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  2008 Sun Microsystems, Inc. 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_WI_INIT,
  51         SRPC_STATE_RUNNING,
  52         SRPC_STATE_STOPPING,
  53 } srpc_state_t;
  54
  55 struct smoketest_rpc {
  56         cfs_spinlock_t    rpc_glock;     /* global lock */
  57         srpc_service_t   *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  58         lnet_handle_eq_t  rpc_lnet_eq;   /* _the_ LNet event queue */
  59         srpc_state_t      rpc_state;
  60         srpc_counters_t   rpc_counters;
  61         __u64             rpc_matchbits; /* matchbits counter */
  62 } srpc_data;
  63
  64 /* forward ref's */
  65 int srpc_handle_rpc (swi_workitem_t *wi);
  66
  67 void srpc_get_counters (srpc_counters_t *cnt)
  68 {
  69         cfs_spin_lock(&srpc_data.rpc_glock);
  70         *cnt = srpc_data.rpc_counters;
  71         cfs_spin_unlock(&srpc_data.rpc_glock);
  72 }
  73
  74 void srpc_set_counters (const srpc_counters_t *cnt)
  75 {
  76         cfs_spin_lock(&srpc_data.rpc_glock);
  77         srpc_data.rpc_counters = *cnt;
  78         cfs_spin_unlock(&srpc_data.rpc_glock);
  79 }
  80
  81 void
  82 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
  83 {
  84         LASSERT (i >= 0 && i < bk->bk_niov);
  85
  86 #ifdef __KERNEL__
  87         bk->bk_iovs[i].kiov_offset = 0;
  88         bk->bk_iovs[i].kiov_page   = pg;
  89         bk->bk_iovs[i].kiov_len    = CFS_PAGE_SIZE;
  90 #else
  91         LASSERT (bk->bk_pages != NULL);
  92
  93         bk->bk_pages[i] = pg;
  94         bk->bk_iovs[i].iov_len  = CFS_PAGE_SIZE;
  95         bk->bk_iovs[i].iov_base = cfs_page_address(pg);
  96 #endif
  97         return;
  98 }
  99
 100 void
 101 srpc_free_bulk (srpc_bulk_t *bk)
 102 {
 103         int         i;
 104         cfs_page_t *pg;
 105
 106         LASSERT (bk != NULL);
 107 #ifndef __KERNEL__
 108         LASSERT (bk->bk_pages != NULL);
 109 #endif
 110
 111         for (i = 0; i < bk->bk_niov; i++) {
 112 #ifdef __KERNEL__
 113                 pg = bk->bk_iovs[i].kiov_page;
 114 #else
 115                 pg = bk->bk_pages[i];
 116 #endif
 117                 if (pg == NULL) break;
 118
 119                 cfs_free_page(pg);
 120         }
 121
 122 #ifndef __KERNEL__
 123         LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
 124 #endif
 125         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 126         return;
 127 }
 128
 129 srpc_bulk_t *
 130 srpc_alloc_bulk (int npages, int sink)
 131 {
 132         srpc_bulk_t  *bk;
 133         cfs_page_t  **pages;
 134         int           i;
 135
 136         LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
 137
 138         LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 139         if (bk == NULL) {
 140                 CERROR ("Can't allocate descriptor for %d pages\n", npages);
 141                 return NULL;
 142         }
 143
 144         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
 145         bk->bk_sink = sink;
 146         bk->bk_niov = npages;
 147         bk->bk_len  = npages * CFS_PAGE_SIZE;
 148 #ifndef __KERNEL__
 149         LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
 150         if (pages == NULL) {
 151                 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 152                 CERROR ("Can't allocate page array for %d pages\n", npages);
 153                 return NULL;
 154         }
 155
 156         memset(pages, 0, sizeof(cfs_page_t *) * npages);
 157         bk->bk_pages = pages;
 158 #else
 159         UNUSED (pages);
 160 #endif
 161
 162         for (i = 0; i < npages; i++) {
 163                 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
 164
 165                 if (pg == NULL) {
 166                         CERROR ("Can't allocate page %d of %d\n", i, npages);
 167                         srpc_free_bulk(bk);
 168                         return NULL;
 169                 }
 170
 171                 srpc_add_bulk_page(bk, pg, i);
 172         }
 173
 174         return bk;
 175 }
 176
 177 static inline __u64
 178 srpc_next_id (void)
 179 {
 180         __u64 id;
 181
 182         cfs_spin_lock(&srpc_data.rpc_glock);
 183         id = srpc_data.rpc_matchbits++;
 184         cfs_spin_unlock(&srpc_data.rpc_glock);
 185         return id;
 186 }
 187
 188 void
 189 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
 190                       srpc_service_t *sv, srpc_buffer_t *buffer)
 191 {
 192         memset(rpc, 0, sizeof(*rpc));
 193         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc);
 194
 195         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 196
 197         rpc->srpc_service  = sv;
 198         rpc->srpc_reqstbuf = buffer;
 199         rpc->srpc_peer     = buffer->buf_peer;
 200         rpc->srpc_self     = buffer->buf_self;
 201         LNetInvalidateHandle(&rpc->srpc_replymdh);
 202 }
 203
 204 int
 205 srpc_add_service (srpc_service_t *sv)
 206 {
 207         int                id = sv->sv_id;
 208         int                i;
 209         srpc_server_rpc_t *rpc;
 210
 211         LASSERT (sv->sv_concur > 0);
 212         LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
 213
 214         cfs_spin_lock(&srpc_data.rpc_glock);
 215
 216         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
 217
 218         if (srpc_data.rpc_services[id] != NULL) {
 219                 cfs_spin_unlock(&srpc_data.rpc_glock);
 220                 return -EBUSY;
 221         }
 222
 223         srpc_data.rpc_services[id] = sv;
 224         cfs_spin_unlock(&srpc_data.rpc_glock);
 225
 226         sv->sv_nprune       = 0;
 227         sv->sv_nposted_msg  = 0;
 228         sv->sv_shuttingdown = 0;
 229         cfs_spin_lock_init(&sv->sv_lock);
 230         CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
 231         CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
 232         CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
 233         CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
 234
 235         sv->sv_ev.ev_data = sv;
 236         sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
 237
 238         for (i = 0; i < sv->sv_concur; i++) {
 239                 LIBCFS_ALLOC(rpc, sizeof(*rpc));
 240                 if (rpc == NULL) goto enomem;
 241
 242                 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 243         }
 244
 245         CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
 246                 id, sv->sv_name, sv->sv_concur);
 247         return 0;
 248
