drivers/staging/lustre/lnet/selftest/rpc.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  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  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  41  * - percpt data for service to improve smp performance
  42  * - code cleanup
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_LNET
  46
  47 #include "selftest.h"
  48
  49 typedef enum {
  50         SRPC_STATE_NONE,
  51         SRPC_STATE_NI_INIT,
  52         SRPC_STATE_EQ_INIT,
  53         SRPC_STATE_RUNNING,
  54         SRPC_STATE_STOPPING,
  55 } srpc_state_t;
  56
  57 static struct smoketest_rpc {
  58         spinlock_t       rpc_glock;     /* global lock */
  59         srpc_service_t  *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  60         lnet_handle_eq_t rpc_lnet_eq;   /* _the_ LNet event queue */
  61         srpc_state_t     rpc_state;
  62         srpc_counters_t  rpc_counters;
  63         __u64            rpc_matchbits; /* matchbits counter */
  64 } srpc_data;
  65
  66 static inline int
  67 srpc_serv_portal(int svc_id)
  68 {
  69         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  70                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  71 }
  72
  73 /* forward ref's */
  74 int srpc_handle_rpc(swi_workitem_t *wi);
  75
  76 void srpc_get_counters(srpc_counters_t *cnt)
  77 {
  78         spin_lock(&srpc_data.rpc_glock);
  79         *cnt = srpc_data.rpc_counters;
  80         spin_unlock(&srpc_data.rpc_glock);
  81 }
  82
  83 void srpc_set_counters(const srpc_counters_t *cnt)
  84 {
  85         spin_lock(&srpc_data.rpc_glock);
  86         srpc_data.rpc_counters = *cnt;
  87         spin_unlock(&srpc_data.rpc_glock);
  88 }
  89
  90 static int
  91 srpc_add_bulk_page(srpc_bulk_t *bk, struct page *pg, int i, int nob)
  92 {
  93         nob = min_t(int, nob, PAGE_CACHE_SIZE);
  94
  95         LASSERT(nob > 0);
  96         LASSERT(i >= 0 && i < bk->bk_niov);
  97
  98         bk->bk_iovs[i].kiov_offset = 0;
  99         bk->bk_iovs[i].kiov_page   = pg;
 100         bk->bk_iovs[i].kiov_len    = nob;
 101         return nob;
 102 }
 103
 104 void
 105 srpc_free_bulk(srpc_bulk_t *bk)
 106 {
 107         int i;
 108         struct page *pg;
 109
 110         LASSERT(bk);
 111
 112         for (i = 0; i < bk->bk_niov; i++) {
 113                 pg = bk->bk_iovs[i].kiov_page;
 114                 if (!pg)
 115                         break;
 116
 117                 __free_page(pg);
 118         }
 119
 120         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 121         return;
 122 }
 123
 124 srpc_bulk_t *
 125 srpc_alloc_bulk(int cpt, unsigned bulk_npg, unsigned bulk_len, int sink)
 126 {
 127         srpc_bulk_t *bk;
 128         int i;
 129
 130         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 131
 132         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 133                          offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 134         if (!bk) {
 135                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 136                 return NULL;
 137         }
 138
 139         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 140         bk->bk_sink   = sink;
 141         bk->bk_len    = bulk_len;
 142         bk->bk_niov   = bulk_npg;
 143
 144         for (i = 0; i < bulk_npg; i++) {
 145                 struct page *pg;
 146                 int nob;
 147
 148                 pg = alloc_pages_node(cfs_cpt_spread_node(lnet_cpt_table(), cpt),
 149                                       GFP_KERNEL, 0);
 150                 if (!pg) {
 151                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 152                         srpc_free_bulk(bk);
 153                         return NULL;
 154                 }
 155
 156                 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
 157                 bulk_len -= nob;
 158         }
 159
 160         return bk;
 161 }
 162
 163 static inline __u64
 164 srpc_next_id(void)
 165 {
 166         __u64 id;
 167
 168         spin_lock(&srpc_data.rpc_glock);
 169         id = srpc_data.rpc_matchbits++;
 170         spin_unlock(&srpc_data.rpc_glock);
 171         return id;
 172 }
 173
 174 static void
 175 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 176                      struct srpc_service_cd *scd,
 177                      struct srpc_buffer *buffer)
 178 {
 179         memset(rpc, 0, sizeof(*rpc));
 180         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
 181                           srpc_serv_is_framework(scd->scd_svc) ?
 182                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 183
 184         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 185
 186         rpc->srpc_scd      = scd;
 187         rpc->srpc_reqstbuf = buffer;
 188         rpc->srpc_peer     = buffer->buf_peer;
 189         rpc->srpc_self     = buffer->buf_self;
 190         LNetInvalidateHandle(&rpc->srpc_replymdh);
 191 }
 192
 193 static void
 194 srpc_service_fini(struct srpc_service *svc)
 195 {
 196         struct srpc_service_cd *scd;
 197         struct srpc_server_rpc *rpc;
 198         struct srpc_buffer *buf;
 199         struct list_head *q;
 200         int i;
 201
 202         if (!svc->sv_cpt_data)
 203                 return;
 204
 205         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 206                 while (1) {
 207                         if (!list_empty(&scd->scd_buf_posted))
 208                                 q = &scd->scd_buf_posted;
 209                         else if (!list_empty(&scd->scd_buf_blocked))
 210                                 q = &scd->scd_buf_blocked;
 211                         else
 212                                 break;
 213
 214                         while (!list_empty(q)) {
 215                                 buf = list_entry(q->next, struct srpc_buffer,
 216                                                  buf_list);
 217                                 list_del(&buf->buf_list);
 218                                 LIBCFS_FREE(buf, sizeof(*buf));
 219                         }
 220                 }
 221
 222                 LASSERT(list_empty(&scd->scd_rpc_active));
 223
 224                 while (!list_empty(&scd->scd_rpc_free)) {
 225                         rpc = list_entry(scd->scd_rpc_free.next,
 226                                          struct srpc_server_rpc,
 227                                          srpc_list);
 228                         list_del(&rpc->srpc_list);
 229                         LIBCFS_FREE(rpc, sizeof(*rpc));
 230                 }
 231         }
 232
 233         cfs_percpt_free(svc->sv_cpt_data);
 234         svc->sv_cpt_data = NULL;
 235 }
 236
 237 static int
 238 srpc_service_nrpcs(struct srpc_service *svc)
 239 {
 240         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 241
 242         return srpc_serv_is_framework(svc) ?
