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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/selftest/rpc.c
  33  *
  34  * Author: Isaac Huang <isaac@clusterfs.com>
  35  *
  36  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  37  * - percpt data for service to improve smp performance
  38  * - code cleanup
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include "selftest.h"
  44
  45 enum srpc_state {
  46         SRPC_STATE_NONE,
  47         SRPC_STATE_NI_INIT,
  48         SRPC_STATE_EQ_INIT,
  49         SRPC_STATE_RUNNING,
  50         SRPC_STATE_STOPPING,
  51 };
  52
  53 static struct smoketest_rpc {
  54         spinlock_t       rpc_glock;     /* global lock */
  55         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  56         struct lnet_handle_eq    rpc_lnet_eq;   /* _the_ LNet event queue */
  57         enum srpc_state          rpc_state;
  58         struct srpc_counters     rpc_counters;
  59         __u64                    rpc_matchbits; /* matchbits counter */
  60 } srpc_data;
  61
  62 static inline int
  63 srpc_serv_portal(int svc_id)
  64 {
  65         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  66                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  67 }
  68
  69 /* forward ref's */
  70 static int srpc_handle_rpc(struct swi_workitem *wi);
  71
  72 void srpc_get_counters(struct srpc_counters *cnt)
  73 {
  74         spin_lock(&srpc_data.rpc_glock);
  75         *cnt = srpc_data.rpc_counters;
  76         spin_unlock(&srpc_data.rpc_glock);
  77 }
  78
  79 void srpc_set_counters(const struct srpc_counters *cnt)
  80 {
  81         spin_lock(&srpc_data.rpc_glock);
  82         srpc_data.rpc_counters = *cnt;
  83         spin_unlock(&srpc_data.rpc_glock);
  84 }
  85
  86 static int
  87 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
  88                    int nob)
  89 {
  90         LASSERT(off < PAGE_SIZE);
  91         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  92
  93         bk->bk_iovs[i].kiov_offset = off;
  94         bk->bk_iovs[i].kiov_page   = pg;
  95         bk->bk_iovs[i].kiov_len    = nob;
  96         return nob;
  97 }
  98
  99 void
 100 srpc_free_bulk(struct srpc_bulk *bk)
 101 {
 102         int i;
 103         struct page *pg;
 104
 105         LASSERT(bk != NULL);
 106
 107         for (i = 0; i < bk->bk_niov; i++) {
 108                 pg = bk->bk_iovs[i].kiov_page;
 109                 if (pg == NULL)
 110                         break;
 111
 112                 __free_page(pg);
 113         }
 114
 115         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
 116         return;
 117 }
 118
 119 struct srpc_bulk *
 120 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 121                 unsigned bulk_len, int sink)
 122 {
 123         struct srpc_bulk *bk;
 124         int i;
 125
 126         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 127
 128         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 129                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 130         if (bk == NULL) {
 131                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 132                 return NULL;
 133         }
 134
 135         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 136         bk->bk_sink   = sink;
 137         bk->bk_len    = bulk_len;
 138         bk->bk_niov   = bulk_npg;
 139
 140         for (i = 0; i < bulk_npg; i++) {
 141                 struct page *pg;
 142                 int nob;
 143
 144                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 145                 if (pg == NULL) {
 146                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 147                         srpc_free_bulk(bk);
 148                         return NULL;
 149                 }
 150
 151                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 152                       bulk_off;
 153
 154                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 155                 bulk_len -= nob;
 156                 bulk_off = 0;
 157         }
 158
 159         return bk;
 160 }
 161
 162 static inline __u64
 163 srpc_next_id (void)
 164 {
 165         __u64 id;
 166
 167         spin_lock(&srpc_data.rpc_glock);
 168         id = srpc_data.rpc_matchbits++;
 169         spin_unlock(&srpc_data.rpc_glock);
 170         return id;
 171 }
 172
 173 static void
 174 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 175                      struct srpc_service_cd *scd,
 176                      struct srpc_buffer *buffer)
 177 {
 178         memset(rpc, 0, sizeof(*rpc));
 179         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 180                           srpc_serv_is_framework(scd->scd_svc) ?
 181                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 182
 183         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 184
 185         rpc->srpc_scd      = scd;
 186         rpc->srpc_reqstbuf = buffer;
 187         rpc->srpc_peer     = buffer->buf_peer;
 188         rpc->srpc_self     = buffer->buf_self;
 189         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 190 }
 191
 192 static void
 193 srpc_service_fini(struct srpc_service *svc)
 194 {
 195         struct srpc_service_cd *scd;
 196         struct srpc_server_rpc *rpc;
 197         struct srpc_buffer *buf;
 198         struct list_head *q;
 199         int i;
 200
 201         if (svc->sv_cpt_data == NULL)
 202                 return;
 203
 204         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 205                 while (1) {
 206                         if (!list_empty(&scd->scd_buf_posted))
 207                                 q = &scd->scd_buf_posted;
 208                         else if (!list_empty(&scd->scd_buf_blocked))
 209                                 q = &scd->scd_buf_blocked;
 210                         else
 211                                 break;
 212
 213                         while (!list_empty(q)) {
 214                                 buf = list_entry(q->next,
 215                                                  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 == NULL)
 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                 /* NB: don't use lst_sched_serial for adding buffer,
 281                  * see details in srpc_service_add_buffers() */
 282                 swi_init_workitem(&scd->scd_buf_wi,
 283                                   srpc_add_buffer, lst_sched_test[i]);
 284
 285                 if (i != 0 && srpc_serv_is_framework(svc)) {
 286                         /* NB: framework service only needs srpc_service_cd for
 287                          * one partition, but we allocate for all to make
 288                          * it easier to implement, it will waste a little
 289                          * memory but nobody should care about this */
 290                         continue;
 291                 }
 292
 293                 for (j = 0; j < nrpcs; j++) {
 294                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 295                                          i, sizeof(*rpc));
 296                         if (rpc == NULL) {
 297                                 srpc_service_fini(svc);
 298                                 return -ENOMEM;
 299                         }
 300                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 301                 }
 302         }
 303
 304         return 0;
 305 }
 306
 307 int
 308 srpc_add_service(struct srpc_service *sv)
 309 {
 310         int id = sv->sv_id;
 311
 312         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 313
 314         if (srpc_service_init(sv) != 0)
