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  *
  31  * lnet/selftest/rpc.c
  32  *
  33  * Author: Isaac Huang <isaac@clusterfs.com>
  34  *
  35  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  36  * - percpt data for service to improve smp performance
  37  * - code cleanup
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LNET
  41
  42 #include "selftest.h"
  43
  44 enum srpc_state {
  45         SRPC_STATE_NONE,
  46         SRPC_STATE_NI_INIT,
  47         SRPC_STATE_EQ_INIT,
  48         SRPC_STATE_RUNNING,
  49         SRPC_STATE_STOPPING,
  50 };
  51
  52 static struct smoketest_rpc {
  53         spinlock_t       rpc_glock;     /* global lock */
  54         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  55         lnet_handler_t           rpc_lnet_handler;/* _the_ LNet event handler */
  56         enum srpc_state          rpc_state;
  57         struct srpc_counters     rpc_counters;
  58         __u64                    rpc_matchbits; /* matchbits counter */
  59 } srpc_data;
  60
  61 static inline int
  62 srpc_serv_portal(int svc_id)
  63 {
  64         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  65                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  66 }
  67
  68 /* forward ref's */
  69 static int srpc_handle_rpc(struct swi_workitem *wi);
  70
  71 void srpc_get_counters(struct srpc_counters *cnt)
  72 {
  73         spin_lock(&srpc_data.rpc_glock);
  74         *cnt = srpc_data.rpc_counters;
  75         spin_unlock(&srpc_data.rpc_glock);
  76 }
  77
  78 void srpc_set_counters(const struct srpc_counters *cnt)
  79 {
  80         spin_lock(&srpc_data.rpc_glock);
  81         srpc_data.rpc_counters = *cnt;
  82         spin_unlock(&srpc_data.rpc_glock);
  83 }
  84
  85 static int
  86 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
  87                    int nob)
  88 {
  89         LASSERT(off < PAGE_SIZE);
  90         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  91
  92         bk->bk_iovs[i].bv_offset = off;
  93         bk->bk_iovs[i].bv_page   = pg;
  94         bk->bk_iovs[i].bv_len    = nob;
  95         return nob;
  96 }
  97
  98 void
  99 srpc_free_bulk(struct srpc_bulk *bk)
 100 {
 101         int i;
 102         struct page *pg;
 103
 104         LASSERT(bk != NULL);
 105
 106         for (i = 0; i < bk->bk_niov; i++) {
 107                 pg = bk->bk_iovs[i].bv_page;
 108                 if (pg == NULL)
 109                         break;
 110
 111                 __free_page(pg);
 112         }
 113
 114         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
 115 }
 116
 117 struct srpc_bulk *
 118 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 119                 unsigned bulk_len, int sink)
 120 {
 121         struct srpc_bulk *bk;
 122         int i;
 123
 124         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 125
 126         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 127                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 128         if (bk == NULL) {
 129                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 130                 return NULL;
 131         }
 132
 133         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 134         bk->bk_sink   = sink;
 135         bk->bk_len    = bulk_len;
 136         bk->bk_niov   = bulk_npg;
 137
 138         for (i = 0; i < bulk_npg; i++) {
 139                 struct page *pg;
 140                 int nob;
 141
 142                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 143                 if (pg == NULL) {
 144                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 145                         srpc_free_bulk(bk);
 146                         return NULL;
 147                 }
 148
 149                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 150                       bulk_off;
 151
 152                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 153                 bulk_len -= nob;
 154                 bulk_off = 0;
 155         }
 156
 157         return bk;
 158 }
 159
 160 static inline __u64
 161 srpc_next_id (void)
 162 {
 163         __u64 id;
 164
 165         spin_lock(&srpc_data.rpc_glock);
 166         id = srpc_data.rpc_matchbits++;
 167         spin_unlock(&srpc_data.rpc_glock);
 168         return id;
 169 }
 170
 171 static void
 172 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 173                      struct srpc_service_cd *scd,
 174                      struct srpc_buffer *buffer)
 175 {
 176         memset(rpc, 0, sizeof(*rpc));
 177         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 178                           srpc_serv_is_framework(scd->scd_svc) ?