 243                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 244 }
 245
 246 int srpc_add_buffer(struct swi_workitem *wi);
 247
 248 static int
 249 srpc_service_init(struct srpc_service *svc)
 250 {
 251         struct srpc_service_cd *scd;
 252         struct srpc_server_rpc *rpc;
 253         int nrpcs;
 254         int i;
 255         int j;
 256
 257         svc->sv_shuttingdown = 0;
 258
 259         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 260                                             sizeof(struct srpc_service_cd));
 261         if (!svc->sv_cpt_data)
 262                 return -ENOMEM;
 263
 264         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 265                         1 : cfs_cpt_number(lnet_cpt_table());
 266         nrpcs = srpc_service_nrpcs(svc);
 267
 268         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 269                 scd->scd_cpt = i;
 270                 scd->scd_svc = svc;
 271                 spin_lock_init(&scd->scd_lock);
 272                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 273                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 274                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 275                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 276
 277                 scd->scd_ev.ev_data = scd;
 278                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 279
 280                 /*
 281                  * NB: don't use lst_sched_serial for adding buffer,
 282                  * see details in srpc_service_add_buffers()
 283                  */
 284                 swi_init_workitem(&scd->scd_buf_wi, scd,
 285                                   srpc_add_buffer, lst_sched_test[i]);
 286
 287                 if (i && srpc_serv_is_framework(svc)) {
 288                         /*
 289                          * NB: framework service only needs srpc_service_cd for
 290                          * one partition, but we allocate for all to make
 291                          * it easier to implement, it will waste a little
 292                          * memory but nobody should care about this
 293                          */
 294                         continue;
 295                 }
 296
 297                 for (j = 0; j < nrpcs; j++) {
 298                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 299                                          i, sizeof(*rpc));
 300                         if (!rpc) {
 301                                 srpc_service_fini(svc);
 302                                 return -ENOMEM;
 303                         }
 304                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 305                 }
 306         }
 307
 308         return 0;
 309 }
 310
 311 int
 312 srpc_add_service(struct srpc_service *sv)
 313 {
 314         int id = sv->sv_id;
 315
 316         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 317
 318         if (srpc_service_init(sv))
 319                 return -ENOMEM;
 320
 321         spin_lock(&srpc_data.rpc_glock);
 322
 323         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 324
 325         if (srpc_data.rpc_services[id]) {
 326                 spin_unlock(&srpc_data.rpc_glock);
 327                 goto failed;
 328         }
 329
 330         srpc_data.rpc_services[id] = sv;
 331         spin_unlock(&srpc_data.rpc_glock);
 332
 333         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 334         return 0;
 335
 336  failed:
 337         srpc_service_fini(sv);
 338         return -EBUSY;
 339 }
 340
 341 int
 342 srpc_remove_service(srpc_service_t *sv)
 343 {
 344         int id = sv->sv_id;
 345
 346         spin_lock(&srpc_data.rpc_glock);
 347
 348         if (srpc_data.rpc_services[id] != sv) {
 349                 spin_unlock(&srpc_data.rpc_glock);
 350                 return -ENOENT;
 351         }
 352
 353         srpc_data.rpc_services[id] = NULL;
 354         spin_unlock(&srpc_data.rpc_glock);
 355         return 0;
 356 }
 357
 358 static int
 359 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 360                        int len, int options, lnet_process_id_t peer,
 361                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 362 {
 363         int rc;
 364         lnet_md_t md;
 365         lnet_handle_me_t meh;
 366
 367         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 368                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 369         if (rc) {
 370                 CERROR("LNetMEAttach failed: %d\n", rc);
 371                 LASSERT(rc == -ENOMEM);
 372                 return -ENOMEM;
 373         }
 374
 375         md.threshold = 1;
 376         md.user_ptr  = ev;
 377         md.start     = buf;
 378         md.length    = len;
 379         md.options   = options;
 380         md.eq_handle = srpc_data.rpc_lnet_eq;
 381
 382         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 383         if (rc) {
 384                 CERROR("LNetMDAttach failed: %d\n", rc);
 385                 LASSERT(rc == -ENOMEM);
 386
 387                 rc = LNetMEUnlink(meh);
 388                 LASSERT(!rc);
 389                 return -ENOMEM;
 390         }
 391
 392         CDEBUG(D_NET, "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 393                libcfs_id2str(peer), portal, matchbits);
 394         return 0;
 395 }
 396
 397 static int
 398 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 399                       int options, lnet_process_id_t peer, lnet_nid_t self,
 400                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 401 {
 402         int rc;
 403         lnet_md_t md;
 404
 405         md.user_ptr  = ev;
 406         md.start     = buf;
 407         md.length    = len;
 408         md.eq_handle = srpc_data.rpc_lnet_eq;
 409         md.threshold = options & LNET_MD_OP_GET ? 2 : 1;
 410         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 411
 412         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 413         if (rc) {
 414                 CERROR("LNetMDBind failed: %d\n", rc);
 415                 LASSERT(rc == -ENOMEM);
 416                 return -ENOMEM;
 417         }
 418
 419         /*
 420          * this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 421          * they're only meaningful for MDs attached to an ME (i.e. passive
 422          * buffers...
 423          */
 424         if (options & LNET_MD_OP_PUT) {
 425                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 426                              portal, matchbits, 0, 0);
 427         } else {
 428                 LASSERT(options & LNET_MD_OP_GET);
 429
 430                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 431         }
 432
 433         if (rc) {
 434                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 435                        options & LNET_MD_OP_PUT ? "Put" : "Get",
 436                        libcfs_id2str(peer), portal, matchbits, rc);
 437
 438                 /*
 439                  * The forthcoming unlink event will complete this operation
 440                  * with failure, so fall through and return success here.