 315                 return -ENOMEM;
 316
 317         spin_lock(&srpc_data.rpc_glock);
 318
 319         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 320
 321         if (srpc_data.rpc_services[id] != NULL) {
 322                 spin_unlock(&srpc_data.rpc_glock);
 323                 goto failed;
 324         }
 325
 326         srpc_data.rpc_services[id] = sv;
 327         spin_unlock(&srpc_data.rpc_glock);
 328
 329         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 330         return 0;
 331
 332 failed:
 333         srpc_service_fini(sv);
 334         return -EBUSY;
 335 }
 336
 337 int
 338 srpc_remove_service(struct srpc_service *sv)
 339 {
 340         int id = sv->sv_id;
 341
 342         spin_lock(&srpc_data.rpc_glock);
 343
 344         if (srpc_data.rpc_services[id] != sv) {
 345                 spin_unlock(&srpc_data.rpc_glock);
 346                 return -ENOENT;
 347         }
 348
 349         srpc_data.rpc_services[id] = NULL;
 350         spin_unlock(&srpc_data.rpc_glock);
 351         return 0;
 352 }
 353
 354 static int
 355 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 356                        int len, int options, struct lnet_process_id peer,
 357                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 358 {
 359         int rc;
 360         struct lnet_md md;
 361         struct lnet_handle_me meh;
 362
 363         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 364                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 365         if (rc != 0) {
 366                 CERROR("LNetMEAttach failed: %d\n", rc);
 367                 LASSERT(rc == -ENOMEM);
 368                 return -ENOMEM;
 369         }
 370
 371         md.threshold = 1;
 372         md.user_ptr  = ev;
 373         md.start     = buf;
 374         md.length    = len;
 375         md.options   = options;
 376         md.eq_handle = srpc_data.rpc_lnet_eq;
 377
 378         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 379         if (rc != 0) {
 380                 CERROR("LNetMDAttach failed: %d\n", rc);
 381                 LASSERT(rc == -ENOMEM);
 382
 383                 rc = LNetMEUnlink(meh);
 384                 LASSERT(rc == 0);
 385                 return -ENOMEM;
 386         }
 387
 388         CDEBUG(D_NET,
 389                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 390                libcfs_id2str(peer), portal, matchbits);
 391         return 0;
 392 }
 393
 394 static int
 395 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 396                       int options, struct lnet_process_id peer,
 397                       lnet_nid_t self, struct lnet_handle_md *mdh,
 398                       struct srpc_event *ev)
 399 {
 400         int rc;
 401         struct lnet_md md;
 402
 403         md.user_ptr  = ev;
 404         md.start     = buf;
 405         md.length    = len;
 406         md.eq_handle = srpc_data.rpc_lnet_eq;
 407         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 408         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 409
 410         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 411         if (rc != 0) {
 412                 CERROR("LNetMDBind failed: %d\n", rc);
 413                 LASSERT(rc == -ENOMEM);
 414                 return -ENOMEM;
 415         }
 416
 417         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 418          * they're only meaningful for MDs attached to an ME (i.e. passive
 419          * buffers...
 420          */
 421         if ((options & LNET_MD_OP_PUT) != 0) {
 422                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 423                              portal, matchbits, 0, 0);
 424         } else {
 425                 LASSERT((options & LNET_MD_OP_GET) != 0);
 426
 427                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0, false);
 428         }
 429
 430         if (rc != 0) {
 431                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 432                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 433                        libcfs_id2str(peer), portal, matchbits, rc);
 434
 435                 /* The forthcoming unlink event will complete this operation
 436                  * with failure, so fall through and return success here.
 437                  */
 438                 rc = LNetMDUnlink(*mdh);
 439                 LASSERT(rc == 0);
 440         } else {
 441                 CDEBUG(D_NET,
 442                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 443                        libcfs_id2str(peer), portal, matchbits);
 444         }
 445         return 0;
 446 }
 447
 448 static int
 449 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 450                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 451 {
 452         struct lnet_process_id any = {0};
 453
 454         any.nid = LNET_NID_ANY;
 455         any.pid = LNET_PID_ANY;
 456
 457         return srpc_post_passive_rdma(srpc_serv_portal(service),
 458                                       local, service, buf, len,
 459                                       LNET_MD_OP_PUT, any, mdh, ev);
 460 }
 461
 462 static int
 463 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 464 __must_hold(&scd->scd_lock)
 465 {
 466         struct srpc_service *sv = scd->scd_svc;
 467         struct srpc_msg *msg = &buf->buf_msg;
 468         int rc;
 469
 470         LNetInvalidateMDHandle(&buf->buf_mdh);
 471         list_add(&buf->buf_list, &scd->scd_buf_posted);
 472         scd->scd_buf_nposted++;
 473         spin_unlock(&scd->scd_lock);
 474
 475         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 476                                       !srpc_serv_is_framework(sv),
 477                                       msg, sizeof(*msg), &buf->buf_mdh,
 478                                       &scd->scd_ev);
 479
 480         /* At this point, a RPC (new or delayed) may have arrived in
 481          * msg and its event handler has been called. So we must add
 482          * buf to scd_buf_posted _before_ dropping scd_lock */
 483
 484         spin_lock(&scd->scd_lock);
 485
 486         if (rc == 0) {
 487                 if (!sv->sv_shuttingdown)
 488                         return 0;
 489
 490                 spin_unlock(&scd->scd_lock);
 491                 /* srpc_shutdown_service might have tried to unlink me
 492                  * when my buf_mdh was still invalid */
 493                 LNetMDUnlink(buf->buf_mdh);
 494                 spin_lock(&scd->scd_lock);
 495                 return 0;
 496         }
 497
 498         scd->scd_buf_nposted--;
 499         if (sv->sv_shuttingdown)
 500                 return rc; /* don't allow to change scd_buf_posted */
 501
 502         list_del(&buf->buf_list);
 503         spin_unlock(&scd->scd_lock);
 504
 505         LIBCFS_FREE(buf, sizeof(*buf));
 506
 507         spin_lock(&scd->scd_lock);
 508         return rc;
 509 }
 510
 511 int
 512 srpc_add_buffer(struct swi_workitem *wi)
 513 {
 514         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 515                                                    scd_buf_wi);
 516         struct srpc_buffer *buf;
 517         int rc = 0;
 518
 519         /* it's called by workitem scheduler threads, these threads
 520          * should have been set CPT affinity, so buffers will be posted
 521          * on CPT local list of Portal */
 522         spin_lock(&scd->scd_lock);
 523
 524         while (scd->scd_buf_adjust > 0 &&
 525                !scd->scd_svc->sv_shuttingdown) {
 526                 scd->scd_buf_adjust--; /* consume it */
 527                 scd->scd_buf_posting++;
 528
 529                 spin_unlock(&scd->scd_lock);
 530