 179                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 180
 181         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 182
 183         rpc->srpc_scd      = scd;
 184         rpc->srpc_reqstbuf = buffer;
 185         rpc->srpc_peer     = buffer->buf_peer;
 186         rpc->srpc_self     = buffer->buf_self;
 187         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 188 }
 189
 190 static void
 191 srpc_service_fini(struct srpc_service *svc)
 192 {
 193         struct srpc_service_cd *scd;
 194         struct srpc_server_rpc *rpc;
 195         struct srpc_buffer *buf;
 196         struct list_head *q;
 197         int i;
 198
 199         if (svc->sv_cpt_data == NULL)
 200                 return;
 201
 202         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 203                 while (1) {
 204                         if (!list_empty(&scd->scd_buf_posted))
 205                                 q = &scd->scd_buf_posted;
 206                         else if (!list_empty(&scd->scd_buf_blocked))
 207                                 q = &scd->scd_buf_blocked;
 208                         else
 209                                 break;
 210
 211                         while (!list_empty(q)) {
 212                                 buf = list_entry(q->next,
 213                                                  struct srpc_buffer,
 214                                                  buf_list);
 215                                 list_del(&buf->buf_list);
 216                                 LIBCFS_FREE(buf, sizeof(*buf));
 217                         }
 218                 }
 219
 220                 LASSERT(list_empty(&scd->scd_rpc_active));
 221
 222                 while (!list_empty(&scd->scd_rpc_free)) {
 223                         rpc = list_entry(scd->scd_rpc_free.next,
 224                                          struct srpc_server_rpc,
 225                                          srpc_list);
 226                         list_del(&rpc->srpc_list);
 227                         LIBCFS_FREE(rpc, sizeof(*rpc));
 228                 }
 229         }
 230
 231         cfs_percpt_free(svc->sv_cpt_data);
 232         svc->sv_cpt_data = NULL;
 233 }
 234
 235 static int
 236 srpc_service_nrpcs(struct srpc_service *svc)
 237 {
 238         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 239
 240         return srpc_serv_is_framework(svc) ?
 241                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 242 }
 243
 244 int srpc_add_buffer(struct swi_workitem *wi);
 245
 246 static int
 247 srpc_service_init(struct srpc_service *svc)
 248 {
 249         struct srpc_service_cd *scd;
 250         struct srpc_server_rpc *rpc;
 251         int nrpcs;
 252         int i;
 253         int j;
 254
 255         svc->sv_shuttingdown = 0;
 256
 257         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 258                                             sizeof(struct srpc_service_cd));
 259         if (svc->sv_cpt_data == NULL)
 260                 return -ENOMEM;
 261
 262         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 263                         1 : cfs_cpt_number(lnet_cpt_table());
 264         nrpcs = srpc_service_nrpcs(svc);
 265
 266         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 267                 scd->scd_cpt = i;
 268                 scd->scd_svc = svc;
 269                 spin_lock_init(&scd->scd_lock);
 270                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 271                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 272                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 273                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 274
 275                 scd->scd_ev.ev_data = scd;
 276                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 277
 278                 /* NB: don't use lst_sched_serial for adding buffer,
 279                  * see details in srpc_service_add_buffers() */
 280                 swi_init_workitem(&scd->scd_buf_wi,
 281                                   srpc_add_buffer, lst_sched_test[i]);
 282
 283                 if (i != 0 && srpc_serv_is_framework(svc)) {
 284                         /* NB: framework service only needs srpc_service_cd for
 285                          * one partition, but we allocate for all to make
 286                          * it easier to implement, it will waste a little
 287                          * memory but nobody should care about this */
 288                         continue;
 289                 }
 290
 291                 for (j = 0; j < nrpcs; j++) {
 292                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 293                                          i, sizeof(*rpc));
 294                         if (rpc == NULL) {
 295                                 srpc_service_fini(svc);
 296                                 return -ENOMEM;
 297                         }
 298                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 299                 }
 300         }
 301
 302         return 0;
 303 }
 304
 305 int
 306 srpc_add_service(struct srpc_service *sv)
 307 {
 308         int id = sv->sv_id;
 309
 310         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 311
 312         if (srpc_service_init(sv) != 0)
 313                 return -ENOMEM;
 314
 315         spin_lock(&srpc_data.rpc_glock);
 316
 317         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 318
 319         if (srpc_data.rpc_services[id] != NULL) {
 320                 spin_unlock(&srpc_data.rpc_glock);
 321                 goto failed;
 322         }
 323
 324         srpc_data.rpc_services[id] = sv;
 325         spin_unlock(&srpc_data.rpc_glock);
 326
 327         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 328         return 0;
 329
 330 failed:
 331         srpc_service_fini(sv);
 332         return -EBUSY;
 333 }
 334
 335 int
 336 srpc_remove_service(struct srpc_service *sv)
 337 {
 338         int id = sv->sv_id;
 339
 340         spin_lock(&srpc_data.rpc_glock);
 341
 342         if (srpc_data.rpc_services[id] != sv) {
 343                 spin_unlock(&srpc_data.rpc_glock);