 441                  */
 442                 rc = LNetMDUnlink(*mdh);
 443                 LASSERT(!rc);
 444         } else {
 445                 CDEBUG(D_NET, "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 446                        libcfs_id2str(peer), portal, matchbits);
 447         }
 448         return 0;
 449 }
 450
 451 static int
 452 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 453                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 454 {
 455         lnet_process_id_t any = {0};
 456
 457         any.nid = LNET_NID_ANY;
 458         any.pid = LNET_PID_ANY;
 459
 460         return srpc_post_passive_rdma(srpc_serv_portal(service),
 461                                       local, service, buf, len,
 462                                       LNET_MD_OP_PUT, any, mdh, ev);
 463 }
 464
 465 static int
 466 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 467         __must_hold(&scd->scd_lock)
 468 {
 469         struct srpc_service *sv = scd->scd_svc;
 470         struct srpc_msg *msg = &buf->buf_msg;
 471         int rc;
 472
 473         LNetInvalidateHandle(&buf->buf_mdh);
 474         list_add(&buf->buf_list, &scd->scd_buf_posted);
 475         scd->scd_buf_nposted++;
 476         spin_unlock(&scd->scd_lock);
 477
 478         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 479                                       !srpc_serv_is_framework(sv),
 480                                       msg, sizeof(*msg), &buf->buf_mdh,
 481                                       &scd->scd_ev);
 482
 483         /*
 484          * At this point, a RPC (new or delayed) may have arrived in
 485          * msg and its event handler has been called. So we must add
 486          * buf to scd_buf_posted _before_ dropping scd_lock
 487          */
 488         spin_lock(&scd->scd_lock);
 489
 490         if (!rc) {
 491                 if (!sv->sv_shuttingdown)
 492                         return 0;
 493
 494                 spin_unlock(&scd->scd_lock);
 495                 /*
 496                  * srpc_shutdown_service might have tried to unlink me
 497                  * when my buf_mdh was still invalid
 498                  */
 499                 LNetMDUnlink(buf->buf_mdh);
 500                 spin_lock(&scd->scd_lock);
 501                 return 0;
 502         }
 503
 504         scd->scd_buf_nposted--;
 505         if (sv->sv_shuttingdown)
 506                 return rc; /* don't allow to change scd_buf_posted */
 507
 508         list_del(&buf->buf_list);
 509         spin_unlock(&scd->scd_lock);
 510
 511         LIBCFS_FREE(buf, sizeof(*buf));
 512
 513         spin_lock(&scd->scd_lock);
 514         return rc;
 515 }
 516
 517 int
 518 srpc_add_buffer(struct swi_workitem *wi)
 519 {
 520         struct srpc_service_cd *scd = wi->swi_workitem.wi_data;
 521         struct srpc_buffer *buf;
 522         int rc = 0;
 523
 524         /*
 525          * it's called by workitem scheduler threads, these threads
 526          * should have been set CPT affinity, so buffers will be posted
 527          * on CPT local list of Portal
 528          */
 529         spin_lock(&scd->scd_lock);
 530
 531         while (scd->scd_buf_adjust > 0 &&
 532                !scd->scd_svc->sv_shuttingdown) {
 533                 scd->scd_buf_adjust--; /* consume it */
 534                 scd->scd_buf_posting++;
 535
 536                 spin_unlock(&scd->scd_lock);
 537
 538                 LIBCFS_ALLOC(buf, sizeof(*buf));
 539                 if (!buf) {
 540                         CERROR("Failed to add new buf to service: %s\n",
 541                                scd->scd_svc->sv_name);
 542                         spin_lock(&scd->scd_lock);
 543                         rc = -ENOMEM;
 544                         break;
 545                 }
 546
 547                 spin_lock(&scd->scd_lock);
 548                 if (scd->scd_svc->sv_shuttingdown) {
 549                         spin_unlock(&scd->scd_lock);
 550                         LIBCFS_FREE(buf, sizeof(*buf));
 551
 552                         spin_lock(&scd->scd_lock);
 553                         rc = -ESHUTDOWN;
 554                         break;
 555                 }
 556
 557                 rc = srpc_service_post_buffer(scd, buf);
 558                 if (rc)
 559                         break; /* buf has been freed inside */
 560
 561                 LASSERT(scd->scd_buf_posting > 0);
 562                 scd->scd_buf_posting--;
 563                 scd->scd_buf_total++;
 564                 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
 565         }
 566
 567         if (rc) {
 568                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 569                 scd->scd_buf_err = rc;
 570
 571                 LASSERT(scd->scd_buf_posting > 0);
 572                 scd->scd_buf_posting--;
 573         }
 574
 575         spin_unlock(&scd->scd_lock);
 576         return 0;
 577 }
 578
 579 int
 580 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 581 {
 582         struct srpc_service_cd *scd;
 583         int rc = 0;
 584         int i;
 585
 586         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 587
 588         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 589                 spin_lock(&scd->scd_lock);
 590
 591                 scd->scd_buf_err = 0;
 592                 scd->scd_buf_err_stamp = 0;
 593                 scd->scd_buf_posting = 0;
 594                 scd->scd_buf_adjust = nbuffer;
 595                 /* start to post buffers */
 596                 swi_schedule_workitem(&scd->scd_buf_wi);
 597                 spin_unlock(&scd->scd_lock);
 598
 599                 /* framework service only post buffer for one partition  */
 600                 if (srpc_serv_is_framework(sv))
 601                         break;
 602         }
 603
 604         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 605                 spin_lock(&scd->scd_lock);
 606                 /*
 607                  * NB: srpc_service_add_buffers() can be called inside
 608                  * thread context of lst_sched_serial, and we don't normally
 609                  * allow to sleep inside thread context of WI scheduler
 610                  * because it will block current scheduler thread from doing
 611                  * anything else, even worse, it could deadlock if it's
 612                  * waiting on result from another WI of the same scheduler.
 613                  * However, it's safe at here because scd_buf_wi is scheduled
 614                  * by thread in a different WI scheduler (lst_sched_test),
 615                  * so we don't have any risk of deadlock, though this could
 616                  * block all WIs pending on lst_sched_serial for a moment
 617                  * which is not good but not fatal.
