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                         CNETERR("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.swi_state),
 524                                 rpc->srpc_wi.swi_workitem.wi_scheduled,
 525                                 rpc->srpc_wi.swi_workitem.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         swi_schedule_workitem(&rpc->srpc_wi);
 588 }
 589
 590 void
 591 srpc_abort_service (srpc_service_t *sv)
 592 {
 593         srpc_server_rpc_t *rpc;
 594
 595         cfs_spin_lock(&sv->sv_lock);
 596
 597         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 598                sv->sv_id, sv->sv_name);
 599
 600         /* schedule in-flight RPCs to notice the abort, NB:
 601          * racing with incoming RPCs; complete fix should make test
 602          * RPCs carry session ID in its headers */
 603         cfs_list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
 604                 rpc->srpc_aborted = 1;
 605                 srpc_schedule_server_rpc(rpc);
 606         }
 607
 608         cfs_spin_unlock(&sv->sv_lock);
 609         return;
 610 }
 611
 612 void
 613 srpc_shutdown_service (srpc_service_t *sv)
 614 {
 615         srpc_server_rpc_t *rpc;
 616         srpc_buffer_t     *buf;
 617
 618         cfs_spin_lock(&sv->sv_lock);
 619
 620         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 621                sv->sv_id, sv->sv_name);
 622
 623         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 624
 625         /* schedule in-flight RPCs to notice the shutdown */
 626         cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
 627                                        srpc_server_rpc_t, srpc_list) {
 628                 srpc_schedule_server_rpc(rpc);
 629         }
 630
 631         cfs_spin_unlock(&sv->sv_lock);
 632
 633         /* OK to traverse sv_posted_msgq without lock, since no one
 634          * touches sv_posted_msgq now */
 635         cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
 636                                        srpc_buffer_t, buf_list)
 637                 LNetMDUnlink(buf->buf_mdh);
 638
 639         return;
 640 }
 641
 642 int
 643 srpc_send_request (srpc_client_rpc_t *rpc)
 644 {
 645         srpc_event_t *ev = &rpc->crpc_reqstev;
 646         int           rc;
 647
 648         ev->ev_fired = 0;
 649         ev->ev_data  = rpc;
 650         ev->ev_type  = SRPC_REQUEST_SENT;
 651
 652         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 653                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 654                                      &rpc->crpc_reqstmdh, ev);
 655         if (rc != 0) {
 656                 LASSERT (rc == -ENOMEM);
 657                 ev->ev_fired = 1;  /* no more event expected */
 658         }
 659         return rc;
 660 }
 661
 662 int
 663 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 664 {
 665         srpc_event_t *ev = &rpc->crpc_replyev;
 666         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 667         int           rc;
 668
 669         ev->ev_fired = 0;
 670         ev->ev_data  = rpc;
 671         ev->ev_type  = SRPC_REPLY_RCVD;
 672
 673         *id = srpc_next_id();
 674
 675         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 676                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 677                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 678                                     &rpc->crpc_replymdh, ev);
 679         if (rc != 0) {
 680                 LASSERT (rc == -ENOMEM);
 681                 ev->ev_fired = 1;  /* no more event expected */
 682         }
 683         return rc;
 684 }
 685
 686 int
 687 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 688 {
 689         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 690         srpc_event_t *ev = &rpc->crpc_bulkev;
 691         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 692         int           rc;
 693         int           opt;
 694
 695         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 696
 697         if (bk->bk_niov == 0) return 0; /* nothing to do */
 698
 699         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 700 #ifdef __KERNEL__
 701         opt |= LNET_MD_KIOV;
 702 #else
 703         opt |= LNET_MD_IOVEC;
 704 #endif
 705
 706         ev->ev_fired = 0;
 707         ev->ev_data  = rpc;
 708         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 709
 710         *id = srpc_next_id();
 711
 712         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 713                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 714                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 715         if (rc != 0) {