 249 enomem:
 250         while (!cfs_list_empty(&sv->sv_free_rpcq)) {
 251                 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
 252                                      srpc_server_rpc_t, srpc_list);
 253                 cfs_list_del(&rpc->srpc_list);
 254                 LIBCFS_FREE(rpc, sizeof(*rpc));
 255         }
 256
 257         cfs_spin_lock(&srpc_data.rpc_glock);
 258         srpc_data.rpc_services[id] = NULL;
 259         cfs_spin_unlock(&srpc_data.rpc_glock);
 260         return -ENOMEM;
 261 }
 262
 263 int
 264 srpc_remove_service (srpc_service_t *sv)
 265 {
 266         int id = sv->sv_id;
 267
 268         cfs_spin_lock(&srpc_data.rpc_glock);
 269
 270         if (srpc_data.rpc_services[id] != sv) {
 271                 cfs_spin_unlock(&srpc_data.rpc_glock);
 272                 return -ENOENT;
 273         }
 274
 275         srpc_data.rpc_services[id] = NULL;
 276         cfs_spin_unlock(&srpc_data.rpc_glock);
 277         return 0;
 278 }
 279
 280 int
 281 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
 282                        int len, int options, lnet_process_id_t peer,
 283                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 284 {
 285         int              rc;
 286         lnet_md_t        md;
 287         lnet_handle_me_t meh;
 288
 289         rc = LNetMEAttach(portal, peer, matchbits, 0,
 290                           LNET_UNLINK, LNET_INS_AFTER, &meh);
 291         if (rc != 0) {
 292                 CERROR ("LNetMEAttach failed: %d\n", rc);
 293                 LASSERT (rc == -ENOMEM);
 294                 return -ENOMEM;
 295         }
 296
 297         md.threshold = 1;
 298         md.user_ptr  = ev;
 299         md.start     = buf;
 300         md.length    = len;
 301         md.options   = options;
 302         md.eq_handle = srpc_data.rpc_lnet_eq;
 303
 304         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 305         if (rc != 0) {
 306                 CERROR ("LNetMDAttach failed: %d\n", rc);
 307                 LASSERT (rc == -ENOMEM);
 308
 309                 rc = LNetMEUnlink(meh);
 310                 LASSERT (rc == 0);
 311                 return -ENOMEM;
 312         }
 313
 314         CDEBUG (D_NET,
 315                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 316                 libcfs_id2str(peer), portal, matchbits);
 317         return 0;
 318 }
 319
 320 int
 321 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 322                       int options, lnet_process_id_t peer, lnet_nid_t self,
 323                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 324 {
 325         int       rc;
 326         lnet_md_t md;
 327
 328         md.user_ptr  = ev;
 329         md.start     = buf;
 330         md.length    = len;
 331         md.eq_handle = srpc_data.rpc_lnet_eq;
 332         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 333         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 334
 335         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 336         if (rc != 0) {
 337                 CERROR ("LNetMDBind failed: %d\n", rc);
 338                 LASSERT (rc == -ENOMEM);
 339                 return -ENOMEM;
 340         }
 341
 342         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 343          * they're only meaningful for MDs attached to an ME (i.e. passive
 344          * buffers... */
 345         if ((options & LNET_MD_OP_PUT) != 0) {
 346                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 347                              portal, matchbits, 0, 0);
 348         } else {
 349                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 350
 351                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 352         }
 353
 354         if (rc != 0) {
 355                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 356                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 357                         libcfs_id2str(peer), portal, matchbits, rc);
 358
 359                 /* The forthcoming unlink event will complete this operation
 360                  * with failure, so fall through and return success here.
 361                  */
 362                 rc = LNetMDUnlink(*mdh);
 363                 LASSERT (rc == 0);
 364         } else {
 365                 CDEBUG (D_NET,
 366                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 367                         libcfs_id2str(peer), portal, matchbits);
 368         }
 369         return 0;
 370 }
 371
 372 int
 373 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 374                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 375 {
 376         int rc;
 377         int portal;
 378
 379         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 380                 portal = SRPC_REQUEST_PORTAL;
 381         else
 382                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 383
 384         rc = srpc_post_active_rdma(portal, service, buf, len,
 385                                    LNET_MD_OP_PUT, peer,
 386                                    LNET_NID_ANY, mdh, ev);
 387         return rc;
 388 }
 389
 390 int
 391 srpc_post_passive_rqtbuf(int service, void *buf, int len,
 392                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 393 {
 394         int               rc;
 395         int               portal;
 396         lnet_process_id_t any = {0};
 397
 398         any.nid = LNET_NID_ANY;
 399         any.pid = LNET_PID_ANY;
 400
 401         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 402                 portal = SRPC_REQUEST_PORTAL;
 403         else
 404                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 405
 406         rc = srpc_post_passive_rdma(portal, service, buf, len,
 407                                     LNET_MD_OP_PUT, any, mdh, ev);
 408         return rc;
 409 }
 410
 411 int
 412 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 413 {
 414         srpc_msg_t *msg = &buf->buf_msg;
 415         int         rc;
 416
 417         LASSERT (!sv->sv_shuttingdown);
 418
 419         LNetInvalidateHandle(&buf->buf_mdh);
 420         cfs_list_add(&buf->buf_list, &sv->sv_posted_msgq);
 421         sv->sv_nposted_msg++;
 422         cfs_spin_unlock(&sv->sv_lock);
 423
 424         rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
 425                                       &buf->buf_mdh, &sv->sv_ev);
 426
 427         /* At this point, a RPC (new or delayed) may have arrived in