 618                  */
 619                 lst_wait_until(scd->scd_buf_err ||
 620                                (!scd->scd_buf_adjust &&
 621                                 !scd->scd_buf_posting),
 622                                scd->scd_lock, "waiting for adding buffer\n");
 623
 624                 if (scd->scd_buf_err && !rc)
 625                         rc = scd->scd_buf_err;
 626
 627                 spin_unlock(&scd->scd_lock);
 628         }
 629
 630         return rc;
 631 }
 632
 633 void
 634 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 635 {
 636         struct srpc_service_cd *scd;
 637         int num;
 638         int i;
 639
 640         LASSERT(!sv->sv_shuttingdown);
 641
 642         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 643                 spin_lock(&scd->scd_lock);
 644
 645                 num = scd->scd_buf_total + scd->scd_buf_posting;
 646                 scd->scd_buf_adjust -= min(nbuffer, num);
 647
 648                 spin_unlock(&scd->scd_lock);
 649         }
 650 }
 651
 652 /* returns 1 if sv has finished, otherwise 0 */
 653 int
 654 srpc_finish_service(struct srpc_service *sv)
 655 {
 656         struct srpc_service_cd *scd;
 657         struct srpc_server_rpc *rpc;
 658         int i;
 659
 660         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 661
 662         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 663                 spin_lock(&scd->scd_lock);
 664                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 665                         spin_unlock(&scd->scd_lock);
 666                         return 0;
 667                 }
 668
 669                 if (scd->scd_buf_nposted > 0) {
 670                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink",
 671                                scd->scd_buf_nposted);
 672                         spin_unlock(&scd->scd_lock);
 673                         return 0;
 674                 }
 675
 676                 if (list_empty(&scd->scd_rpc_active)) {
 677                         spin_unlock(&scd->scd_lock);
 678                         continue;
 679                 }
 680
 681                 rpc = list_entry(scd->scd_rpc_active.next,
 682                                  struct srpc_server_rpc, srpc_list);
 683                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, wi %s scheduled %d running %d, ev fired %d type %d status %d lnet %d\n",
 684                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 685                         swi_state2str(rpc->srpc_wi.swi_state),
 686                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 687                         rpc->srpc_wi.swi_workitem.wi_running,
 688                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 689                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 690                 spin_unlock(&scd->scd_lock);
 691                 return 0;
 692         }
 693
 694         /* no lock needed from now on */
 695         srpc_service_fini(sv);
 696         return 1;
 697 }
 698
 699 /* called with sv->sv_lock held */
 700 static void
 701 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
 702         __must_hold(&scd->scd_lock)
 703 {
 704         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 705                 if (srpc_service_post_buffer(scd, buf)) {
 706                         CWARN("Failed to post %s buffer\n",
 707                               scd->scd_svc->sv_name);
 708                 }
 709                 return;
 710         }
 711
 712         /* service is shutting down, or we want to recycle some buffers */
 713         scd->scd_buf_total--;
 714
 715         if (scd->scd_buf_adjust < 0) {
 716                 scd->scd_buf_adjust++;
 717                 if (scd->scd_buf_adjust < 0 &&
 718                     !scd->scd_buf_total && !scd->scd_buf_posting) {
 719                         CDEBUG(D_INFO,
 720                                "Try to recycle %d buffers but nothing left\n",
 721                                scd->scd_buf_adjust);
 722                         scd->scd_buf_adjust = 0;
 723                 }
 724         }
 725
 726         spin_unlock(&scd->scd_lock);
 727         LIBCFS_FREE(buf, sizeof(*buf));
 728         spin_lock(&scd->scd_lock);
 729 }
 730
 731 void
 732 srpc_abort_service(struct srpc_service *sv)
 733 {
 734         struct srpc_service_cd *scd;
 735         struct srpc_server_rpc *rpc;
 736         int i;
 737
 738         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 739                sv->sv_id, sv->sv_name);
 740
 741         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 742                 spin_lock(&scd->scd_lock);
 743
 744                 /*
 745                  * schedule in-flight RPCs to notice the abort, NB:
 746                  * racing with incoming RPCs; complete fix should make test
 747                  * RPCs carry session ID in its headers
 748                  */
 749                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 750                         rpc->srpc_aborted = 1;
 751                         swi_schedule_workitem(&rpc->srpc_wi);
 752                 }
 753
 754                 spin_unlock(&scd->scd_lock);
 755         }
 756 }
 757
 758 void
 759 srpc_shutdown_service(srpc_service_t *sv)
 760 {
 761         struct srpc_service_cd *scd;
 762         struct srpc_server_rpc *rpc;
 763         srpc_buffer_t *buf;
 764         int i;
 765
 766         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 767                sv->sv_id, sv->sv_name);
 768
 769         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 770                 spin_lock(&scd->scd_lock);
 771
 772         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 773
 774         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 775                 spin_unlock(&scd->scd_lock);
 776
 777         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 778                 spin_lock(&scd->scd_lock);
 779
 780                 /* schedule in-flight RPCs to notice the shutdown */
 781                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 782                         swi_schedule_workitem(&rpc->srpc_wi);
 783
 784                 spin_unlock(&scd->scd_lock);
 785
 786                 /*
 787                  * OK to traverse scd_buf_posted without lock, since no one
 788                  * touches scd_buf_posted now
 789                  */
 790                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 791                         LNetMDUnlink(buf->buf_mdh);
 792         }
 793 }
 794
 795 static int
 796 srpc_send_request(srpc_client_rpc_t *rpc)
 797 {
 798         srpc_event_t *ev = &rpc->crpc_reqstev;
 799         int rc;
 800
 801         ev->ev_fired = 0;
 802         ev->ev_data  = rpc;
 803         ev->ev_type  = SRPC_REQUEST_SENT;
 804
 805          rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 806                                     rpc->crpc_service, &rpc->crpc_reqstmsg,
 807                                     sizeof(srpc_msg_t), LNET_MD_OP_PUT,
 808                                     rpc->crpc_dest, LNET_NID_ANY,
 809                                     &rpc->crpc_reqstmdh, ev);
 810         if (rc) {
 811                 LASSERT(rc == -ENOMEM);
 812                 ev->ev_fired = 1;  /* no more event expected */
 813         }
 814         return rc;
 815 }
 816
 817 static int
 818 srpc_prepare_reply(srpc_client_rpc_t *rpc)
 819 {
 820         srpc_event_t *ev = &rpc->crpc_replyev;
 821         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 822         int rc;
 823
 824         ev->ev_fired = 0;
 825         ev->ev_data  = rpc;
 826         ev->ev_type  = SRPC_REPLY_RCVD;
 827
 828         *id = srpc_next_id();
 829
 830         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 831                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 832                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 833                                     &rpc->crpc_replymdh, ev);
 834         if (rc) {
 835                 LASSERT(rc == -ENOMEM);
 836                 ev->ev_fired = 1;  /* no more event expected */
 837         }
 838         return rc;
 839 }
 840
 841 static int
 842 srpc_prepare_bulk(srpc_client_rpc_t *rpc)
 843 {
 844         srpc_bulk_t *bk = &rpc->crpc_bulk;
 845         srpc_event_t *ev = &rpc->crpc_bulkev;
 846         __u64   *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 847         int rc;
 848         int opt;
 849
 850         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 851
 852         if (!bk->bk_niov)
 853                 return 0; /* nothing to do */
 854
 855         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 856         opt |= LNET_MD_KIOV;
 857
 858         ev->ev_fired = 0;
 859         ev->ev_data  = rpc;
 860         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 861
 862         *id = srpc_next_id();
 863
 864         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 865                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 866                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 867         if (rc) {