 716                 LASSERT (rc == -ENOMEM);
 717                 ev->ev_fired = 1;  /* no more event expected */
 718         }
 719         return rc;
 720 }
 721
 722 int
 723 srpc_do_bulk (srpc_server_rpc_t *rpc)
 724 {
 725         srpc_event_t  *ev = &rpc->srpc_ev;
 726         srpc_bulk_t   *bk = rpc->srpc_bulk;
 727         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 728         int            rc;
 729         int            opt;
 730
 731         LASSERT (bk != NULL);
 732
 733         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 734 #ifdef __KERNEL__
 735         opt |= LNET_MD_KIOV;
 736 #else
 737         opt |= LNET_MD_IOVEC;
 738 #endif
 739
 740         ev->ev_fired = 0;
 741         ev->ev_data  = rpc;
 742         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 743
 744         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 745                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 746                                    rpc->srpc_peer, rpc->srpc_self,
 747                                    &bk->bk_mdh, ev);
 748         if (rc != 0)
 749                 ev->ev_fired = 1;  /* no more event expected */
 750         return rc;
 751 }
 752
 753 /* only called from srpc_handle_rpc */
 754 void
 755 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
 756 {
 757         srpc_service_t *sv = rpc->srpc_service;
 758         srpc_buffer_t  *buffer;
 759
 760         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 761
 762         rpc->srpc_status = status;
 763
 764         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 765                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 766                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 767                 swi_state2str(rpc->srpc_wi.swi_state), status);
 768
 769         if (status != 0) {
 770                 cfs_spin_lock(&srpc_data.rpc_glock);
 771                 srpc_data.rpc_counters.rpcs_dropped++;
 772                 cfs_spin_unlock(&srpc_data.rpc_glock);
 773         }
 774
 775         if (rpc->srpc_done != NULL)
 776                 (*rpc->srpc_done) (rpc);
 777         LASSERT (rpc->srpc_bulk == NULL);
 778
 779         cfs_spin_lock(&sv->sv_lock);
 780
 781         if (rpc->srpc_reqstbuf != NULL) {
 782                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 783                  * sv won't go away for sv_active_rpcq must not be empty */
 784                 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
 785                 rpc->srpc_reqstbuf = NULL;
 786         }
 787
 788         cfs_list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
 789
 790         /*
 791          * No one can schedule me now since:
 792          * - I'm not on sv_active_rpcq.
 793          * - all LNet events have been fired.
 794          * Cancel pending schedules and prevent future schedule attempts:
 795          */
 796         LASSERT (rpc->srpc_ev.ev_fired);
 797         swi_kill_workitem(&rpc->srpc_wi);
 798
 799         if (!sv->sv_shuttingdown && !cfs_list_empty(&sv->sv_blocked_msgq)) {
 800                 buffer = cfs_list_entry(sv->sv_blocked_msgq.next,
 801                                     srpc_buffer_t, buf_list);
 802                 cfs_list_del(&buffer->buf_list);
 803
 804                 srpc_init_server_rpc(rpc, sv, buffer);
 805                 cfs_list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
 806                 srpc_schedule_server_rpc(rpc);
 807         } else {
 808                 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 809         }
 810
 811         cfs_spin_unlock(&sv->sv_lock);
 812         return;
 813 }
 814
 815 /* handles an incoming RPC */
 816 int
 817 srpc_handle_rpc (swi_workitem_t *wi)
 818 {
 819         srpc_server_rpc_t *rpc = wi->swi_workitem.wi_data;
 820         srpc_service_t    *sv = rpc->srpc_service;
 821         srpc_event_t      *ev = &rpc->srpc_ev;
 822         int                rc = 0;
 823
 824         LASSERT (wi == &rpc->srpc_wi);
 825
 826         cfs_spin_lock(&sv->sv_lock);
 827
 828         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 829                 cfs_spin_unlock(&sv->sv_lock);
 830
 831                 if (rpc->srpc_bulk != NULL)
 832                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 833                 LNetMDUnlink(rpc->srpc_replymdh);
 834
 835                 if (ev->ev_fired) { /* no more event, OK to finish */
 836                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 837                         return 1;
 838                 }