 428          * msg and its event handler has been called. So we must add
 429          * buf to sv_posted_msgq _before_ dropping sv_lock */
 430
 431         cfs_spin_lock(&sv->sv_lock);
 432
 433         if (rc == 0) {
 434                 if (sv->sv_shuttingdown) {
 435                         cfs_spin_unlock(&sv->sv_lock);
 436
 437                         /* srpc_shutdown_service might have tried to unlink me
 438                          * when my buf_mdh was still invalid */
 439                         LNetMDUnlink(buf->buf_mdh);
 440
 441                         cfs_spin_lock(&sv->sv_lock);
 442                 }
 443                 return 0;
 444         }
 445
 446         sv->sv_nposted_msg--;
 447         if (sv->sv_shuttingdown) return rc;
 448
 449         cfs_list_del(&buf->buf_list);
 450
 451         cfs_spin_unlock(&sv->sv_lock);
 452         LIBCFS_FREE(buf, sizeof(*buf));
 453         cfs_spin_lock(&sv->sv_lock);
 454         return rc;
 455 }
 456
 457 int
 458 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
 459 {
 460         int                rc;
 461         int                posted;
 462         srpc_buffer_t     *buf;
 463
 464         LASSERTF (nbuffer > 0,
 465                   "nbuffer must be positive: %d\n", nbuffer);
 466
 467         for (posted = 0; posted < nbuffer; posted++) {
 468                 LIBCFS_ALLOC(buf, sizeof(*buf));
 469                 if (buf == NULL) break;
 470
 471                 cfs_spin_lock(&sv->sv_lock);
 472                 rc = srpc_service_post_buffer(sv, buf);
 473                 cfs_spin_unlock(&sv->sv_lock);
 474
 475                 if (rc != 0) break;
 476         }
 477
 478         return posted;
 479 }
 480
 481 void
 482 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
 483 {
 484         LASSERTF (nbuffer > 0,
 485                   "nbuffer must be positive: %d\n", nbuffer);
 486
 487         cfs_spin_lock(&sv->sv_lock);
 488
 489         LASSERT (sv->sv_nprune >= 0);
 490         LASSERT (!sv->sv_shuttingdown);
 491
 492         sv->sv_nprune += nbuffer;
 493
 494         cfs_spin_unlock(&sv->sv_lock);
 495         return;
 496 }
 497
 498 /* returns 1 if sv has finished, otherwise 0 */
 499 int
 500 srpc_finish_service (srpc_service_t *sv)
 501 {
 502         srpc_server_rpc_t *rpc;
 503         srpc_buffer_t     *buf;
 504
 505         cfs_spin_lock(&sv->sv_lock);
 506
 507         LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
 508
 509         if (sv->sv_nposted_msg != 0 || !cfs_list_empty(&sv->sv_active_rpcq)) {
 510                 CDEBUG (D_NET,
 511                         "waiting for %d posted buffers to unlink and "
 512                         "in-flight RPCs to die.\n",
 513                         sv->sv_nposted_msg);
 514
 515                 if (!cfs_list_empty(&sv->sv_active_rpcq)) {
 516                         rpc = cfs_list_entry(sv->sv_active_rpcq.next,
 517                                              srpc_server_rpc_t, srpc_list);
 518                         CDEBUG (D_NETERROR,
 519                                 "Active RPC %p on shutdown: sv %s, peer %s, "
 520                                 "wi %s scheduled %d running %d, "
 521                                 "ev fired %d type %d status %d lnet %d\n",
 522                                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 523                                 swi_state2str(rpc->srpc_wi.wi_state),
 524                                 rpc->srpc_wi.wi_scheduled,
 525                                 rpc->srpc_wi.wi_running,
 526                                 rpc->srpc_ev.ev_fired,
 527                                 rpc->srpc_ev.ev_type,
 528                                 rpc->srpc_ev.ev_status,
 529                                 rpc->srpc_ev.ev_lnet);
 530                 }
 531
 532                 cfs_spin_unlock(&sv->sv_lock);
 533                 return 0;
 534         }
 535
 536         cfs_spin_unlock(&sv->sv_lock); /* no lock needed from now on */
 537
 538         for (;;) {
 539                 cfs_list_t *q;
 540
 541                 if (!cfs_list_empty(&sv->sv_posted_msgq))
 542                         q = &sv->sv_posted_msgq;
 543                 else if (!cfs_list_empty(&sv->sv_blocked_msgq))
 544                         q = &sv->sv_blocked_msgq;
 545                 else
 546                         break;
 547
 548                 buf = cfs_list_entry(q->next, srpc_buffer_t, buf_list);
 549                 cfs_list_del(&buf->buf_list);
 550
 551                 LIBCFS_FREE(buf, sizeof(*buf));
 552         }
 553
 554         while (!cfs_list_empty(&sv->sv_free_rpcq)) {
 555                 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
 556                                      srpc_server_rpc_t, srpc_list);
 557                 cfs_list_del(&rpc->srpc_list);
 558                 LIBCFS_FREE(rpc, sizeof(*rpc));
 559         }
 560
 561         return 1;
 562 }
 563
 564 /* called with sv->sv_lock held */
 565 void
 566 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 567 {
 568         if (sv->sv_shuttingdown) goto free;
 569
 570         if (sv->sv_nprune == 0) {
 571                 if (srpc_service_post_buffer(sv, buf) != 0)
 572                         CWARN ("Failed to post %s buffer\n", sv->sv_name);
 573                 return;
 574         }
 575
 576         sv->sv_nprune--;
 577 free:
 578         cfs_spin_unlock(&sv->sv_lock);
 579         LIBCFS_FREE(buf, sizeof(*buf));
 580         cfs_spin_lock(&sv->sv_lock);
 581 }
 582
 583 /* called with srpc_service_t::sv_lock held */
 584 inline void
 585 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
 586 {
 587         srpc_service_t *sv = rpc->srpc_service;
 588
 589         if (sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 590                 swi_schedule_workitem(&rpc->srpc_wi);
 591         else    /* framework RPCs are handled one by one */
 592                 swi_schedule_serial_workitem(&rpc->srpc_wi);
 593
 594         return;
 595 }
 596
 597 void
 598 srpc_abort_service (srpc_service_t *sv)
 599 {
 600         srpc_server_rpc_t *rpc;
 601
 602         cfs_spin_lock(&sv->sv_lock);
 603
 604         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 605                sv->sv_id, sv->sv_name);
 606
 607         /* schedule in-flight RPCs to notice the abort, NB:
 608          * racing with incoming RPCs; complete fix should make test
 609          * RPCs carry session ID in its headers */
 610         cfs_list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
 611                 rpc->srpc_aborted = 1;