 868                 LASSERT(rc == -ENOMEM);
 869                 ev->ev_fired = 1;  /* no more event expected */
 870         }
 871         return rc;
 872 }
 873
 874 static int
 875 srpc_do_bulk(struct srpc_server_rpc *rpc)
 876 {
 877         srpc_event_t *ev = &rpc->srpc_ev;
 878         srpc_bulk_t *bk = rpc->srpc_bulk;
 879         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 880         int rc;
 881         int opt;
 882
 883         LASSERT(bk);
 884
 885         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 886         opt |= LNET_MD_KIOV;
 887
 888         ev->ev_fired = 0;
 889         ev->ev_data  = rpc;
 890         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 891
 892         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 893                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 894                                    rpc->srpc_peer, rpc->srpc_self,
 895                                    &bk->bk_mdh, ev);
 896         if (rc)
 897                 ev->ev_fired = 1;  /* no more event expected */
 898         return rc;
 899 }
 900
 901 /* only called from srpc_handle_rpc */
 902 static void
 903 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 904 {
 905         struct srpc_service_cd *scd = rpc->srpc_scd;
 906         struct srpc_service *sv  = scd->scd_svc;
 907         srpc_buffer_t *buffer;
 908
 909         LASSERT(status || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 910
 911         rpc->srpc_status = status;
 912
 913         CDEBUG_LIMIT(!status ? D_NET : D_NETERROR,
 914                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 915                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 916                      swi_state2str(rpc->srpc_wi.swi_state), status);
 917
 918         if (status) {
 919                 spin_lock(&srpc_data.rpc_glock);
 920                 srpc_data.rpc_counters.rpcs_dropped++;
 921                 spin_unlock(&srpc_data.rpc_glock);
 922         }
 923
 924         if (rpc->srpc_done)
 925                 (*rpc->srpc_done) (rpc);
 926         LASSERT(!rpc->srpc_bulk);
 927
 928         spin_lock(&scd->scd_lock);
 929
 930         if (rpc->srpc_reqstbuf) {
 931                 /*
 932                  * NB might drop sv_lock in srpc_service_recycle_buffer, but
 933                  * sv won't go away for scd_rpc_active must not be empty
 934                  */
 935                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 936                 rpc->srpc_reqstbuf = NULL;
 937         }
 938
 939         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 940
 941         /*
 942          * No one can schedule me now since:
 943          * - I'm not on scd_rpc_active.
 944          * - all LNet events have been fired.
 945          * Cancel pending schedules and prevent future schedule attempts:
 946          */
 947         LASSERT(rpc->srpc_ev.ev_fired);
 948         swi_exit_workitem(&rpc->srpc_wi);
 949
 950         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 951                 buffer = list_entry(scd->scd_buf_blocked.next,
 952                                     srpc_buffer_t, buf_list);
 953                 list_del(&buffer->buf_list);
 954
 955                 srpc_init_server_rpc(rpc, scd, buffer);
 956                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 957                 swi_schedule_workitem(&rpc->srpc_wi);
 958         } else {
 959                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 960         }
 961
 962         spin_unlock(&scd->scd_lock);
 963         return;
 964 }
 965
 966 /* handles an incoming RPC */
 967 int
 968 srpc_handle_rpc(swi_workitem_t *wi)
 969 {
 970         struct srpc_server_rpc *rpc = wi->swi_workitem.wi_data;
 971         struct srpc_service_cd *scd = rpc->srpc_scd;
 972         struct srpc_service *sv = scd->scd_svc;
 973         srpc_event_t *ev = &rpc->srpc_ev;
 974         int rc = 0;
 975
 976         LASSERT(wi == &rpc->srpc_wi);
 977
 978         spin_lock(&scd->scd_lock);
 979
 980         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 981                 spin_unlock(&scd->scd_lock);
 982
 983                 if (rpc->srpc_bulk)
 984                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 985                 LNetMDUnlink(rpc->srpc_replymdh);
 986
 987                 if (ev->ev_fired) { /* no more event, OK to finish */
 988                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 989                         return 1;
 990                 }
 991                 return 0;
 992         }
 993
 994         spin_unlock(&scd->scd_lock);
 995
 996         switch (wi->swi_state) {
 997         default:
 998                 LBUG();
 999         case SWI_STATE_NEWBORN: {
1000                 srpc_msg_t *msg;
1001                 srpc_generic_reply_t *reply;
1002
1003                 msg = &rpc->srpc_reqstbuf->buf_msg;
1004                 reply = &rpc->srpc_replymsg.msg_body.reply;
1005
1006                 if (!msg->msg_magic) {
1007                         /* moaned already in srpc_lnet_ev_handler */
1008                         srpc_server_rpc_done(rpc, EBADMSG);
1009                         return 1;
1010                 }
1011
1012                 srpc_unpack_msg_hdr(msg);
1013                 if (msg->msg_version != SRPC_MSG_VERSION) {
1014                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1015                               msg->msg_version, SRPC_MSG_VERSION,
1016                               libcfs_id2str(rpc->srpc_peer));
1017                         reply->status = EPROTO;
1018                         /* drop through and send reply */
1019                 } else {
1020                         reply->status = 0;
1021                         rc = (*sv->sv_handler)(rpc);
1022                         LASSERT(!reply->status || !rpc->srpc_bulk);
1023                         if (rc) {
1024                                 srpc_server_rpc_done(rpc, rc);
1025                                 return 1;
1026                         }
1027                 }
1028
1029                 wi->swi_state = SWI_STATE_BULK_STARTED;
1030
1031                 if (rpc->srpc_bulk) {
1032                         rc = srpc_do_bulk(rpc);
1033                         if (!rc)
1034                                 return 0; /* wait for bulk */
1035
1036                         LASSERT(ev->ev_fired);
1037                         ev->ev_status = rc;
1038                 }
1039         }
1040         case SWI_STATE_BULK_STARTED:
1041                 LASSERT(!rpc->srpc_bulk || ev->ev_fired);
1042
1043                 if (rpc->srpc_bulk) {
1044                         rc = ev->ev_status;
1045
1046                         if (sv->sv_bulk_ready)
1047                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1048
1049                         if (rc) {
1050                                 srpc_server_rpc_done(rpc, rc);
1051                                 return 1;
1052                         }
1053                 }
1054
1055                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1056                 rc = srpc_send_reply(rpc);
1057                 if (!rc)
1058                         return 0; /* wait for reply */
1059                 srpc_server_rpc_done(rpc, rc);
1060                 return 1;
1061
1062         case SWI_STATE_REPLY_SUBMITTED:
1063                 if (!ev->ev_fired) {
1064                         CERROR("RPC %p: bulk %p, service %d\n",
1065                                rpc, rpc->srpc_bulk, sv->sv_id);
1066                         CERROR("Event: status %d, type %d, lnet %d\n",
1067                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1068                         LASSERT(ev->ev_fired);
1069                 }
1070
1071                 wi->swi_state = SWI_STATE_DONE;
1072                 srpc_server_rpc_done(rpc, ev->ev_status);
1073                 return 1;
1074         }
1075
1076         return 0;
1077 }
1078
1079 static void
1080 srpc_client_rpc_expired(void *data)
1081 {
1082         srpc_client_rpc_t *rpc = data;
1083
1084         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1085               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1086               rpc->crpc_timeout);
1087
1088         spin_lock(&rpc->crpc_lock);
1089
1090         rpc->crpc_timeout = 0;
1091         srpc_abort_rpc(rpc, -ETIMEDOUT);
1092
1093         spin_unlock(&rpc->crpc_lock);
1094
1095         spin_lock(&srpc_data.rpc_glock);
1096         srpc_data.rpc_counters.rpcs_expired++;
1097         spin_unlock(&srpc_data.rpc_glock);
1098 }
1099
1100 inline void
1101 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1102 {
1103         stt_timer_t *timer = &rpc->crpc_timer;
1104
1105         if (!rpc->crpc_timeout)
1106                 return;
1107
1108         INIT_LIST_HEAD(&timer->stt_list);
1109         timer->stt_data    = rpc;
1110         timer->stt_func    = srpc_client_rpc_expired;
1111         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1112         stt_add_timer(timer);
1113         return;
1114 }
1115
1116 /*
1117  * Called with rpc->crpc_lock held.