 839                 return 0;
 840         }
 841
 842         cfs_spin_unlock(&sv->sv_lock);
 843
 844         switch (wi->swi_state) {
 845         default:
 846                 LBUG ();
 847         case SWI_STATE_NEWBORN: {
 848                 srpc_msg_t           *msg;
 849                 srpc_generic_reply_t *reply;
 850
 851                 msg = &rpc->srpc_reqstbuf->buf_msg;
 852                 reply = &rpc->srpc_replymsg.msg_body.reply;
 853
 854                 if (msg->msg_magic == 0) {
 855                         /* moaned already in srpc_lnet_ev_handler */
 856                         rc = EBADMSG;
 857                 } else if (msg->msg_version != SRPC_MSG_VERSION &&
 858                            msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
 859                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
 860                                msg->msg_version, SRPC_MSG_VERSION,
 861                                libcfs_id2str(rpc->srpc_peer));
 862                         reply->status = EPROTO;
 863                 } else {
 864                         reply->status = 0;
 865                         rc = (*sv->sv_handler) (rpc);
 866                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
 867                 }
 868
 869                 if (rc != 0) {
 870                         srpc_server_rpc_done(rpc, rc);
 871                         return 1;
 872                 }
 873
 874                 wi->swi_state = SWI_STATE_BULK_STARTED;
 875
 876                 if (rpc->srpc_bulk != NULL) {
 877                         rc = srpc_do_bulk(rpc);
 878                         if (rc == 0)
 879                                 return 0; /* wait for bulk */
 880
 881                         LASSERT (ev->ev_fired);
 882                         ev->ev_status = rc;
 883                 }
 884         }
 885         case SWI_STATE_BULK_STARTED:
 886                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
 887
 888                 if (rpc->srpc_bulk != NULL) {
 889                         rc = ev->ev_status;
 890
 891                         if (sv->sv_bulk_ready != NULL)
 892                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
 893
 894                         if (rc != 0) {
 895                                 srpc_server_rpc_done(rpc, rc);
 896                                 return 1;
 897                         }
 898                 }
 899
 900                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
 901                 rc = srpc_send_reply(rpc);
 902                 if (rc == 0)
 903                         return 0; /* wait for reply */
 904                 srpc_server_rpc_done(rpc, rc);
 905                 return 1;
 906
 907         case SWI_STATE_REPLY_SUBMITTED:
 908                 if (!ev->ev_fired) {
 909                         CERROR("RPC %p: bulk %p, service %d\n",
 910                                rpc, rpc->srpc_bulk, rpc->srpc_service->sv_id);
 911                         CERROR("Event: status %d, type %d, lnet %d\n",
 912                                ev->ev_status, ev->ev_type, ev->ev_lnet);
 913                         LASSERT (ev->ev_fired);
 914                 }
 915
 916                 wi->swi_state = SWI_STATE_DONE;
 917                 srpc_server_rpc_done(rpc, ev->ev_status);
 918                 return 1;
 919         }
 920
 921         return 0;
 922 }
 923
 924 void
 925 srpc_client_rpc_expired (void *data)
 926 {
 927         srpc_client_rpc_t *rpc = data;
 928
 929         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
 930                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
 931                rpc->crpc_timeout);
 932
 933         cfs_spin_lock(&rpc->crpc_lock);
 934
 935         rpc->crpc_timeout = 0;
 936         srpc_abort_rpc(rpc, -ETIMEDOUT);
 937
 938         cfs_spin_unlock(&rpc->crpc_lock);
 939
 940         cfs_spin_lock(&srpc_data.rpc_glock);
 941         srpc_data.rpc_counters.rpcs_expired++;
 942         cfs_spin_unlock(&srpc_data.rpc_glock);
 943         return;
 944 }
 945
 946 inline void
 947 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
 948 {
 949         stt_timer_t *timer = &rpc->crpc_timer;
 950
 951         if (rpc->crpc_timeout == 0) return;
 952
 953         CFS_INIT_LIST_HEAD(&timer->stt_list);
 954         timer->stt_data    = rpc;
 955         timer->stt_func    = srpc_client_rpc_expired;
 956         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
 957                                           cfs_time_current_sec());
 958         stt_add_timer(timer);
 959         return;
 960 }
 961
 962 /*
 963  * Called with rpc->crpc_lock held.