 531                 LIBCFS_ALLOC(buf, sizeof(*buf));
 532                 if (buf == NULL) {
 533                         CERROR("Failed to add new buf to service: %s\n",
 534                                scd->scd_svc->sv_name);
 535                         spin_lock(&scd->scd_lock);
 536                         rc = -ENOMEM;
 537                         break;
 538                 }
 539
 540                 spin_lock(&scd->scd_lock);
 541                 if (scd->scd_svc->sv_shuttingdown) {
 542                         spin_unlock(&scd->scd_lock);
 543                         LIBCFS_FREE(buf, sizeof(*buf));
 544
 545                         spin_lock(&scd->scd_lock);
 546                         rc = -ESHUTDOWN;
 547                         break;
 548                 }
 549
 550                 rc = srpc_service_post_buffer(scd, buf);
 551                 if (rc != 0)
 552                         break; /* buf has been freed inside */
 553
 554                 LASSERT(scd->scd_buf_posting > 0);
 555                 scd->scd_buf_posting--;
 556                 scd->scd_buf_total++;
 557                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 558         }
 559
 560         if (rc != 0) {
 561                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 562                 scd->scd_buf_err = rc;
 563
 564                 LASSERT(scd->scd_buf_posting > 0);
 565                 scd->scd_buf_posting--;
 566         }
 567
 568         spin_unlock(&scd->scd_lock);
 569         return 0;
 570 }
 571
 572 int
 573 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 574 {
 575         struct srpc_service_cd *scd;
 576         int rc = 0;
 577         int i;
 578
 579         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 580
 581         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 582                 spin_lock(&scd->scd_lock);
 583
 584                 scd->scd_buf_err = 0;
 585                 scd->scd_buf_err_stamp = 0;
 586                 scd->scd_buf_posting = 0;
 587                 scd->scd_buf_adjust = nbuffer;
 588                 /* start to post buffers */
 589                 swi_schedule_workitem(&scd->scd_buf_wi);
 590                 spin_unlock(&scd->scd_lock);
 591
 592                 /* framework service only post buffer for one partition  */
 593                 if (srpc_serv_is_framework(sv))
 594                         break;
 595         }
 596
 597         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 598                 spin_lock(&scd->scd_lock);
 599                 /*
 600                  * NB: srpc_service_add_buffers() can be called inside
 601                  * thread context of lst_sched_serial, and we don't normally
 602                  * allow to sleep inside thread context of WI scheduler
 603                  * because it will block current scheduler thread from doing
 604                  * anything else, even worse, it could deadlock if it's
 605                  * waiting on result from another WI of the same scheduler.
 606                  * However, it's safe at here because scd_buf_wi is scheduled
 607                  * by thread in a different WI scheduler (lst_sched_test),
 608                  * so we don't have any risk of deadlock, though this could
 609                  * block all WIs pending on lst_sched_serial for a moment
 610                  * which is not good but not fatal.
 611                  */
 612                 lst_wait_until(scd->scd_buf_err != 0 ||
 613                                (scd->scd_buf_adjust == 0 &&
 614                                 scd->scd_buf_posting == 0),
 615                                scd->scd_lock, "waiting for adding buffer\n");
 616
 617                 if (scd->scd_buf_err != 0 && rc == 0)
 618                         rc = scd->scd_buf_err;
 619
 620                 spin_unlock(&scd->scd_lock);
 621         }
 622
 623         return rc;
 624 }
 625
 626 void
 627 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 628 {
 629         struct srpc_service_cd *scd;
 630         int num;
 631         int i;
 632
 633         LASSERT(!sv->sv_shuttingdown);
 634
 635         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 636                 spin_lock(&scd->scd_lock);
 637
 638                 num = scd->scd_buf_total + scd->scd_buf_posting;
 639                 scd->scd_buf_adjust -= min(nbuffer, num);
 640
 641                 spin_unlock(&scd->scd_lock);
 642         }
 643 }
 644
 645 /* returns 1 if sv has finished, otherwise 0 */
 646 int
 647 srpc_finish_service(struct srpc_service *sv)
 648 {
 649         struct srpc_service_cd *scd;
 650         struct srpc_server_rpc *rpc;
 651         int i;
 652
 653         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 654
 655         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 656                 spin_lock(&scd->scd_lock);
 657                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 658                         spin_unlock(&scd->scd_lock);
 659                         return 0;
 660                 }
 661
 662                 if (scd->scd_buf_nposted > 0) {
 663                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 664                                scd->scd_buf_nposted);
 665                         spin_unlock(&scd->scd_lock);
 666                         return 0;
 667                 }
 668
 669                 if (list_empty(&scd->scd_rpc_active)) {
 670                         spin_unlock(&scd->scd_lock);
 671                         continue;
 672                 }
 673
 674                 rpc = list_entry(scd->scd_rpc_active.next,
 675                                  struct srpc_server_rpc, srpc_list);
 676                 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",
 677                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 678                         swi_state2str(rpc->srpc_wi.swi_state),
 679                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 680                         rpc->srpc_wi.swi_workitem.wi_running,
 681                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 682                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 683                 spin_unlock(&scd->scd_lock);
 684                 return 0;
 685         }
 686
 687         /* no lock needed from now on */
 688         srpc_service_fini(sv);
 689         return 1;
 690 }
 691
 692 /* called with sv->sv_lock held */
 693 static void
 694 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 695                             struct srpc_buffer *buf)
 696 __must_hold(&scd->scd_lock)
 697 {
 698         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 699                 if (srpc_service_post_buffer(scd, buf) != 0) {
 700                         CWARN("Failed to post %s buffer\n",
 701                               scd->scd_svc->sv_name);
 702                 }
 703                 return;
 704         }
 705
 706         /* service is shutting down, or we want to recycle some buffers */
 707         scd->scd_buf_total--;
 708
 709         if (scd->scd_buf_adjust < 0) {
 710                 scd->scd_buf_adjust++;
 711                 if (scd->scd_buf_adjust < 0 &&
 712                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 713                         CDEBUG(D_INFO,
 714                                "Try to recyle %d buffers but nothing left\n",
 715                                scd->scd_buf_adjust);
 716                         scd->scd_buf_adjust = 0;
 717                 }
 718         }
 719
 720         spin_unlock(&scd->scd_lock);
 721         LIBCFS_FREE(buf, sizeof(*buf));
 722         spin_lock(&scd->scd_lock);
 723 }
 724