 612                 srpc_schedule_server_rpc(rpc);
 613         }
 614
 615         cfs_spin_unlock(&sv->sv_lock);
 616         return;
 617 }
 618
 619 void
 620 srpc_shutdown_service (srpc_service_t *sv)
 621 {
 622         srpc_server_rpc_t *rpc;
 623         srpc_buffer_t     *buf;
 624
 625         cfs_spin_lock(&sv->sv_lock);
 626
 627         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 628                sv->sv_id, sv->sv_name);
 629
 630         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 631
 632         /* schedule in-flight RPCs to notice the shutdown */
 633         cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
 634                                        srpc_server_rpc_t, srpc_list) {
 635                 srpc_schedule_server_rpc(rpc);
 636         }
 637
 638         cfs_spin_unlock(&sv->sv_lock);
 639
 640         /* OK to traverse sv_posted_msgq without lock, since no one
 641          * touches sv_posted_msgq now */
 642         cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
 643                                        srpc_buffer_t, buf_list)
 644                 LNetMDUnlink(buf->buf_mdh);
 645
 646         return;
 647 }
 648
 649 int
 650 srpc_send_request (srpc_client_rpc_t *rpc)
 651 {
 652         srpc_event_t *ev = &rpc->crpc_reqstev;
 653         int           rc;
 654
 655         ev->ev_fired = 0;
 656         ev->ev_data  = rpc;
 657         ev->ev_type  = SRPC_REQUEST_SENT;
 658
 659         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 660                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 661                                      &rpc->crpc_reqstmdh, ev);
 662         if (rc != 0) {
 663                 LASSERT (rc == -ENOMEM);
 664                 ev->ev_fired = 1;  /* no more event expected */
 665         }
 666         return rc;
 667 }
 668
 669 int
 670 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 671 {
 672         srpc_event_t *ev = &rpc->crpc_replyev;
 673         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 674         int           rc;
 675
 676         ev->ev_fired = 0;
 677         ev->ev_data  = rpc;
 678         ev->ev_type  = SRPC_REPLY_RCVD;
 679
 680         *id = srpc_next_id();
 681
 682         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 683                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 684                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 685                                     &rpc->crpc_replymdh, ev);
 686         if (rc != 0) {
 687                 LASSERT (rc == -ENOMEM);
 688                 ev->ev_fired = 1;  /* no more event expected */
 689         }
 690         return rc;
 691 }
 692
 693 int
 694 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 695 {
 696         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 697         srpc_event_t *ev = &rpc->crpc_bulkev;
 698         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 699         int           rc;
 700         int           opt;
 701
 702         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 703
 704         if (bk->bk_niov == 0) return 0; /* nothing to do */
 705
 706         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 707 #ifdef __KERNEL__
 708         opt |= LNET_MD_KIOV;
 709 #else
 710         opt |= LNET_MD_IOVEC;
 711 #endif
 712
 713         ev->ev_fired = 0;
 714         ev->ev_data  = rpc;
 715         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 716
 717         *id = srpc_next_id();
 718
 719         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 720                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 721                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 722         if (rc != 0) {
 723                 LASSERT (rc == -ENOMEM);
 724                 ev->ev_fired = 1;  /* no more event expected */
 725         }
 726         return rc;
 727 }
 728
 729 int
 730 srpc_do_bulk (srpc_server_rpc_t *rpc)
 731 {
 732         srpc_event_t  *ev = &rpc->srpc_ev;
 733         srpc_bulk_t   *bk = rpc->srpc_bulk;
 734         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 735         int            rc;
 736         int            opt;
 737
 738         LASSERT (bk != NULL);
 739
 740         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 741 #ifdef __KERNEL__
 742         opt |= LNET_MD_KIOV;
 743 #else
 744         opt |= LNET_MD_IOVEC;
 745 #endif
 746
 747         ev->ev_fired = 0;
 748         ev->ev_data  = rpc;
 749         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 750
 751         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 752                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 753                                    rpc->srpc_peer, rpc->srpc_self,
 754                                    &bk->bk_mdh, ev);
 755         if (rc != 0)
 756                 ev->ev_fired = 1;  /* no more event expected */
 757         return rc;
 758 }
 759
 760 /* only called from srpc_handle_rpc */
 761 void
 762 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
 763 {
 764         srpc_service_t *sv = rpc->srpc_service;
 765         srpc_buffer_t  *buffer;
 766
 767         LASSERT (status != 0 || rpc->srpc_wi.wi_state == SWI_STATE_DONE);
 768
 769         rpc->srpc_status = status;
 770
 771         CDEBUG (status == 0 ? D_NET : D_NETERROR,
 772                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 773                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 774                 swi_state2str(rpc->srpc_wi.wi_state), status);
 775
 776         if (status != 0) {
 777                 cfs_spin_lock(&srpc_data.rpc_glock);
 778                 srpc_data.rpc_counters.rpcs_dropped++;
 779                 cfs_spin_unlock(&srpc_data.rpc_glock);
 780         }
 781
 782         if (rpc->srpc_done != NULL)
 783                 (*rpc->srpc_done) (rpc);
 784         LASSERT (rpc->srpc_bulk == NULL);
 785
 786         cfs_spin_lock(&sv->sv_lock);
 787
 788         if (rpc->srpc_reqstbuf != NULL) {
 789                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 790                  * sv won't go away for sv_active_rpcq must not be empty */
 791                 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
 792                 rpc->srpc_reqstbuf = NULL;
 793         }
 794
 795         cfs_list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
 796
 797         /*
 798          * No one can schedule me now since:
 799          * - I'm not on sv_active_rpcq.
 800          * - all LNet events have been fired.