 964  *
 965  * Upon exit the RPC expiry timer is not queued and the handler is not
 966  * running on any CPU. */
 967 void
 968 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
 969 {
 970         /* timer not planted or already exploded */
 971         if (rpc->crpc_timeout == 0) return;
 972
 973         /* timer sucessfully defused */
 974         if (stt_del_timer(&rpc->crpc_timer)) return;
 975
 976 #ifdef __KERNEL__
 977         /* timer detonated, wait for it to explode */
 978         while (rpc->crpc_timeout != 0) {
 979                 cfs_spin_unlock(&rpc->crpc_lock);
 980
 981                 cfs_schedule();
 982
 983                 cfs_spin_lock(&rpc->crpc_lock);
 984         }
 985 #else
 986         LBUG(); /* impossible in single-threaded runtime */
 987 #endif
 988         return;
 989 }
 990
 991 void
 992 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
 993 {
 994         swi_workitem_t *wi = &rpc->crpc_wi;
 995
 996         LASSERT (status != 0 || wi->swi_state == SWI_STATE_DONE);
 997
 998         cfs_spin_lock(&rpc->crpc_lock);
 999
1000         rpc->crpc_closed = 1;
1001         if (rpc->crpc_status == 0)
1002                 rpc->crpc_status = status;
1003
1004         srpc_del_client_rpc_timer(rpc);
1005
1006         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1007                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1008                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1009                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1010
1011         /*
1012          * No one can schedule me now since:
1013          * - RPC timer has been defused.
1014          * - all LNet events have been fired.
1015          * - crpc_closed has been set, preventing srpc_abort_rpc from
1016          *   scheduling me.
1017          * Cancel pending schedules and prevent future schedule attempts:
1018          */
1019         LASSERT (!srpc_event_pending(rpc));
1020         swi_kill_workitem(wi);
1021
1022         cfs_spin_unlock(&rpc->crpc_lock);
1023
1024         (*rpc->crpc_done) (rpc);
1025         return;
1026 }
1027
1028 /* sends an outgoing RPC */
1029 int
1030 srpc_send_rpc (swi_workitem_t *wi)
1031 {
1032         int                rc = 0;
1033         srpc_client_rpc_t *rpc = wi->swi_workitem.wi_data;
1034         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1035         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1036
1037         LASSERT (rpc != NULL);
1038         LASSERT (wi == &rpc->crpc_wi);
1039
1040         cfs_spin_lock(&rpc->crpc_lock);
1041
1042         if (rpc->crpc_aborted) {
1043                 cfs_spin_unlock(&rpc->crpc_lock);
1044                 goto abort;
1045         }
1046
1047         cfs_spin_unlock(&rpc->crpc_lock);
1048
1049         switch (wi->swi_state) {
1050         default:
1051                 LBUG ();
1052         case SWI_STATE_NEWBORN:
1053                 LASSERT (!srpc_event_pending(rpc));
1054
1055                 rc = srpc_prepare_reply(rpc);
1056                 if (rc != 0) {
1057                         srpc_client_rpc_done(rpc, rc);
1058                         return 1;
1059                 }
1060
1061                 rc = srpc_prepare_bulk(rpc);
1062                 if (rc != 0) break;
1063
1064                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1065                 rc = srpc_send_request(rpc);
1066                 break;
1067
1068         case SWI_STATE_REQUEST_SUBMITTED:
1069                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1070                  * order; however, they're processed in a strict order:
1071                  * rqt, rpy, and bulk. */
1072                 if (!rpc->crpc_reqstev.ev_fired) break;
1073
1074                 rc = rpc->crpc_reqstev.ev_status;
1075                 if (rc != 0) break;
1076
1077                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1078                 /* perhaps more events, fall thru */
1079         case SWI_STATE_REQUEST_SENT: {
1080                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1081
1082                 if (!rpc->crpc_replyev.ev_fired) break;
1083
1084                 rc = rpc->crpc_replyev.ev_status;
1085                 if (rc != 0) break;
1086
1087                 if ((reply->msg_type != type &&
1088                      reply->msg_type != __swab32(type)) ||