 725 void
 726 srpc_abort_service(struct srpc_service *sv)
 727 {
 728         struct srpc_service_cd *scd;
 729         struct srpc_server_rpc *rpc;
 730         int i;
 731
 732         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 733                sv->sv_id, sv->sv_name);
 734
 735         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 736                 spin_lock(&scd->scd_lock);
 737
 738                 /* schedule in-flight RPCs to notice the abort, NB:
 739                  * racing with incoming RPCs; complete fix should make test
 740                  * RPCs carry session ID in its headers
 741                  */
 742                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 743                         rpc->srpc_aborted = 1;
 744                         swi_schedule_workitem(&rpc->srpc_wi);
 745                 }
 746
 747                 spin_unlock(&scd->scd_lock);
 748         }
 749 }
 750
 751 void
 752 srpc_shutdown_service(struct srpc_service *sv)
 753 {
 754         struct srpc_service_cd *scd;
 755         struct srpc_server_rpc *rpc;
 756         struct srpc_buffer *buf;
 757         int i;
 758
 759         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 760                sv->sv_id, sv->sv_name);
 761
 762         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 763         spin_lock(&scd->scd_lock);
 764
 765         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 766
 767         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 768         spin_unlock(&scd->scd_lock);
 769
 770         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 771                 spin_lock(&scd->scd_lock);
 772
 773                 /* schedule in-flight RPCs to notice the shutdown */
 774                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 775                 swi_schedule_workitem(&rpc->srpc_wi);
 776
 777                 spin_unlock(&scd->scd_lock);
 778
 779                 /* OK to traverse scd_buf_posted without lock, since no one
 780                  * touches scd_buf_posted now
 781                  */
 782                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 783                 LNetMDUnlink(buf->buf_mdh);
 784         }
 785 }
 786
 787 static int
 788 srpc_send_request(struct srpc_client_rpc *rpc)
 789 {
 790         struct srpc_event *ev = &rpc->crpc_reqstev;
 791         int rc;
 792
 793         ev->ev_fired = 0;
 794         ev->ev_data  = rpc;
 795         ev->ev_type  = SRPC_REQUEST_SENT;
 796
 797         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 798                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 799                                    sizeof(struct srpc_msg), LNET_MD_OP_PUT,
 800                                    rpc->crpc_dest, LNET_NID_ANY,
 801                                    &rpc->crpc_reqstmdh, ev);
 802         if (rc != 0) {
 803                 LASSERT(rc == -ENOMEM);
 804                 ev->ev_fired = 1;  /* no more event expected */
 805         }
 806         return rc;
 807 }
 808
 809 static int
 810 srpc_prepare_reply(struct srpc_client_rpc *rpc)
 811 {
 812         struct srpc_event *ev = &rpc->crpc_replyev;
 813         u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 814         int rc;
 815
 816         ev->ev_fired = 0;
 817         ev->ev_data  = rpc;
 818         ev->ev_type  = SRPC_REPLY_RCVD;
 819
 820         *id = srpc_next_id();
 821
 822         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 823                                     &rpc->crpc_replymsg,
 824                                     sizeof(struct srpc_msg),
 825                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 826                                     &rpc->crpc_replymdh, ev);
 827         if (rc != 0) {
 828                 LASSERT(rc == -ENOMEM);
 829                 ev->ev_fired = 1;  /* no more event expected */
 830         }
 831         return rc;
 832 }
 833
 834 static int
 835 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 836 {
 837         struct srpc_bulk *bk = &rpc->crpc_bulk;
 838         struct srpc_event *ev = &rpc->crpc_bulkev;
 839         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 840         int rc;
 841         int opt;
 842
 843         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 844
 845         /* nothing to do */
 846         if (bk->bk_niov == 0)
 847                 return 0;
 848
 849         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 850         opt |= LNET_MD_KIOV;
 851
 852         ev->ev_fired = 0;
 853         ev->ev_data  = rpc;
 854         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 855
 856         *id = srpc_next_id();
 857
 858         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 859                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 860                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 861         if (rc != 0) {
 862                 LASSERT(rc == -ENOMEM);
 863                 ev->ev_fired = 1;  /* no more event expected */
 864         }
 865         return rc;
 866 }
 867
 868 static int
 869 srpc_do_bulk(struct srpc_server_rpc *rpc)
 870 {
 871         struct srpc_event *ev = &rpc->srpc_ev;
 872         struct srpc_bulk *bk = rpc->srpc_bulk;
 873         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 874         int rc;
 875         int opt;
 876
 877         LASSERT(bk != NULL);
 878
 879         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 880         opt |= LNET_MD_KIOV;
 881
 882         ev->ev_fired = 0;
 883         ev->ev_data  = rpc;
 884         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 885
 886         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 887                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 888                                    rpc->srpc_peer, rpc->srpc_self,
 889                                    &bk->bk_mdh, ev);
 890         if (rc != 0)
 891                 ev->ev_fired = 1;  /* no more event expected */
 892         return rc;
 893 }
 894
 895 /* only called from srpc_handle_rpc */
 896 static void
 897 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 898 {
 899         struct srpc_service_cd  *scd = rpc->srpc_scd;
 900         struct srpc_service     *sv  = scd->scd_svc;
 901         struct srpc_buffer *buffer;
 902
 903         LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 904
 905         rpc->srpc_status = status;
 906
 907         CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
 908                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 909                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 910                      swi_state2str(rpc->srpc_wi.swi_state), status);
 911
 912         if (status != 0) {
 913                 spin_lock(&srpc_data.rpc_glock);
 914                 srpc_data.rpc_counters.rpcs_dropped++;
 915                 spin_unlock(&srpc_data.rpc_glock);
 916         }
 917
 918         if (rpc->srpc_done != NULL)
 919                 (*rpc->srpc_done) (rpc);
 920         LASSERT(rpc->srpc_bulk == NULL);
 921
 922         spin_lock(&scd->scd_lock);
 923
 924         if (rpc->srpc_reqstbuf != NULL) {
 925                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 926                  * sv won't go away for scd_rpc_active must not be empty
 927                  */
 928                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 929                 rpc->srpc_reqstbuf = NULL;
 930         }
 931
 932         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 933
 934         /*
 935          * No one can schedule me now since:
 936          * - I'm not on scd_rpc_active.