 801          * Cancel pending schedules and prevent future schedule attempts:
 802          */
 803         LASSERT (rpc->srpc_ev.ev_fired);
 804         swi_kill_workitem(&rpc->srpc_wi);
 805
 806         if (!sv->sv_shuttingdown && !cfs_list_empty(&sv->sv_blocked_msgq)) {
 807                 buffer = cfs_list_entry(sv->sv_blocked_msgq.next,
 808                                     srpc_buffer_t, buf_list);
 809                 cfs_list_del(&buffer->buf_list);
 810
 811                 srpc_init_server_rpc(rpc, sv, buffer);
 812                 cfs_list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
 813                 srpc_schedule_server_rpc(rpc);
 814         } else {
 815                 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 816         }
 817
 818         cfs_spin_unlock(&sv->sv_lock);
 819         return;
 820 }
 821
 822 /* handles an incoming RPC */
 823 int
 824 srpc_handle_rpc (swi_workitem_t *wi)
 825 {
 826         srpc_server_rpc_t *rpc = wi->wi_data;
 827         srpc_service_t    *sv = rpc->srpc_service;
 828         srpc_event_t      *ev = &rpc->srpc_ev;
 829         int                rc = 0;
 830
 831         LASSERT (wi == &rpc->srpc_wi);
 832
 833         cfs_spin_lock(&sv->sv_lock);
 834
 835         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 836                 cfs_spin_unlock(&sv->sv_lock);
 837
 838                 if (rpc->srpc_bulk != NULL)
 839                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 840                 LNetMDUnlink(rpc->srpc_replymdh);
 841
 842                 if (ev->ev_fired) { /* no more event, OK to finish */
 843                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 844                         return 1;
 845                 }
 846                 return 0;
 847         }
 848
 849         cfs_spin_unlock(&sv->sv_lock);
 850
 851         switch (wi->wi_state) {
 852         default:
 853                 LBUG ();
 854         case SWI_STATE_NEWBORN: {
 855                 srpc_msg_t           *msg;
 856                 srpc_generic_reply_t *reply;
 857
 858                 msg = &rpc->srpc_reqstbuf->buf_msg;
 859                 reply = &rpc->srpc_replymsg.msg_body.reply;
 860
 861                 if (msg->msg_magic == 0) {
 862                         /* moaned already in srpc_lnet_ev_handler */
 863                         rc = EBADMSG;
 864                 } else if (msg->msg_version != SRPC_MSG_VERSION &&
 865                            msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
 866                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
 867                                msg->msg_version, SRPC_MSG_VERSION,
 868                                libcfs_id2str(rpc->srpc_peer));
 869                         reply->status = EPROTO;
 870                 } else {
 871                         reply->status = 0;
 872                         rc = (*sv->sv_handler) (rpc);
 873                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
 874                 }
 875
 876                 if (rc != 0) {
 877                         srpc_server_rpc_done(rpc, rc);
 878                         return 1;
 879                 }
 880
 881                 wi->wi_state = SWI_STATE_BULK_STARTED;
 882
 883                 if (rpc->srpc_bulk != NULL) {
 884                         rc = srpc_do_bulk(rpc);
 885                         if (rc == 0)
 886                                 return 0; /* wait for bulk */
 887
 888                         LASSERT (ev->ev_fired);
 889                         ev->ev_status = rc;
 890                 }
 891         }
 892         case SWI_STATE_BULK_STARTED:
 893                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
 894
 895                 if (rpc->srpc_bulk != NULL) {
 896                         rc = ev->ev_status;
 897
 898                         if (sv->sv_bulk_ready != NULL)
 899                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
 900
 901                         if (rc != 0) {
 902                                 srpc_server_rpc_done(rpc, rc);
 903                                 return 1;
 904                         }
 905                 }
 906
 907                 wi->wi_state = SWI_STATE_REPLY_SUBMITTED;
 908                 rc = srpc_send_reply(rpc);
 909                 if (rc == 0)
 910                         return 0; /* wait for reply */
 911                 srpc_server_rpc_done(rpc, rc);
 912                 return 1;
 913
 914         case SWI_STATE_REPLY_SUBMITTED:
 915                 if (!ev->ev_fired) {
 916                         CERROR("RPC %p: bulk %p, service %d\n",
 917                                rpc, rpc->srpc_bulk, rpc->srpc_service->sv_id);
 918                         CERROR("Event: status %d, type %d, lnet %d\n",
 919                                ev->ev_status, ev->ev_type, ev->ev_lnet);
 920                         LASSERT (ev->ev_fired);
 921                 }
 922
 923                 wi->wi_state = SWI_STATE_DONE;
 924                 srpc_server_rpc_done(rpc, ev->ev_status);
 925                 return 1;
 926         }
 927
 928         return 0;
 929 }
 930
 931 void
 932 srpc_client_rpc_expired (void *data)
 933 {
 934         srpc_client_rpc_t *rpc = data;
 935
 936         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
 937                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
 938                rpc->crpc_timeout);
 939
 940         cfs_spin_lock(&rpc->crpc_lock);
 941
 942         rpc->crpc_timeout = 0;
 943         srpc_abort_rpc(rpc, -ETIMEDOUT);
 944
 945         cfs_spin_unlock(&rpc->crpc_lock);
 946
 947         cfs_spin_lock(&srpc_data.rpc_glock);
 948         srpc_data.rpc_counters.rpcs_expired++;
 949         cfs_spin_unlock(&srpc_data.rpc_glock);
 950         return;
 951 }
 952
 953 inline void
 954 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
 955 {
 956         stt_timer_t *timer = &rpc->crpc_timer;
 957
 958         if (rpc->crpc_timeout == 0) return;
 959
 960         CFS_INIT_LIST_HEAD(&timer->stt_list);
 961         timer->stt_data    = rpc;
 962         timer->stt_func    = srpc_client_rpc_expired;
 963         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
 964                                           cfs_time_current_sec());
 965         stt_add_timer(timer);
 966         return;
 967 }
 968
 969 /*
 970  * Called with rpc->crpc_lock held.