1089                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1090                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1091                         CWARN ("Bad message from %s: type %u (%d expected),"
1092                                " magic %u (%d expected).\n",
1093                                libcfs_id2str(rpc->crpc_dest),
1094                                reply->msg_type, type,
1095                                reply->msg_magic, SRPC_MSG_MAGIC);
1096                         rc = -EBADMSG;
1097                         break;
1098                 }
1099
1100                 if (do_bulk && reply->msg_body.reply.status != 0) {
1101                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1102                                "case peer didn't initiate bulk transfer\n",
1103                                reply->msg_body.reply.status,
1104                                libcfs_id2str(rpc->crpc_dest));
1105                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1106                 }
1107
1108                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1109         }
1110         case SWI_STATE_REPLY_RECEIVED:
1111                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1112
1113                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1114
1115                 /* Bulk buffer was unlinked due to remote error. Clear error
1116                  * since reply buffer still contains valid data.
1117                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1118                  * remote error. */
1119                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1120                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1121                         rc = 0;
1122
1123                 wi->swi_state = SWI_STATE_DONE;
1124                 srpc_client_rpc_done(rpc, rc);
1125                 return 1;
1126         }
1127
1128         if (rc != 0) {
1129                 cfs_spin_lock(&rpc->crpc_lock);
1130                 srpc_abort_rpc(rpc, rc);
1131                 cfs_spin_unlock(&rpc->crpc_lock);
1132         }
1133
1134 abort:
1135         if (rpc->crpc_aborted) {
1136                 LNetMDUnlink(rpc->crpc_reqstmdh);
1137                 LNetMDUnlink(rpc->crpc_replymdh);
1138                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1139
1140                 if (!srpc_event_pending(rpc)) {
1141                         srpc_client_rpc_done(rpc, -EINTR);
1142                         return 1;
1143                 }
1144         }
1145         return 0;
1146 }
1147
1148 srpc_client_rpc_t *
1149 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1150                         int nbulkiov, int bulklen,
1151                         void (*rpc_done)(srpc_client_rpc_t *),
1152                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1153 {
1154         srpc_client_rpc_t *rpc;
1155
1156         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1157                                    crpc_bulk.bk_iovs[nbulkiov]));
1158         if (rpc == NULL)
1159                 return NULL;
1160
1161         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1162                              bulklen, rpc_done, rpc_fini, priv);
1163         return rpc;
1164 }
1165
1166 /* called with rpc->crpc_lock held */
1167 void
1168 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1169 {
1170         LASSERT (why != 0);
1171
1172         if (rpc->crpc_aborted || /* already aborted */
1173             rpc->crpc_closed)    /* callback imminent */
1174                 return;
1175
1176         CDEBUG (D_NET,
1177                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1178                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1179                 swi_state2str(rpc->crpc_wi.swi_state), why);
1180
1181         rpc->crpc_aborted = 1;
1182         rpc->crpc_status  = why;
1183         swi_schedule_workitem(&rpc->crpc_wi);
1184         return;
1185 }
1186
1187 /* called with rpc->crpc_lock held */
1188 void
1189 srpc_post_rpc (srpc_client_rpc_t *rpc)
1190 {
1191         LASSERT (!rpc->crpc_aborted);
1192         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1193         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1194