 937          * - all LNet events have been fired.
 938          * Cancel pending schedules and prevent future schedule attempts:
 939          */
 940         LASSERT(rpc->srpc_ev.ev_fired);
 941         swi_exit_workitem(&rpc->srpc_wi);
 942
 943         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 944                 buffer = list_entry(scd->scd_buf_blocked.next,
 945                                     struct srpc_buffer, buf_list);
 946                 list_del(&buffer->buf_list);
 947
 948                 srpc_init_server_rpc(rpc, scd, buffer);
 949                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 950                 swi_schedule_workitem(&rpc->srpc_wi);
 951         } else {
 952                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 953         }
 954
 955         spin_unlock(&scd->scd_lock);
 956         return;
 957 }
 958
 959 /* handles an incoming RPC */
 960 static int srpc_handle_rpc(struct swi_workitem *wi)
 961 {
 962         struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc,
 963                                                    srpc_wi);
 964         struct srpc_service_cd *scd = rpc->srpc_scd;
 965         struct srpc_service *sv = scd->scd_svc;
 966         struct srpc_event *ev = &rpc->srpc_ev;
 967         int rc = 0;
 968
 969         LASSERT(wi == &rpc->srpc_wi);
 970
 971         spin_lock(&scd->scd_lock);
 972
 973         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 974                 spin_unlock(&scd->scd_lock);
 975
 976                 if (rpc->srpc_bulk != NULL)
 977                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 978                 LNetMDUnlink(rpc->srpc_replymdh);
 979
 980                 if (ev->ev_fired) { /* no more event, OK to finish */
 981                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 982                         return 1;
 983                 }
 984                 return 0;
 985         }
 986
 987         spin_unlock(&scd->scd_lock);
 988
 989         switch (wi->swi_state) {
 990         default:
 991                 LBUG();
 992                 /* fallthrough */
 993         case SWI_STATE_NEWBORN: {
 994                 struct srpc_msg *msg;
 995                 struct srpc_generic_reply *reply;
 996
 997                 msg = &rpc->srpc_reqstbuf->buf_msg;
 998                 reply = &rpc->srpc_replymsg.msg_body.reply;
 999
1000                 if (msg->msg_magic == 0) {
1001                         /* moaned already in srpc_lnet_ev_handler */
1002                         srpc_server_rpc_done(rpc, EBADMSG);
1003                         return 1;
1004                 }
1005
1006                 srpc_unpack_msg_hdr(msg);
1007                 if (msg->msg_version != SRPC_MSG_VERSION) {
1008                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1009                               msg->msg_version, SRPC_MSG_VERSION,
1010                               libcfs_id2str(rpc->srpc_peer));
1011                         reply->status = EPROTO;
1012                         /* drop through and send reply */
1013                 } else {
1014                         reply->status = 0;
1015                         rc = (*sv->sv_handler)(rpc);
1016                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1017                         if (rc != 0) {
1018                                 srpc_server_rpc_done(rpc, rc);
1019                                 return 1;
1020                         }
1021                 }
1022
1023                 wi->swi_state = SWI_STATE_BULK_STARTED;
1024
1025                 if (rpc->srpc_bulk != NULL) {
1026                         rc = srpc_do_bulk(rpc);
1027                         if (rc == 0)
1028                                 return 0; /* wait for bulk */
1029
1030                         LASSERT(ev->ev_fired);
1031                         ev->ev_status = rc;
1032                 }
1033         }
1034         /* fallthrough */
1035         case SWI_STATE_BULK_STARTED:
1036                 LASSERT(rpc->srpc_bulk == NULL || ev->ev_fired);
1037
1038                 if (rpc->srpc_bulk != NULL) {
1039                         rc = ev->ev_status;
1040
1041                         if (sv->sv_bulk_ready != NULL)
1042                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1043
1044                         if (rc != 0) {
1045                                 srpc_server_rpc_done(rpc, rc);
1046                                 return 1;
1047                         }
1048                 }
1049
1050                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1051                 rc = srpc_send_reply(rpc);
1052                 if (rc == 0)
1053                         return 0; /* wait for reply */
1054                 srpc_server_rpc_done(rpc, rc);
1055                 return 1;
1056
1057         case SWI_STATE_REPLY_SUBMITTED:
1058                 if (!ev->ev_fired) {
1059                         CERROR("RPC %p: bulk %p, service %d\n",
1060                                rpc, rpc->srpc_bulk, sv->sv_id);
1061                         CERROR("Event: status %d, type %d, lnet %d\n",
1062                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1063                         LASSERT(ev->ev_fired);
1064                 }
1065
1066                 wi->swi_state = SWI_STATE_DONE;
1067                 srpc_server_rpc_done(rpc, ev->ev_status);
1068                 return 1;
1069         }
1070
1071         return 0;
1072 }
1073
1074 static void
1075 srpc_client_rpc_expired (void *data)
1076 {
1077         struct srpc_client_rpc *rpc = data;
1078
1079         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1080               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1081               rpc->crpc_timeout);
1082
1083         spin_lock(&rpc->crpc_lock);
1084
1085         rpc->crpc_timeout = 0;
1086         srpc_abort_rpc(rpc, -ETIMEDOUT);
1087
1088         spin_unlock(&rpc->crpc_lock);
1089
1090         spin_lock(&srpc_data.rpc_glock);
1091         srpc_data.rpc_counters.rpcs_expired++;
1092         spin_unlock(&srpc_data.rpc_glock);
1093 }
1094
1095 static void
1096 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1097 {
1098         struct stt_timer *timer = &rpc->crpc_timer;
1099
1100         if (rpc->crpc_timeout == 0)
1101                 return;
1102
1103         INIT_LIST_HEAD(&timer->stt_list);
1104         timer->stt_data    = rpc;
1105         timer->stt_func    = srpc_client_rpc_expired;
1106         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1107         stt_add_timer(timer);
1108         return;
1109 }
1110
1111 /*
1112  * Called with rpc->crpc_lock held.