 971  *
 972  * Upon exit the RPC expiry timer is not queued and the handler is not
 973  * running on any CPU. */
 974 void
 975 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
 976 {
 977         /* timer not planted or already exploded */
 978         if (rpc->crpc_timeout == 0) return;
 979
 980         /* timer sucessfully defused */
 981         if (stt_del_timer(&rpc->crpc_timer)) return;
 982
 983 #ifdef __KERNEL__
 984         /* timer detonated, wait for it to explode */
 985         while (rpc->crpc_timeout != 0) {
 986                 cfs_spin_unlock(&rpc->crpc_lock);
 987
 988                 cfs_schedule();
 989
 990                 cfs_spin_lock(&rpc->crpc_lock);
 991         }
 992 #else
 993         LBUG(); /* impossible in single-threaded runtime */
 994 #endif
 995         return;
 996 }
 997
 998 void
 999 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1000 {
1001         swi_workitem_t *wi = &rpc->crpc_wi;
1002
1003         LASSERT (status != 0 || wi->wi_state == SWI_STATE_DONE);
1004
1005         cfs_spin_lock(&rpc->crpc_lock);
1006
1007         rpc->crpc_closed = 1;
1008         if (rpc->crpc_status == 0)
1009                 rpc->crpc_status = status;
1010
1011         srpc_del_client_rpc_timer(rpc);
1012
1013         CDEBUG ((status == 0) ? D_NET : D_NETERROR,
1014                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1015                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1016                 swi_state2str(wi->wi_state), rpc->crpc_aborted, status);
1017
1018         /*
1019          * No one can schedule me now since:
1020          * - RPC timer has been defused.
1021          * - all LNet events have been fired.
1022          * - crpc_closed has been set, preventing srpc_abort_rpc from
1023          *   scheduling me.
1024          * Cancel pending schedules and prevent future schedule attempts:
1025          */
1026         LASSERT (!srpc_event_pending(rpc));
1027         swi_kill_workitem(wi);
1028
1029         cfs_spin_unlock(&rpc->crpc_lock);
1030
1031         (*rpc->crpc_done) (rpc);
1032         return;
1033 }
1034
1035 /* sends an outgoing RPC */
1036 int
1037 srpc_send_rpc (swi_workitem_t *wi)
1038 {
1039         int                rc = 0;
1040         srpc_client_rpc_t *rpc = wi->wi_data;
1041         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1042         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1043
1044         LASSERT (rpc != NULL);
1045         LASSERT (wi == &rpc->crpc_wi);
1046
1047         cfs_spin_lock(&rpc->crpc_lock);
1048
1049         if (rpc->crpc_aborted) {
1050                 cfs_spin_unlock(&rpc->crpc_lock);
1051                 goto abort;
1052         }
1053
1054         cfs_spin_unlock(&rpc->crpc_lock);
1055
1056         switch (wi->wi_state) {
1057         default:
1058                 LBUG ();
1059         case SWI_STATE_NEWBORN:
1060                 LASSERT (!srpc_event_pending(rpc));
1061
1062                 rc = srpc_prepare_reply(rpc);
1063                 if (rc != 0) {
1064                         srpc_client_rpc_done(rpc, rc);
1065                         return 1;
1066                 }
1067
1068                 rc = srpc_prepare_bulk(rpc);
1069                 if (rc != 0) break;
1070
1071                 wi->wi_state = SWI_STATE_REQUEST_SUBMITTED;
1072                 rc = srpc_send_request(rpc);
1073                 break;
1074
1075         case SWI_STATE_REQUEST_SUBMITTED:
1076                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1077                  * order; however, they're processed in a strict order:
1078                  * rqt, rpy, and bulk. */
1079                 if (!rpc->crpc_reqstev.ev_fired) break;
1080
1081                 rc = rpc->crpc_reqstev.ev_status;
1082                 if (rc != 0) break;
1083
1084                 wi->wi_state = SWI_STATE_REQUEST_SENT;
1085                 /* perhaps more events, fall thru */
1086         case SWI_STATE_REQUEST_SENT: {
1087                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1088
1089                 if (!rpc->crpc_replyev.ev_fired) break;
1090
1091                 rc = rpc->crpc_replyev.ev_status;
1092                 if (rc != 0) break;
1093
1094                 if ((reply->msg_type != type &&
1095                      reply->msg_type != __swab32(type)) ||
1096                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1097                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1098                         CWARN ("Bad message from %s: type %u (%d expected),"
1099                                " magic %u (%d expected).\n",
1100                                libcfs_id2str(rpc->crpc_dest),
1101                                reply->msg_type, type,
1102                                reply->msg_magic, SRPC_MSG_MAGIC);
1103                         rc = -EBADMSG;
1104                         break;
1105                 }
1106
1107                 if (do_bulk && reply->msg_body.reply.status != 0) {
1108                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1109                                "case peer didn't initiate bulk transfer\n",
1110                                reply->msg_body.reply.status,
1111                                libcfs_id2str(rpc->crpc_dest));
1112                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1113                 }
1114
1115                 wi->wi_state = SWI_STATE_REPLY_RECEIVED;
1116         }
1117         case SWI_STATE_REPLY_RECEIVED:
1118                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1119
1120                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1121
1122                 /* Bulk buffer was unlinked due to remote error. Clear error
1123                  * since reply buffer still contains valid data.