1195         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1196                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1197                 rpc->crpc_timeout);
1198
1199         srpc_add_client_rpc_timer(rpc);
1200         swi_schedule_workitem(&rpc->crpc_wi);
1201         return;
1202 }
1203
1204
1205 int
1206 srpc_send_reply (srpc_server_rpc_t *rpc)
1207 {
1208         srpc_event_t   *ev = &rpc->srpc_ev;
1209         srpc_msg_t     *msg = &rpc->srpc_replymsg;
1210         srpc_buffer_t  *buffer = rpc->srpc_reqstbuf;
1211         srpc_service_t *sv = rpc->srpc_service;
1212         __u64           rpyid;
1213         int             rc;
1214
1215         LASSERT (buffer != NULL);
1216         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1217
1218         cfs_spin_lock(&sv->sv_lock);
1219
1220         if (!sv->sv_shuttingdown &&
1221             sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1222                 /* Repost buffer before replying since test client
1223                  * might send me another RPC once it gets the reply */
1224                 if (srpc_service_post_buffer(sv, buffer) != 0)
1225                         CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1226                 rpc->srpc_reqstbuf = NULL;
1227         }
1228
1229         cfs_spin_unlock(&sv->sv_lock);
1230
1231         ev->ev_fired = 0;
1232         ev->ev_data  = rpc;
1233         ev->ev_type  = SRPC_REPLY_SENT;
1234
1235         msg->msg_magic   = SRPC_MSG_MAGIC;
1236         msg->msg_version = SRPC_MSG_VERSION;
1237         msg->msg_type    = srpc_service2reply(sv->sv_id);
1238
1239         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1240                                    sizeof(*msg), LNET_MD_OP_PUT,
1241                                    rpc->srpc_peer, rpc->srpc_self,
1242                                    &rpc->srpc_replymdh, ev);
1243         if (rc != 0)
1244                 ev->ev_fired = 1;  /* no more event expected */
1245         return rc;
1246 }
1247
1248 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1249 void
1250 srpc_lnet_ev_handler (lnet_event_t *ev)
1251 {
1252         srpc_event_t      *rpcev = ev->md.user_ptr;
1253         srpc_client_rpc_t *crpc;
1254         srpc_server_rpc_t *srpc;
1255         srpc_buffer_t     *buffer;
1256         srpc_service_t    *sv;
1257         srpc_msg_t        *msg;
1258         srpc_msg_type_t    type;
1259
1260         LASSERT (!cfs_in_interrupt());
1261
1262         if (ev->status != 0) {
1263                 cfs_spin_lock(&srpc_data.rpc_glock);
1264                 srpc_data.rpc_counters.errors++;
1265                 cfs_spin_unlock(&srpc_data.rpc_glock);
1266         }
1267
1268         rpcev->ev_lnet = ev->type;
1269
1270         switch (rpcev->ev_type) {
1271         default:
1272                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1273                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1274                 LBUG ();
1275         case SRPC_REQUEST_SENT:
1276                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1277                         cfs_spin_lock(&srpc_data.rpc_glock);
1278                         srpc_data.rpc_counters.rpcs_sent++;
1279                         cfs_spin_unlock(&srpc_data.rpc_glock);
1280                 }
1281         case SRPC_REPLY_RCVD:
1282         case SRPC_BULK_REQ_RCVD:
1283                 crpc = rpcev->ev_data;
1284
1285                 if (rpcev != &crpc->crpc_reqstev &&
1286                     rpcev != &crpc->crpc_replyev &&
1287                     rpcev != &crpc->crpc_bulkev) {
1288                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1289                                rpcev, crpc, &crpc->crpc_reqstev,
1290                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1291                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1292                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1293                         LBUG ();
1294                 }
1295
1296                 cfs_spin_lock(&crpc->crpc_lock);
1297
1298                 LASSERT (rpcev->ev_fired == 0);
1299                 rpcev->ev_fired  = 1;
1300                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1301                                                 -EINTR : ev->status;
1302                 swi_schedule_workitem(&crpc->crpc_wi);