1113  *
1114  * Upon exit the RPC expiry timer is not queued and the handler is not
1115  * running on any CPU.
1116  */
1117 static void
1118 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1119 {
1120         /* timer not planted or already exploded */
1121         if (rpc->crpc_timeout == 0)
1122                 return;
1123
1124         /* timer successfully defused */
1125         if (stt_del_timer(&rpc->crpc_timer))
1126                 return;
1127
1128         /* timer detonated, wait for it to explode */
1129         while (rpc->crpc_timeout != 0) {
1130                 spin_unlock(&rpc->crpc_lock);
1131
1132                 schedule();
1133
1134                 spin_lock(&rpc->crpc_lock);
1135         }
1136 }
1137
1138 static void
1139 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1140 {
1141         struct swi_workitem *wi = &rpc->crpc_wi;
1142
1143         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1144
1145         spin_lock(&rpc->crpc_lock);
1146
1147         rpc->crpc_closed = 1;
1148         if (rpc->crpc_status == 0)
1149                 rpc->crpc_status = status;
1150
1151         srpc_del_client_rpc_timer(rpc);
1152
1153         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1154                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1155                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1156                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1157
1158         /*
1159          * No one can schedule me now since:
1160          * - RPC timer has been defused.
1161          * - all LNet events have been fired.
1162          * - crpc_closed has been set, preventing srpc_abort_rpc from
1163          *   scheduling me.
1164          * Cancel pending schedules and prevent future schedule attempts:
1165          */
1166         LASSERT(!srpc_event_pending(rpc));
1167         swi_exit_workitem(wi);
1168
1169         spin_unlock(&rpc->crpc_lock);
1170
1171         (*rpc->crpc_done)(rpc);
1172         return;
1173 }
1174
1175 /* sends an outgoing RPC */
1176 int
1177 srpc_send_rpc(struct swi_workitem *wi)
1178 {
1179         int rc = 0;
1180         struct srpc_client_rpc *rpc;
1181         struct srpc_msg *reply;
1182         int do_bulk;
1183
1184         LASSERT(wi != NULL);
1185
1186         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1187
1188         LASSERT(rpc != NULL);
1189         LASSERT(wi == &rpc->crpc_wi);
1190
1191         reply = &rpc->crpc_replymsg;
1192         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1193
1194         spin_lock(&rpc->crpc_lock);
1195
1196         if (rpc->crpc_aborted) {
1197                 spin_unlock(&rpc->crpc_lock);
1198                 goto abort;
1199         }
1200
1201         spin_unlock(&rpc->crpc_lock);
1202
1203         switch (wi->swi_state) {
1204         default:
1205                 LBUG();
1206         case SWI_STATE_NEWBORN:
1207                 LASSERT(!srpc_event_pending(rpc));
1208
1209                 rc = srpc_prepare_reply(rpc);
1210                 if (rc != 0) {
1211                         srpc_client_rpc_done(rpc, rc);
1212                         return 1;
1213                 }
1214
1215                 rc = srpc_prepare_bulk(rpc);
1216                 if (rc != 0)
1217                         break;
1218
1219                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1220                 rc = srpc_send_request(rpc);
1221                 break;
1222
1223         case SWI_STATE_REQUEST_SUBMITTED:
1224                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1225                  * order; however, they're processed in a strict order:
1226                  * rqt, rpy, and bulk.
1227                  */
1228                 if (!rpc->crpc_reqstev.ev_fired)
1229                         break;
1230
1231                 rc = rpc->crpc_reqstev.ev_status;
1232                 if (rc != 0)
1233                         break;
1234
1235                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1236                 /* fallthrough */
1237         case SWI_STATE_REQUEST_SENT: {
1238                 enum srpc_msg_type type;
1239
1240                 type = srpc_service2reply(rpc->crpc_service);
1241
1242                 if (!rpc->crpc_replyev.ev_fired)
1243                         break;
1244
1245                 rc = rpc->crpc_replyev.ev_status;
1246                 if (rc != 0)
1247                         break;
1248
1249                 srpc_unpack_msg_hdr(reply);
1250                 if (reply->msg_type != type ||
1251                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1252                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1253                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1254                               libcfs_id2str(rpc->crpc_dest),
1255                               reply->msg_type, type,
1256                               reply->msg_magic, SRPC_MSG_MAGIC);
1257                         rc = -EBADMSG;
1258                         break;
1259                 }
1260
1261                 if (do_bulk && reply->msg_body.reply.status != 0) {
1262                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1263                               reply->msg_body.reply.status,
1264                               libcfs_id2str(rpc->crpc_dest));
1265                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1266                 }
1267
1268                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1269         }
1270         /* fallthrough */
1271         case SWI_STATE_REPLY_RECEIVED:
1272                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1273                         break;
1274
1275                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1276
1277                 /* Bulk buffer was unlinked due to remote error. Clear error
1278                  * since reply buffer still contains valid data.
1279                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1280                  * remote error.