1124                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1125                  * remote error. */
1126                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1127                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1128                         rc = 0;
1129
1130                 wi->wi_state = SWI_STATE_DONE;
1131                 srpc_client_rpc_done(rpc, rc);
1132                 return 1;
1133         }
1134
1135         if (rc != 0) {
1136                 cfs_spin_lock(&rpc->crpc_lock);
1137                 srpc_abort_rpc(rpc, rc);
1138                 cfs_spin_unlock(&rpc->crpc_lock);
1139         }
1140
1141 abort:
1142         if (rpc->crpc_aborted) {
1143                 LNetMDUnlink(rpc->crpc_reqstmdh);
1144                 LNetMDUnlink(rpc->crpc_replymdh);
1145                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1146
1147                 if (!srpc_event_pending(rpc)) {
1148                         srpc_client_rpc_done(rpc, -EINTR);
1149                         return 1;
1150                 }
1151         }
1152         return 0;
1153 }
1154
1155 srpc_client_rpc_t *
1156 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1157                         int nbulkiov, int bulklen,
1158                         void (*rpc_done)(srpc_client_rpc_t *),
1159                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1160 {
1161         srpc_client_rpc_t *rpc;
1162
1163         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1164                                    crpc_bulk.bk_iovs[nbulkiov]));
1165         if (rpc == NULL)
1166                 return NULL;
1167
1168         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1169                              bulklen, rpc_done, rpc_fini, priv);
1170         return rpc;
1171 }
1172
1173 /* called with rpc->crpc_lock held */
1174 void
1175 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1176 {
1177         LASSERT (why != 0);
1178
1179         if (rpc->crpc_aborted || /* already aborted */
1180             rpc->crpc_closed)    /* callback imminent */
1181                 return;
1182
1183         CDEBUG (D_NET,
1184                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1185                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1186                 swi_state2str(rpc->crpc_wi.wi_state), why);
1187
1188         rpc->crpc_aborted = 1;
1189         rpc->crpc_status  = why;
1190         swi_schedule_workitem(&rpc->crpc_wi);
1191         return;
1192 }
1193
1194 /* called with rpc->crpc_lock held */
1195 void
1196 srpc_post_rpc (srpc_client_rpc_t *rpc)
1197 {
1198         LASSERT (!rpc->crpc_aborted);
1199         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1200         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1201
1202         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1203                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1204                 rpc->crpc_timeout);
1205
1206         srpc_add_client_rpc_timer(rpc);
1207         swi_schedule_workitem(&rpc->crpc_wi);
1208         return;
1209 }
1210
1211
1212 int
1213 srpc_send_reply (srpc_server_rpc_t *rpc)
1214 {
1215         srpc_event_t   *ev = &rpc->srpc_ev;
1216         srpc_msg_t     *msg = &rpc->srpc_replymsg;
1217         srpc_buffer_t  *buffer = rpc->srpc_reqstbuf;
1218         srpc_service_t *sv = rpc->srpc_service;
1219         __u64           rpyid;
1220         int             rc;
1221
1222         LASSERT (buffer != NULL);
1223         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1224
1225         cfs_spin_lock(&sv->sv_lock);
1226
1227         if (!sv->sv_shuttingdown &&
1228             sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1229                 /* Repost buffer before replying since test client
1230                  * might send me another RPC once it gets the reply */
1231                 if (srpc_service_post_buffer(sv, buffer) != 0)
1232                         CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1233                 rpc->srpc_reqstbuf = NULL;
1234         }
1235
1236         cfs_spin_unlock(&sv->sv_lock);
1237
1238         ev->ev_fired = 0;
1239         ev->ev_data  = rpc;
1240         ev->ev_type  = SRPC_REPLY_SENT;
1241
1242         msg->msg_magic   = SRPC_MSG_MAGIC;
1243         msg->msg_version = SRPC_MSG_VERSION;
1244         msg->msg_type    = srpc_service2reply(sv->sv_id);
1245
1246         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1247                                    sizeof(*msg), LNET_MD_OP_PUT,
1248                                    rpc->srpc_peer, rpc->srpc_self,
1249                                    &rpc->srpc_replymdh, ev);
1250         if (rc != 0)
1251                 ev->ev_fired = 1;  /* no more event expected */
1252         return rc;
1253 }
1254
1255 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1256 void
1257 srpc_lnet_ev_handler (lnet_event_t *ev)
1258 {
1259         srpc_event_t      *rpcev = ev->md.user_ptr;
1260         srpc_client_rpc_t *crpc;
1261         srpc_server_rpc_t *srpc;
1262         srpc_buffer_t     *buffer;
1263         srpc_service_t    *sv;
1264         srpc_msg_t        *msg;
1265         srpc_msg_type_t    type;
1266
1267         LASSERT (!cfs_in_interrupt());
1268
1269         if (ev->status != 0) {
1270                 cfs_spin_lock(&srpc_data.rpc_glock);
1271                 srpc_data.rpc_counters.errors++;
1272                 cfs_spin_unlock(&srpc_data.rpc_glock);
1273         }
1274
1275         rpcev->ev_lnet = ev->type;
1276
1277         switch (rpcev->ev_type) {
1278         default:
1279                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1280                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1281                 LBUG ();
1282         case SRPC_REQUEST_SENT:
1283                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1284                         cfs_spin_lock(&srpc_data.rpc_glock);
1285                         srpc_data.rpc_counters.rpcs_sent++;
1286                         cfs_spin_unlock(&srpc_data.rpc_glock);
1287                 }
1288         case SRPC_REPLY_RCVD:
1289         case SRPC_BULK_REQ_RCVD:
1290                 crpc = rpcev->ev_data;
1291
1292                 if (rpcev != &crpc->crpc_reqstev &&
1293                     rpcev != &crpc->crpc_replyev &&
1294                     rpcev != &crpc->crpc_bulkev) {
1295                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1296                                rpcev, crpc, &crpc->crpc_reqstev,
1297                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1298                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1299                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1300                         LBUG ();
1301                 }
1302
1303                 cfs_spin_lock(&crpc->crpc_lock);
1304
1305                 LASSERT (rpcev->ev_fired == 0);
1306                 rpcev->ev_fired  = 1;
1307                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1308                                                 -EINTR : ev->status;
1309                 swi_schedule_workitem(&crpc->crpc_wi);
1310
1311                 cfs_spin_unlock(&crpc->crpc_lock);
1312                 break;
1313
1314         case SRPC_REQUEST_RCVD:
1315                 sv = rpcev->ev_data;
1316
1317                 LASSERT (rpcev == &sv->sv_ev);
1318
1319                 cfs_spin_lock(&sv->sv_lock);
1320
1321                 LASSERT (ev->unlinked);
1322                 LASSERT (ev->type == LNET_EVENT_PUT ||
1323                          ev->type == LNET_EVENT_UNLINK);
1324                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1325                          sv->sv_shuttingdown);
1326
1327                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1328                 buffer->buf_peer = ev->initiator;
1329                 buffer->buf_self = ev->target.nid;
1330
1331                 sv->sv_nposted_msg--;
1332                 LASSERT (sv->sv_nposted_msg >= 0);