1303
1304                 cfs_spin_unlock(&crpc->crpc_lock);
1305                 break;
1306
1307         case SRPC_REQUEST_RCVD:
1308                 sv = rpcev->ev_data;
1309
1310                 LASSERT (rpcev == &sv->sv_ev);
1311
1312                 cfs_spin_lock(&sv->sv_lock);
1313
1314                 LASSERT (ev->unlinked);
1315                 LASSERT (ev->type == LNET_EVENT_PUT ||
1316                          ev->type == LNET_EVENT_UNLINK);
1317                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1318                          sv->sv_shuttingdown);
1319
1320                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1321                 buffer->buf_peer = ev->initiator;
1322                 buffer->buf_self = ev->target.nid;
1323
1324                 sv->sv_nposted_msg--;
1325                 LASSERT (sv->sv_nposted_msg >= 0);
1326
1327                 if (sv->sv_shuttingdown) {
1328                         /* Leave buffer on sv->sv_posted_msgq since
1329                          * srpc_finish_service needs to traverse it. */
1330                         cfs_spin_unlock(&sv->sv_lock);
1331                         break;
1332                 }
1333
1334                 cfs_list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1335                 msg = &buffer->buf_msg;
1336                 type = srpc_service2request(sv->sv_id);
1337
1338                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1339                     (msg->msg_type != type &&
1340                      msg->msg_type != __swab32(type)) ||
1341                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1342                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1343                         CERROR ("Dropping RPC (%s) from %s: "
1344                                 "status %d mlength %d type %u magic %u.\n",
1345                                 sv->sv_name, libcfs_id2str(ev->initiator),
1346                                 ev->status, ev->mlength,
1347                                 msg->msg_type, msg->msg_magic);
1348
1349                         /* NB can't call srpc_service_recycle_buffer here since
1350                          * it may call LNetM[DE]Attach. The invalid magic tells
1351                          * srpc_handle_rpc to drop this RPC */
1352                         msg->msg_magic = 0;
1353                 }
1354
1355                 if (!cfs_list_empty(&sv->sv_free_rpcq)) {
1356                         srpc = cfs_list_entry(sv->sv_free_rpcq.next,
1357                                               srpc_server_rpc_t, srpc_list);
1358                         cfs_list_del(&srpc->srpc_list);
1359
1360                         srpc_init_server_rpc(srpc, sv, buffer);
1361                         cfs_list_add_tail(&srpc->srpc_list,
1362                                           &sv->sv_active_rpcq);
1363                         srpc_schedule_server_rpc(srpc);
1364                 } else {
1365                         cfs_list_add_tail(&buffer->buf_list,
1366                                           &sv->sv_blocked_msgq);
1367                 }
1368
1369                 cfs_spin_unlock(&sv->sv_lock);
1370
1371                 cfs_spin_lock(&srpc_data.rpc_glock);
1372                 srpc_data.rpc_counters.rpcs_rcvd++;
1373                 cfs_spin_unlock(&srpc_data.rpc_glock);
1374                 break;
1375
1376         case SRPC_BULK_GET_RPLD:
1377                 LASSERT (ev->type == LNET_EVENT_SEND ||
1378                          ev->type == LNET_EVENT_REPLY ||
1379                          ev->type == LNET_EVENT_UNLINK);
1380
1381                 if (!ev->unlinked)
1382                         break; /* wait for final event */
1383
1384         case SRPC_BULK_PUT_SENT:
1385                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1386                         cfs_spin_lock(&srpc_data.rpc_glock);
1387
1388                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1389                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1390                         else
1391                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1392
1393                         cfs_spin_unlock(&srpc_data.rpc_glock);
1394                 }
1395         case SRPC_REPLY_SENT:
1396                 srpc = rpcev->ev_data;