1118  *
1119  * Upon exit the RPC expiry timer is not queued and the handler is not
1120  * running on any CPU.
1121  */
1122 static void
1123 srpc_del_client_rpc_timer(srpc_client_rpc_t *rpc)
1124 {
1125         /* timer not planted or already exploded */
1126         if (!rpc->crpc_timeout)
1127                 return;
1128
1129         /* timer successfully defused */
1130         if (stt_del_timer(&rpc->crpc_timer))
1131                 return;
1132
1133         /* timer detonated, wait for it to explode */
1134         while (rpc->crpc_timeout) {
1135                 spin_unlock(&rpc->crpc_lock);
1136
1137                 schedule();
1138
1139                 spin_lock(&rpc->crpc_lock);
1140         }
1141 }
1142
1143 static void
1144 srpc_client_rpc_done(srpc_client_rpc_t *rpc, int status)
1145 {
1146         swi_workitem_t *wi = &rpc->crpc_wi;
1147
1148         LASSERT(status || wi->swi_state == SWI_STATE_DONE);
1149
1150         spin_lock(&rpc->crpc_lock);
1151
1152         rpc->crpc_closed = 1;
1153         if (!rpc->crpc_status)
1154                 rpc->crpc_status = status;
1155
1156         srpc_del_client_rpc_timer(rpc);
1157
1158         CDEBUG_LIMIT(!status ? D_NET : D_NETERROR,
1159                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1160                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1161                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1162
1163         /*
1164          * No one can schedule me now since:
1165          * - RPC timer has been defused.
1166          * - all LNet events have been fired.
1167          * - crpc_closed has been set, preventing srpc_abort_rpc from
1168          *   scheduling me.
1169          * Cancel pending schedules and prevent future schedule attempts:
1170          */
1171         LASSERT(!srpc_event_pending(rpc));
1172         swi_exit_workitem(wi);
1173
1174         spin_unlock(&rpc->crpc_lock);
1175
1176         (*rpc->crpc_done)(rpc);
1177         return;
1178 }
1179
1180 /* sends an outgoing RPC */
1181 int
1182 srpc_send_rpc(swi_workitem_t *wi)
1183 {
1184         int rc = 0;
1185         srpc_client_rpc_t *rpc;
1186         srpc_msg_t *reply;
1187         int do_bulk;
1188
1189         LASSERT(wi);
1190
1191         rpc = wi->swi_workitem.wi_data;
1192
1193         LASSERT(rpc);
1194         LASSERT(wi == &rpc->crpc_wi);
1195
1196         reply = &rpc->crpc_replymsg;
1197         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1198
1199         spin_lock(&rpc->crpc_lock);
1200
1201         if (rpc->crpc_aborted) {
1202                 spin_unlock(&rpc->crpc_lock);
1203                 goto abort;
1204         }
1205
1206         spin_unlock(&rpc->crpc_lock);
1207
1208         switch (wi->swi_state) {
1209         default:
1210                 LBUG();
1211         case SWI_STATE_NEWBORN:
1212                 LASSERT(!srpc_event_pending(rpc));
1213
1214                 rc = srpc_prepare_reply(rpc);
1215                 if (rc) {
1216                         srpc_client_rpc_done(rpc, rc);
1217                         return 1;
1218                 }
1219
1220                 rc = srpc_prepare_bulk(rpc);
1221                 if (rc)
1222                         break;
1223
1224                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1225                 rc = srpc_send_request(rpc);
1226                 break;
1227
1228         case SWI_STATE_REQUEST_SUBMITTED:
1229                 /*
1230                  * CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1231                  * order; however, they're processed in a strict order:
1232                  * rqt, rpy, and bulk.
1233                  */
1234                 if (!rpc->crpc_reqstev.ev_fired)
1235                         break;
1236
1237                 rc = rpc->crpc_reqstev.ev_status;
1238                 if (rc)
1239                         break;
1240
1241                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1242                 /* perhaps more events, fall thru */
1243         case SWI_STATE_REQUEST_SENT: {
1244                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1245
1246                 if (!rpc->crpc_replyev.ev_fired)
1247                         break;
1248
1249                 rc = rpc->crpc_replyev.ev_status;
1250                 if (rc)
1251                         break;
1252
1253                 srpc_unpack_msg_hdr(reply);
1254                 if (reply->msg_type != type ||
1255                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1256                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1257                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1258                               libcfs_id2str(rpc->crpc_dest),
1259                               reply->msg_type, type,
1260                               reply->msg_magic, SRPC_MSG_MAGIC);
1261                         rc = -EBADMSG;
1262                         break;
1263                 }
1264
1265                 if (do_bulk && reply->msg_body.reply.status) {
1266                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1267                               reply->msg_body.reply.status,
1268                               libcfs_id2str(rpc->crpc_dest));
1269                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1270                 }
1271
1272                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1273         }
1274         case SWI_STATE_REPLY_RECEIVED:
1275                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1276                         break;
1277
1278                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1279
1280                 /*
1281                  * Bulk buffer was unlinked due to remote error. Clear error
1282                  * since reply buffer still contains valid data.
1283                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1284                  * remote error.