1333
1334                 if (sv->sv_shuttingdown) {
1335                         /* Leave buffer on sv->sv_posted_msgq since
1336                          * srpc_finish_service needs to traverse it. */
1337                         cfs_spin_unlock(&sv->sv_lock);
1338                         break;
1339                 }
1340
1341                 cfs_list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1342                 msg = &buffer->buf_msg;
1343                 type = srpc_service2request(sv->sv_id);
1344
1345                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1346                     (msg->msg_type != type &&
1347                      msg->msg_type != __swab32(type)) ||
1348                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1349                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1350                         CERROR ("Dropping RPC (%s) from %s: "
1351                                 "status %d mlength %d type %u magic %u.\n",
1352                                 sv->sv_name, libcfs_id2str(ev->initiator),
1353                                 ev->status, ev->mlength,
1354                                 msg->msg_type, msg->msg_magic);
1355
1356                         /* NB can't call srpc_service_recycle_buffer here since
1357                          * it may call LNetM[DE]Attach. The invalid magic tells
1358                          * srpc_handle_rpc to drop this RPC */
1359                         msg->msg_magic = 0;
1360                 }
1361
1362                 if (!cfs_list_empty(&sv->sv_free_rpcq)) {
1363                         srpc = cfs_list_entry(sv->sv_free_rpcq.next,
1364                                               srpc_server_rpc_t, srpc_list);
1365                         cfs_list_del(&srpc->srpc_list);
1366
1367                         srpc_init_server_rpc(srpc, sv, buffer);
1368                         cfs_list_add_tail(&srpc->srpc_list,
1369                                           &sv->sv_active_rpcq);
1370                         srpc_schedule_server_rpc(srpc);
1371                 } else {
1372                         cfs_list_add_tail(&buffer->buf_list,
1373                                           &sv->sv_blocked_msgq);
1374                 }
1375
1376                 cfs_spin_unlock(&sv->sv_lock);
1377
1378                 cfs_spin_lock(&srpc_data.rpc_glock);
1379                 srpc_data.rpc_counters.rpcs_rcvd++;
1380                 cfs_spin_unlock(&srpc_data.rpc_glock);
1381                 break;
1382
1383         case SRPC_BULK_GET_RPLD:
1384                 LASSERT (ev->type == LNET_EVENT_SEND ||
1385                          ev->type == LNET_EVENT_REPLY ||
1386                          ev->type == LNET_EVENT_UNLINK);
1387
1388                 if (!ev->unlinked)
1389                         break; /* wait for final event */
1390
1391         case SRPC_BULK_PUT_SENT:
1392                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1393                         cfs_spin_lock(&srpc_data.rpc_glock);
1394
1395                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1396                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1397                         else
1398                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1399
1400                         cfs_spin_unlock(&srpc_data.rpc_glock);
1401                 }
1402         case SRPC_REPLY_SENT:
1403                 srpc = rpcev->ev_data;
1404                 sv = srpc->srpc_service;
1405
1406                 LASSERT (rpcev == &srpc->srpc_ev);
1407
1408                 cfs_spin_lock(&sv->sv_lock);
1409
1410                 rpcev->ev_fired  = 1;
1411                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1412                                                 -EINTR : ev->status;
1413                 srpc_schedule_server_rpc(srpc);
1414
1415                 cfs_spin_unlock(&sv->sv_lock);
1416                 break;
1417         }
1418
1419         return;
1420 }
1421
1422 #ifndef __KERNEL__
1423
1424 int
1425 srpc_check_event (int timeout)
1426 {
1427         lnet_event_t ev;
1428         int          rc;
1429         int          i;
1430
1431         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1432                         timeout * 1000, &ev, &i);
1433         if (rc == 0) return 0;
1434
1435         LASSERT (rc == -EOVERFLOW || rc == 1);
1436
1437         /* We can't affort to miss any events... */
1438         if (rc == -EOVERFLOW) {
1439                 CERROR ("Dropped an event!!!\n");
1440                 abort();
1441         }
1442
1443         srpc_lnet_ev_handler(&ev);
1444         return 1;
1445 }
1446
1447 #endif
1448
1449 int
1450 srpc_startup (void)
1451 {
1452         int rc;
1453
1454         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1455         cfs_spin_lock_init(&srpc_data.rpc_glock);
1456
1457         /* 1 second pause to avoid timestamp reuse */
1458         cfs_pause(cfs_time_seconds(1));
1459         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1460
1461         srpc_data.rpc_state = SRPC_STATE_NONE;
1462
1463 #ifdef __KERNEL__
1464         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1465 #else
1466         if (the_lnet.ln_server_mode_flag)
1467                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1468         else
1469                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1470 #endif
1471         if (rc < 0) {
1472                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1473                 return rc;
1474         }
1475
1476         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1477
1478         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1479 #ifdef __KERNEL__
1480         rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1481 #else
1482         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1483 #endif
1484         if (rc != 0) {
1485                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1486                 goto bail;
1487         }
1488
1489         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1490         LASSERT (rc == 0);
1491
1492         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1493
1494         rc = swi_startup();
1495         if (rc != 0)
1496                 goto bail;
1497
1498         srpc_data.rpc_state = SRPC_STATE_WI_INIT;
1499
1500         rc = stt_startup();
1501
1502 bail:
1503         if (rc != 0)
1504                 srpc_shutdown();
1505         else
1506                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1507
1508         return rc;
1509 }
1510
1511 void
1512 srpc_shutdown (void)
1513 {
1514         int i;
1515         int rc;
1516         int state;
1517
1518         state = srpc_data.rpc_state;
1519         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1520
1521         switch (state) {
1522         default:
1523                 LBUG ();
1524         case SRPC_STATE_RUNNING:
1525                 cfs_spin_lock(&srpc_data.rpc_glock);
1526
1527                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1528                         srpc_service_t *sv = srpc_data.rpc_services[i];
1529
1530                         LASSERTF (sv == NULL,
1531                                   "service not empty: id %d, name %s\n",
1532                                   i, sv->sv_name);
1533                 }
1534
1535                 cfs_spin_unlock(&srpc_data.rpc_glock);
1536
1537                 stt_shutdown();
1538
1539         case SRPC_STATE_WI_INIT:
1540                 swi_shutdown();
1541
1542         case SRPC_STATE_EQ_INIT:
1543                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1544                 LASSERT (rc == 0);
1545                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1546                 LASSERT (rc == 0); /* the EQ should have no user by now */
1547
1548         case SRPC_STATE_NI_INIT:
1549                 LNetNIFini();
1550                 break;
1551         }
1552
1553         return;
1554 }