlock(&sv->sv_lock);
1371
1372                 cfs_spin_lock(&srpc_data.rpc_glock);
1373                 srpc_data.rpc_counters.rpcs_rcvd++;
1374                 cfs_spin_unlock(&srpc_data.rpc_glock);
1375                 break;
1376
1377         case SRPC_BULK_GET_RPLD:
1378                 LASSERT (ev->type == LNET_EVENT_SEND ||
1379                          ev->type == LNET_EVENT_REPLY ||
1380                          ev->type == LNET_EVENT_UNLINK);
1381
1382                 if (!ev->unlinked)
1383                         break; /* wait for final event */
1384
1385         case SRPC_BULK_PUT_SENT:
1386                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1387                         cfs_spin_lock(&srpc_data.rpc_glock);
1388
1389                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1390                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1391                         else
1392                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1393
1394                         cfs_spin_unlock(&srpc_data.rpc_glock);
1395                 }
1396         case SRPC_REPLY_SENT:
1397                 srpc = rpcev->ev_data;
1398                 sv = srpc->srpc_service;
1399
1400                 LASSERT (rpcev == &srpc->srpc_ev);
1401
1402                 cfs_spin_lock(&sv->sv_lock);
1403
1404                 rpcev->ev_fired  = 1;
1405                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1406                                                 -EINTR : ev->status;
1407                 srpc_schedule_server_rpc(srpc);
1408
1409                 cfs_spin_unlock(&sv->sv_lock);
1410                 break;
1411         }
1412
1413         return;
1414 }
1415
1416 #ifndef __KERNEL__
1417
1418 int
1419 srpc_check_event (int timeout)
1420 {
1421         lnet_event_t ev;
1422         int          rc;
1423         int          i;
1424
1425         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1426                         timeout * 1000, &ev, &i);
1427         if (rc == 0) return 0;
1428
1429         LASSERT (rc == -EOVERFLOW || rc == 1);
1430
1431         /* We can't affort to miss any events... */
1432         if (rc == -EOVERFLOW) {
1433                 CERROR ("Dropped an event!!!\n");
1434                 abort();
1435         }
1436
1437         srpc_lnet_ev_handler(&ev);
1438         return 1;
1439 }
1440
1441 #endif
1442
1443 int
1444 srpc_startup (void)
1445 {
1446         int rc;
1447
1448         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1449         cfs_spin_lock_init(&srpc_data.rpc_glock);
1450
1451         /* 1 second pause to avoid timestamp reuse */
1452         cfs_pause(cfs_time_seconds(1));
1453         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1454
1455         srpc_data.rpc_state = SRPC_STATE_NONE;
1456
1457 #ifdef __KERNEL__
1458         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1459 #else
1460         if (the_lnet.ln_server_mode_flag)
1461                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1462         else
1463                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1464 #endif
1465         if (rc < 0) {
1466                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1467                 return rc;
1468         }
1469
1470         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1471
1472         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1473 #ifdef __KERNEL__
1474         rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1475 #else
1476         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1477 #endif
1478         if (rc != 0) {
1479                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1480                 goto bail;
1481         }
1482
1483         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1484         LASSERT (rc == 0);
1485
1486         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1487
1488         rc = stt_startup();
1489
1490 bail:
1491         if (rc != 0)
1492                 srpc_shutdown();
1493         else
1494                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1495
1496         return rc;
1497 }
1498
1499 void
1500 srpc_shutdown (void)
1501 {
1502         int i;
1503         int rc;
1504         int state;
1505
1506         state = srpc_data.rpc_state;
1507         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1508
1509         switch (state) {
1510         default:
1511                 LBUG ();
1512         case SRPC_STATE_RUNNING:
1513                 cfs_spin_lock(&srpc_data.rpc_glock);
1514
1515                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1516                         srpc_service_t *sv = srpc_data.rpc_services[i];
1517
1518                         LASSERTF (sv == NULL,
1519                                   "service not empty: id %d, name %s\n",
1520                                   i, sv->sv_name);
1521                 }
1522
1523                 cfs_spin_unlock(&srpc_data.rpc_glock);
1524
1525                 stt_shutdown();
1526
1527         case SRPC_STATE_EQ_INIT:
1528                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1529                 LASSERT (rc == 0);
1530                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1531                 LASSERT (rc == 0); /* the EQ should have no user by now */
1532
1533         case SRPC_STATE_NI_INIT:
1534                 LNetNIFini();
1535                 break;
1536         }
1537
1538         return;
1539 }