1397                 sv = srpc->srpc_service;
1398
1399                 LASSERT (rpcev == &srpc->srpc_ev);
1400
1401                 cfs_spin_lock(&sv->sv_lock);
1402
1403                 rpcev->ev_fired  = 1;
1404                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1405                                                 -EINTR : ev->status;
1406                 srpc_schedule_server_rpc(srpc);
1407
1408                 cfs_spin_unlock(&sv->sv_lock);
1409                 break;
1410         }
1411
1412         return;
1413 }
1414
1415 #ifndef __KERNEL__
1416
1417 int
1418 srpc_check_event (int timeout)
1419 {
1420         lnet_event_t ev;
1421         int          rc;
1422         int          i;
1423
1424         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1425                         timeout * 1000, &ev, &i);
1426         if (rc == 0) return 0;
1427
1428         LASSERT (rc == -EOVERFLOW || rc == 1);
1429
1430         /* We can't affort to miss any events... */
1431         if (rc == -EOVERFLOW) {
1432                 CERROR ("Dropped an event!!!\n");
1433                 abort();
1434         }
1435
1436         srpc_lnet_ev_handler(&ev);
1437         return 1;
1438 }
1439
1440 #endif
1441
1442 int
1443 srpc_startup (void)
1444 {
1445         int rc;
1446
1447         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1448         cfs_spin_lock_init(&srpc_data.rpc_glock);
1449
1450         /* 1 second pause to avoid timestamp reuse */
1451         cfs_pause(cfs_time_seconds(1));
1452         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1453
1454         srpc_data.rpc_state = SRPC_STATE_NONE;
1455
1456 #ifdef __KERNEL__
1457         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1458 #else
1459         if (the_lnet.ln_server_mode_flag)
1460                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1461         else
1462                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1463 #endif
1464         if (rc < 0) {
1465                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1466                 return rc;
1467         }
1468
1469         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1470
1471         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1472 #ifdef __KERNEL__
1473         rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1474 #else
1475         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1476 #endif
1477         if (rc != 0) {
1478                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1479                 goto bail;
1480         }
1481
1482         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1483         LASSERT (rc == 0);
1484
1485         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1486
1487         rc = stt_startup();
1488
1489 bail:
1490         if (rc != 0)
1491                 srpc_shutdown();
1492         else
1493                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1494
1495         return rc;
1496 }
1497
1498 void
1499 srpc_shutdown (void)
1500 {
1501         int i;
1502         int rc;
1503         int state;
1504
1505         state = srpc_data.rpc_state;
1506         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1507
1508         switch (state) {
1509         default:
1510                 LBUG ();
1511         case SRPC_STATE_RUNNING:
1512                 cfs_spin_lock(&srpc_data.rpc_glock);
1513
1514                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1515                         srpc_service_t *sv = srpc_data.rpc_services[i];
1516
1517                         LASSERTF (sv == NULL,
1518                                   "service not empty: id %d, name %s\n",
1519                                   i, sv->sv_name);
1520                 }
1521
1522                 cfs_spin_unlock(&srpc_data.rpc_glock);
1523
1524                 stt_shutdown();
1525
1526         case SRPC_STATE_EQ_INIT:
1527                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1528                 LASSERT (rc == 0);
1529                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1530                 LASSERT (rc == 0); /* the EQ should have no user by now */
1531
1532         case SRPC_STATE_NI_INIT:
1533                 LNetNIFini();
1534                 break;
1535         }
1536
1537         return;
1538 }