1281                  */
1282                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1283                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1284                         rc = 0;
1285
1286                 wi->swi_state = SWI_STATE_DONE;
1287                 srpc_client_rpc_done(rpc, rc);
1288                 return 1;
1289         }
1290
1291         if (rc != 0) {
1292                 spin_lock(&rpc->crpc_lock);
1293                 srpc_abort_rpc(rpc, rc);
1294                 spin_unlock(&rpc->crpc_lock);
1295         }
1296
1297 abort:
1298         if (rpc->crpc_aborted) {
1299                 LNetMDUnlink(rpc->crpc_reqstmdh);
1300                 LNetMDUnlink(rpc->crpc_replymdh);
1301                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1302
1303                 if (!srpc_event_pending(rpc)) {
1304                         srpc_client_rpc_done(rpc, -EINTR);
1305                         return 1;
1306                 }
1307         }
1308         return 0;
1309 }
1310
1311 struct srpc_client_rpc *
1312 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1313                        int nbulkiov, int bulklen,
1314                        void (*rpc_done)(struct srpc_client_rpc *),
1315                        void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1316 {
1317         struct srpc_client_rpc *rpc;
1318
1319         LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1320                                    crpc_bulk.bk_iovs[nbulkiov]));
1321         if (rpc == NULL)
1322                 return NULL;
1323
1324         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1325                              bulklen, rpc_done, rpc_fini, priv);
1326         return rpc;
1327 }
1328
1329 /* called with rpc->crpc_lock held */
1330 void
1331 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1332 {
1333         LASSERT(why != 0);
1334
1335         if (rpc->crpc_aborted || /* already aborted */
1336             rpc->crpc_closed)    /* callback imminent */
1337                 return;
1338
1339         CDEBUG(D_NET,
1340                "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1341                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1342                swi_state2str(rpc->crpc_wi.swi_state), why);
1343
1344         rpc->crpc_aborted = 1;
1345         rpc->crpc_status  = why;
1346         swi_schedule_workitem(&rpc->crpc_wi);
1347         return;
1348 }
1349
1350 /* called with rpc->crpc_lock held */
1351 void
1352 srpc_post_rpc(struct srpc_client_rpc *rpc)
1353 {
1354         LASSERT(!rpc->crpc_aborted);
1355         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1356
1357         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1358                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1359                rpc->crpc_timeout);
1360
1361         srpc_add_client_rpc_timer(rpc);
1362         swi_schedule_workitem(&rpc->crpc_wi);
1363         return;
1364 }
1365
1366
1367 int
1368 srpc_send_reply(struct srpc_server_rpc *rpc)
1369 {
1370         struct srpc_event *ev = &rpc->srpc_ev;
1371         struct srpc_msg *msg = &rpc->srpc_replymsg;
1372         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1373         struct srpc_service_cd *scd = rpc->srpc_scd;
1374         struct srpc_service *sv = scd->scd_svc;
1375         __u64 rpyid;
1376         int rc;
1377
1378         LASSERT(buffer != NULL);
1379         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1380
1381         spin_lock(&scd->scd_lock);
1382
1383         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1384                 /* Repost buffer before replying since test client
1385                  * might send me another RPC once it gets the reply
1386                  */
1387                 if (srpc_service_post_buffer(scd, buffer) != 0)
1388                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1389                 rpc->srpc_reqstbuf = NULL;
1390         }
1391
1392         spin_unlock(&scd->scd_lock);
1393
1394         ev->ev_fired = 0;
1395         ev->ev_data  = rpc;
1396         ev->ev_type  = SRPC_REPLY_SENT;
1397
1398         msg->msg_magic   = SRPC_MSG_MAGIC;
1399         msg->msg_version = SRPC_MSG_VERSION;
1400         msg->msg_type    = srpc_service2reply(sv->sv_id);
1401
1402         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1403                                    sizeof(*msg), LNET_MD_OP_PUT,
1404                                    rpc->srpc_peer, rpc->srpc_self,
1405                                    &rpc->srpc_replymdh, ev);
1406         if (rc != 0)
1407                 ev->ev_fired = 1;  /* no more event expected */
1408         return rc;
1409 }
1410
1411 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1412 static void
1413 srpc_lnet_ev_handler(struct lnet_event *ev)
1414 {
1415         struct srpc_service_cd *scd;
1416         struct srpc_event *rpcev = ev->md.user_ptr;
1417         struct srpc_client_rpc *crpc;
1418         struct srpc_server_rpc *srpc;
1419         struct srpc_buffer *buffer;
1420         struct srpc_service *sv;
1421         struct srpc_msg *msg;
1422         enum srpc_msg_type type;
1423
1424         LASSERT(!in_interrupt());
1425
1426         if (ev->status != 0) {
1427                 __u32 errors;
1428
1429                 spin_lock(&srpc_data.rpc_glock);
1430                 if (ev->status != -ECANCELED) /* cancellation is not error */
1431                         srpc_data.rpc_counters.errors++;
1432                 errors = srpc_data.rpc_counters.errors;
1433                 spin_unlock(&srpc_data.rpc_glock);
1434
1435                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1436                         ev->status, ev->type, errors);
1437         }
1438
1439         rpcev->ev_lnet = ev->type;
1440
1441         switch (rpcev->ev_type) {
1442         default:
1443                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1444                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1445                 LBUG();
1446                 /* fallthrough */
1447         case SRPC_REQUEST_SENT:
1448                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1449                         spin_lock(&srpc_data.rpc_glock);
1450                         srpc_data.rpc_counters.rpcs_sent++;
1451                         spin_unlock(&srpc_data.rpc_glock);
1452                 }
1453                 /* fallthrough */
1454         case SRPC_REPLY_RCVD:
1455         case SRPC_BULK_REQ_RCVD:
1456                 crpc = rpcev->ev_data;
1457
1458                 if (rpcev != &crpc->crpc_reqstev &&
1459                     rpcev != &crpc->crpc_replyev &&
1460                     rpcev != &crpc->crpc_bulkev) {
1461                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1462                                rpcev, crpc, &crpc->crpc_reqstev,
1463                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1464                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1465                                rpcev->ev_status, rpcev->ev_type,
1466                                rpcev->ev_lnet);
1467                         LBUG();
1468                 }
1469
1470                 spin_lock(&crpc->crpc_lock);
1471
1472                 LASSERT(rpcev->ev_fired == 0);
1473                 rpcev->ev_fired  = 1;
1474                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1475                                    -EINTR : ev->status;
1476                 swi_schedule_workitem(&crpc->crpc_wi);
1477
1478                 spin_unlock(&crpc->crpc_lock);
1479                 break;
1480
1481         case SRPC_REQUEST_RCVD:
1482                 scd = rpcev->ev_data;
1483                 sv = scd->scd_svc;
1484
1485                 LASSERT(rpcev == &scd->scd_ev);
1486
1487                 spin_lock(&scd->scd_lock);
1488
1489                 LASSERT(ev->unlinked);
1490                 LASSERT(ev->type == LNET_EVENT_PUT ||
1491                         ev->type == LNET_EVENT_UNLINK);
1492                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1493                         sv->sv_shuttingdown);
1494
1495                 buffer = container_of(ev->md.start, struct srpc_buffer,
1496                                       buf_msg);
1497                 buffer->buf_peer = ev->source;
1498                 buffer->buf_self = ev->target.nid;
1499
1500                 LASSERT(scd->scd_buf_nposted > 0);
1501                 scd->scd_buf_nposted--;
1502
1503                 if (sv->sv_shuttingdown) {
1504                         /* Leave buffer on scd->scd_buf_nposted since
1505                          * srpc_finish_service needs to traverse it.