1285                  */
1286                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1287                     !rpc->crpc_status && reply->msg_body.reply.status)
1288                         rc = 0;
1289
1290                 wi->swi_state = SWI_STATE_DONE;
1291                 srpc_client_rpc_done(rpc, rc);
1292                 return 1;
1293         }
1294
1295         if (rc) {
1296                 spin_lock(&rpc->crpc_lock);
1297                 srpc_abort_rpc(rpc, rc);
1298                 spin_unlock(&rpc->crpc_lock);
1299         }
1300
1301 abort:
1302         if (rpc->crpc_aborted) {
1303                 LNetMDUnlink(rpc->crpc_reqstmdh);
1304                 LNetMDUnlink(rpc->crpc_replymdh);
1305                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1306
1307                 if (!srpc_event_pending(rpc)) {
1308                         srpc_client_rpc_done(rpc, -EINTR);
1309                         return 1;
1310                 }
1311         }
1312         return 0;
1313 }
1314
1315 srpc_client_rpc_t *
1316 srpc_create_client_rpc(lnet_process_id_t peer, int service,
1317                        int nbulkiov, int bulklen,
1318                        void (*rpc_done)(srpc_client_rpc_t *),
1319                        void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1320 {
1321         srpc_client_rpc_t *rpc;
1322
1323         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1324                                    crpc_bulk.bk_iovs[nbulkiov]));
1325         if (!rpc)
1326                 return NULL;
1327
1328         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1329                              bulklen, rpc_done, rpc_fini, priv);
1330         return rpc;
1331 }
1332
1333 /* called with rpc->crpc_lock held */
1334 void
1335 srpc_abort_rpc(srpc_client_rpc_t *rpc, int why)
1336 {
1337         LASSERT(why);
1338
1339         if (rpc->crpc_aborted || /* already aborted */
1340             rpc->crpc_closed)    /* callback imminent */
1341                 return;
1342
1343         CDEBUG(D_NET, "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1344                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1345                swi_state2str(rpc->crpc_wi.swi_state), why);
1346
1347         rpc->crpc_aborted = 1;
1348         rpc->crpc_status  = why;
1349         swi_schedule_workitem(&rpc->crpc_wi);
1350         return;
1351 }
1352
1353 /* called with rpc->crpc_lock held */
1354 void
1355 srpc_post_rpc(srpc_client_rpc_t *rpc)
1356 {
1357         LASSERT(!rpc->crpc_aborted);
1358         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1359
1360         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1361                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1362                rpc->crpc_timeout);
1363
1364         srpc_add_client_rpc_timer(rpc);
1365         swi_schedule_workitem(&rpc->crpc_wi);
1366         return;
1367 }
1368
1369 int
1370 srpc_send_reply(struct srpc_server_rpc *rpc)
1371 {
1372         srpc_event_t *ev = &rpc->srpc_ev;
1373         struct srpc_msg *msg = &rpc->srpc_replymsg;
1374         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1375         struct srpc_service_cd *scd = rpc->srpc_scd;
1376         struct srpc_service *sv = scd->scd_svc;
1377         __u64 rpyid;
1378         int rc;
1379
1380         LASSERT(buffer);
1381         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1382
1383         spin_lock(&scd->scd_lock);
1384
1385         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1386                 /*
1387                  * Repost buffer before replying since test client
1388                  * might send me another RPC once it gets the reply
1389                  */
1390                 if (srpc_service_post_buffer(scd, buffer))
1391                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1392                 rpc->srpc_reqstbuf = NULL;
1393         }
1394
1395         spin_unlock(&scd->scd_lock);
1396
1397         ev->ev_fired = 0;
1398         ev->ev_data  = rpc;
1399         ev->ev_type  = SRPC_REPLY_SENT;
1400
1401         msg->msg_magic   = SRPC_MSG_MAGIC;
1402         msg->msg_version = SRPC_MSG_VERSION;
1403         msg->msg_type    = srpc_service2reply(sv->sv_id);
1404
1405         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1406                                    sizeof(*msg), LNET_MD_OP_PUT,
1407                                    rpc->srpc_peer, rpc->srpc_self,
1408                                    &rpc->srpc_replymdh, ev);
1409         if (rc)
1410                 ev->ev_fired = 1;  /* no more event expected */
1411         return rc;
1412 }
1413
1414 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1415 static void
1416 srpc_lnet_ev_handler(lnet_event_t *ev)
1417 {
1418         struct srpc_service_cd *scd;
1419         srpc_event_t *rpcev = ev->md.user_ptr;
1420         srpc_client_rpc_t *crpc;
1421         struct srpc_server_rpc *srpc;
1422         srpc_buffer_t *buffer;
1423         srpc_service_t *sv;
1424         srpc_msg_t *msg;
1425         srpc_msg_type_t type;
1426
1427         LASSERT(!in_interrupt());
1428
1429         if (ev->status) {
1430                 spin_lock(&srpc_data.rpc_glock);
1431                 srpc_data.rpc_counters.errors++;
1432                 spin_unlock(&srpc_data.rpc_glock);
1433         }
1434
1435         rpcev->ev_lnet = ev->type;
1436
1437         switch (rpcev->ev_type) {
1438         default:
1439                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1440                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1441                 LBUG();
1442         case SRPC_REQUEST_SENT:
1443                 if (!ev->status && ev->type != LNET_EVENT_UNLINK) {
1444                         spin_lock(&srpc_data.rpc_glock);
1445                         srpc_data.rpc_counters.rpcs_sent++;
1446                         spin_unlock(&srpc_data.rpc_glock);
1447                 }
1448         case SRPC_REPLY_RCVD:
1449         case SRPC_BULK_REQ_RCVD:
1450                 crpc = rpcev->ev_data;
1451
1452                 if (rpcev != &crpc->crpc_reqstev &&
1453                     rpcev != &crpc->crpc_replyev &&
1454                     rpcev != &crpc->crpc_bulkev) {
1455                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1456                                rpcev, crpc, &crpc->crpc_reqstev,
1457                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1458                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1459                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1460                         LBUG();
1461                 }
1462
1463                 spin_lock(&crpc->crpc_lock);
1464
1465                 LASSERT(!rpcev->ev_fired);
1466                 rpcev->ev_fired  = 1;
1467                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1468                                                 -EINTR : ev->status;
1469                 swi_schedule_workitem(&crpc->crpc_wi);
1470
1471                 spin_unlock(&crpc->crpc_lock);
1472                 break;
1473
1474         case SRPC_REQUEST_RCVD:
1475                 scd = rpcev->ev_data;
1476                 sv = scd->scd_svc;
1477
1478                 LASSERT(rpcev == &scd->scd_ev);
1479
1480                 spin_lock(&scd->scd_lock);
1481
1482                 LASSERT(ev->unlinked);
1483                 LASSERT(ev->type == LNET_EVENT_PUT ||
1484                         ev->type == LNET_EVENT_UNLINK);
1485                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1486                         sv->sv_shuttingdown);
1487
1488                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1489                 buffer->buf_peer = ev->initiator;
1490                 buffer->buf_self = ev->target.nid;
1491
1492                 LASSERT(scd->scd_buf_nposted > 0);
1493                 scd->scd_buf_nposted--;
1494
1495                 if (sv->sv_shuttingdown) {
1496                         /*
1497                          * Leave buffer on scd->scd_buf_nposted since
1498                          * srpc_finish_service needs to traverse it.