1506                          */
1507                         spin_unlock(&scd->scd_lock);
1508                         break;
1509                 }
1510
1511                 if (scd->scd_buf_err_stamp != 0 &&
1512                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1513                         /* re-enable adding buffer */
1514                         scd->scd_buf_err_stamp = 0;
1515                         scd->scd_buf_err = 0;
1516                 }
1517
1518                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1519                     scd->scd_buf_adjust == 0 &&
1520                     scd->scd_buf_nposted < scd->scd_buf_low) {
1521                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1522                                                   SFW_TEST_WI_MIN);
1523                         swi_schedule_workitem(&scd->scd_buf_wi);
1524                 }
1525
1526                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1527                 msg = &buffer->buf_msg;
1528                 type = srpc_service2request(sv->sv_id);
1529
1530                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1531                     (msg->msg_type != type &&
1532                      msg->msg_type != __swab32(type)) ||
1533                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1534                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1535                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1536                                sv->sv_name, libcfs_id2str(ev->initiator),
1537                                ev->status, ev->mlength,
1538                                msg->msg_type, msg->msg_magic);
1539
1540                         /* NB can't call srpc_service_recycle_buffer here since
1541                          * it may call LNetM[DE]Attach. The invalid magic tells
1542                          * srpc_handle_rpc to drop this RPC
1543                          */
1544                         msg->msg_magic = 0;
1545                 }
1546
1547                 if (!list_empty(&scd->scd_rpc_free)) {
1548                         srpc = list_entry(scd->scd_rpc_free.next,
1549                                           struct srpc_server_rpc,
1550                                           srpc_list);
1551                         list_del(&srpc->srpc_list);
1552
1553                         srpc_init_server_rpc(srpc, scd, buffer);
1554                         list_add_tail(&srpc->srpc_list,
1555                                       &scd->scd_rpc_active);
1556                         swi_schedule_workitem(&srpc->srpc_wi);
1557                 } else {
1558                         list_add_tail(&buffer->buf_list,
1559                                       &scd->scd_buf_blocked);
1560                 }
1561
1562                 spin_unlock(&scd->scd_lock);
1563
1564                 spin_lock(&srpc_data.rpc_glock);
1565                 srpc_data.rpc_counters.rpcs_rcvd++;
1566                 spin_unlock(&srpc_data.rpc_glock);
1567                 break;
1568
1569         case SRPC_BULK_GET_RPLD:
1570                 LASSERT(ev->type == LNET_EVENT_SEND ||
1571                         ev->type == LNET_EVENT_REPLY ||
1572                         ev->type == LNET_EVENT_UNLINK);
1573
1574                 if (!ev->unlinked)
1575                         break; /* wait for final event */
1576                 /* fallthrough */
1577         case SRPC_BULK_PUT_SENT:
1578                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1579                         spin_lock(&srpc_data.rpc_glock);
1580
1581                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1582                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1583                         else
1584                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1585
1586                         spin_unlock(&srpc_data.rpc_glock);
1587                 }
1588                 /* fallthrough */
1589         case SRPC_REPLY_SENT:
1590                 srpc = rpcev->ev_data;
1591                 scd  = srpc->srpc_scd;
1592
1593                 LASSERT(rpcev == &srpc->srpc_ev);
1594
1595                 spin_lock(&scd->scd_lock);
1596
1597                 rpcev->ev_fired  = 1;
1598                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1599                                    -EINTR : ev->status;
1600                 swi_schedule_workitem(&srpc->srpc_wi);
1601
1602                 spin_unlock(&scd->scd_lock);
1603                 break;
1604         }
1605 }
1606
1607
1608 int
1609 srpc_startup (void)
1610 {
1611         int rc;
1612
1613         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1614         spin_lock_init(&srpc_data.rpc_glock);
1615
1616         /* 1 second pause to avoid timestamp reuse */
1617         set_current_state(TASK_UNINTERRUPTIBLE);
1618         schedule_timeout(cfs_time_seconds(1));
1619         srpc_data.rpc_matchbits = ((__u64) ktime_get_real_seconds()) << 48;
1620
1621         srpc_data.rpc_state = SRPC_STATE_NONE;
1622
1623         rc = LNetNIInit(LNET_PID_LUSTRE);
1624         if (rc < 0) {
1625                 CERROR("LNetNIInit() has failed: %d\n", rc);
1626                 return rc;
1627         }
1628
1629         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1630
1631         LNetInvalidateEQHandle(&srpc_data.rpc_lnet_eq);
1632         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1633         if (rc != 0) {
1634                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1635                 goto bail;
1636         }
1637
1638         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1639         LASSERT(rc == 0);
1640         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1641         LASSERT(rc == 0);
1642
1643         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1644
1645         rc = stt_startup();
1646
1647 bail:
1648         if (rc != 0)
1649                 srpc_shutdown();
1650         else
1651                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1652
1653         return rc;
1654 }
1655
1656 void
1657 srpc_shutdown (void)
1658 {
1659         int i;
1660         int rc;
1661         int state;
1662
1663         state = srpc_data.rpc_state;
1664         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1665
1666         switch (state) {
1667         default:
1668                 LBUG();
1669                 /* fallthrough */
1670         case SRPC_STATE_RUNNING:
1671                 spin_lock(&srpc_data.rpc_glock);
1672
1673                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1674                         struct srpc_service *sv = srpc_data.rpc_services[i];
1675
1676                         LASSERTF(sv == NULL,
1677                                  "service not empty: id %d, name %s\n",
1678                                  i, sv->sv_name);
1679                 }
1680
1681                 spin_unlock(&srpc_data.rpc_glock);
1682
1683                 stt_shutdown();
1684                 /* fallthrough */
1685
1686         case SRPC_STATE_EQ_INIT:
1687                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1688                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1689                 LASSERT(rc == 0);
1690                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1691                 LASSERT(rc == 0); /* the EQ should have no user by now */
1692                 /* fallthrough */
1693
1694         case SRPC_STATE_NI_INIT:
1695                 LNetNIFini();
1696         }
1697
1698         return;
1699 }