1499                          */
1500                         spin_unlock(&scd->scd_lock);
1501                         break;
1502                 }
1503
1504                 if (scd->scd_buf_err_stamp &&
1505                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1506                         /* re-enable adding buffer */
1507                         scd->scd_buf_err_stamp = 0;
1508                         scd->scd_buf_err = 0;
1509                 }
1510
1511                 if (!scd->scd_buf_err && /* adding buffer is enabled */
1512                     !scd->scd_buf_adjust &&
1513                     scd->scd_buf_nposted < scd->scd_buf_low) {
1514                         scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1515                                                   SFW_TEST_WI_MIN);
1516                         swi_schedule_workitem(&scd->scd_buf_wi);
1517                 }
1518
1519                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1520                 msg = &buffer->buf_msg;
1521                 type = srpc_service2request(sv->sv_id);
1522
1523                 if (ev->status || ev->mlength != sizeof(*msg) ||
1524                     (msg->msg_type != type &&
1525                      msg->msg_type != __swab32(type)) ||
1526                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1527                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1528                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1529                                sv->sv_name, libcfs_id2str(ev->initiator),
1530                                ev->status, ev->mlength,
1531                                msg->msg_type, msg->msg_magic);
1532
1533                         /*
1534                          * NB can't call srpc_service_recycle_buffer here since
1535                          * it may call LNetM[DE]Attach. The invalid magic tells
1536                          * srpc_handle_rpc to drop this RPC
1537                          */
1538                         msg->msg_magic = 0;
1539                 }
1540
1541                 if (!list_empty(&scd->scd_rpc_free)) {
1542                         srpc = list_entry(scd->scd_rpc_free.next,
1543                                           struct srpc_server_rpc,
1544                                           srpc_list);
1545                         list_del(&srpc->srpc_list);
1546
1547                         srpc_init_server_rpc(srpc, scd, buffer);
1548                         list_add_tail(&srpc->srpc_list,
1549                                       &scd->scd_rpc_active);
1550                         swi_schedule_workitem(&srpc->srpc_wi);
1551                 } else {
1552                         list_add_tail(&buffer->buf_list,
1553                                       &scd->scd_buf_blocked);
1554                 }
1555
1556                 spin_unlock(&scd->scd_lock);
1557
1558                 spin_lock(&srpc_data.rpc_glock);
1559                 srpc_data.rpc_counters.rpcs_rcvd++;
1560                 spin_unlock(&srpc_data.rpc_glock);
1561                 break;
1562
1563         case SRPC_BULK_GET_RPLD:
1564                 LASSERT(ev->type == LNET_EVENT_SEND ||
1565                         ev->type == LNET_EVENT_REPLY ||
1566                         ev->type == LNET_EVENT_UNLINK);
1567
1568                 if (!ev->unlinked)
1569                         break; /* wait for final event */
1570
1571         case SRPC_BULK_PUT_SENT:
1572                 if (!ev->status && ev->type != LNET_EVENT_UNLINK) {
1573                         spin_lock(&srpc_data.rpc_glock);
1574
1575                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1576                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1577                         else
1578                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1579
1580                         spin_unlock(&srpc_data.rpc_glock);
1581                 }
1582         case SRPC_REPLY_SENT:
1583                 srpc = rpcev->ev_data;
1584                 scd  = srpc->srpc_scd;
1585
1586                 LASSERT(rpcev == &srpc->srpc_ev);
1587
1588                 spin_lock(&scd->scd_lock);
1589
1590                 rpcev->ev_fired  = 1;
1591                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1592                                    -EINTR : ev->status;
1593                 swi_schedule_workitem(&srpc->srpc_wi);
1594
1595                 spin_unlock(&scd->scd_lock);
1596                 break;
1597         }
1598 }
1599
1600 int
1601 srpc_startup(void)
1602 {
1603         int rc;
1604
1605         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1606         spin_lock_init(&srpc_data.rpc_glock);
1607
1608         /* 1 second pause to avoid timestamp reuse */
1609         set_current_state(TASK_UNINTERRUPTIBLE);
1610         schedule_timeout(cfs_time_seconds(1));
1611         srpc_data.rpc_matchbits = ((__u64)ktime_get_real_seconds()) << 48;
1612
1613         srpc_data.rpc_state = SRPC_STATE_NONE;
1614
1615         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1616         if (rc < 0) {
1617                 CERROR("LNetNIInit() has failed: %d\n", rc);
1618                 return rc;
1619         }
1620
1621         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1622
1623         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1624         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1625         if (rc) {
1626                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1627                 goto bail;
1628         }
1629
1630         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1631         LASSERT(!rc);
1632         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1633         LASSERT(!rc);
1634
1635         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1636
1637         rc = stt_startup();
1638
1639 bail:
1640         if (rc)
1641                 srpc_shutdown();
1642         else
1643                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1644
1645         return rc;
1646 }
1647
1648 void
1649 srpc_shutdown(void)
1650 {
1651         int i;
1652         int rc;
1653         int state;
1654
1655         state = srpc_data.rpc_state;
1656         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1657
1658         switch (state) {
1659         default:
1660                 LBUG();
1661         case SRPC_STATE_RUNNING:
1662                 spin_lock(&srpc_data.rpc_glock);
1663
1664                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1665                         srpc_service_t *sv = srpc_data.rpc_services[i];
1666
1667                         LASSERTF(!sv, "service not empty: id %d, name %s\n",
1668                                  i, sv->sv_name);
1669                 }
1670
1671                 spin_unlock(&srpc_data.rpc_glock);
1672
1673                 stt_shutdown();
1674
1675         case SRPC_STATE_EQ_INIT:
1676                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1677                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1678                 LASSERT(!rc);
1679                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1680                 LASSERT(!rc); /* the EQ should have no user by now */
1681
1682         case SRPC_STATE_NI_INIT:
1683                 LNetNIFini();
1684         }
1685
1686         return;
1687 }