 344                 return -ENOENT;
 345         }
 346
 347         srpc_data.rpc_services[id] = NULL;
 348         spin_unlock(&srpc_data.rpc_glock);
 349         return 0;
 350 }
 351
 352 static int
 353 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 354                        int len, int options, struct lnet_process_id peer4,
 355                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 356 {
 357         int rc;
 358         struct lnet_md md;
 359         struct lnet_me *me;
 360         struct lnet_processid peer;
 361
 362         peer.pid = peer4.pid;
 363         lnet_nid4_to_nid(peer4.nid, &peer.nid);
 364
 365         me = LNetMEAttach(portal, &peer, matchbits, 0, LNET_UNLINK,
 366                           local ? LNET_INS_LOCAL : LNET_INS_AFTER);
 367         if (IS_ERR(me)) {
 368                 rc = PTR_ERR(me);
 369                 CERROR("LNetMEAttach failed: %d\n", rc);
 370                 LASSERT(rc == -ENOMEM);
 371                 return -ENOMEM;
 372         }
 373
 374         md.threshold = 1;
 375         md.user_ptr  = ev;
 376         md.start     = buf;
 377         md.length    = len;
 378         md.options   = options;
 379         md.handler   = srpc_data.rpc_lnet_handler;
 380
 381         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdh);
 382         if (rc != 0) {
 383                 CERROR("LNetMDAttach failed: %d\n", rc);
 384                 LASSERT(rc == -ENOMEM);
 385
 386                 return -ENOMEM;
 387         }
 388
 389         CDEBUG(D_NET,
 390                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 391                libcfs_id2str(peer4), portal, matchbits);
 392         return 0;
 393 }
 394
 395 static int
 396 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 397                       int options, struct lnet_process_id peer,
 398                       lnet_nid_t self, struct lnet_handle_md *mdh,
 399                       struct srpc_event *ev)
 400 {
 401         int rc;
 402         struct lnet_md md;
 403
 404         md.user_ptr  = ev;
 405         md.start     = buf;
 406         md.length    = len;
 407         md.handler   = srpc_data.rpc_lnet_handler;
 408         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 409         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 410
 411         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
 412         if (rc != 0) {
 413                 CERROR("LNetMDBind failed: %d\n", rc);
 414                 LASSERT(rc == -ENOMEM);
 415                 return -ENOMEM;
 416         }
 417
 418         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 419          * they're only meaningful for MDs attached to an ME (i.e. passive
 420          * buffers...
 421          */
 422         if ((options & LNET_MD_OP_PUT) != 0) {
 423                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 424                              portal, matchbits, 0, 0);
 425         } else {
 426                 LASSERT((options & LNET_MD_OP_GET) != 0);
 427
 428                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0, false);
 429         }
 430
 431         if (rc != 0) {
 432                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 433                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 434                        libcfs_id2str(peer), portal, matchbits, rc);
 435
 436                 /* The forthcoming unlink event will complete this operation
 437                  * with failure, so fall through and return success here.
 438                  */
 439                 rc = LNetMDUnlink(*mdh);
 440                 LASSERT(rc == 0);
 441         } else {
 442                 CDEBUG(D_NET,
 443                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 444                        libcfs_id2str(peer), portal, matchbits);
 445         }
 446         return 0;
 447 }
 448
 449 static int
 450 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 451                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 452 {
 453         struct lnet_process_id any = {0};
 454
 455         any.nid = LNET_NID_ANY;
 456         any.pid = LNET_PID_ANY;
 457
 458         return srpc_post_passive_rdma(srpc_serv_portal(service),
 459                                       local, service, buf, len,
 460                                       LNET_MD_OP_PUT, any, mdh, ev);
 461 }
 462
 463 static int
 464 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 465 __must_hold(&scd->scd_lock)
 466 {
 467         struct srpc_service *sv = scd->scd_svc;
 468         struct srpc_msg *msg = &buf->buf_msg;
 469         int rc;
 470
 471         LNetInvalidateMDHandle(&buf->buf_mdh);
 472         list_add(&buf->buf_list, &scd->scd_buf_posted);
 473         scd->scd_buf_nposted++;
 474         spin_unlock(&scd->scd_lock);
 475
 476         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 477                                       !srpc_serv_is_framework(sv),
 478                                       msg, sizeof(*msg), &buf->buf_mdh,
 479                                       &scd->scd_ev);
 480
 481         /* At this point, a RPC (new or delayed) may have arrived in
 482          * msg and its event handler has been called. So we must add
 483          * buf to scd_buf_posted _before_ dropping scd_lock */
 484
 485         spin_lock(&scd->scd_lock);
 486
 487         if (rc == 0) {
 488                 if (!sv->sv_shuttingdown)
 489                         return 0;
 490
 491                 spin_unlock(&scd->scd_lock);
 492                 /* srpc_shutdown_service might have tried to unlink me
 493                  * when my buf_mdh was still invalid */
 494                 LNetMDUnlink(buf->buf_mdh);
 495                 spin_lock(&scd->scd_lock);
 496                 return 0;
 497         }
 498
 499         scd->scd_buf_nposted--;
 500         if (sv->sv_shuttingdown)
 501                 return rc; /* don't allow to change scd_buf_posted */
 502
 503         list_del(&buf->buf_list);
 504         spin_unlock(&scd->scd_lock);
 505
 506         LIBCFS_FREE(buf, sizeof(*buf));
 507
 508         spin_lock(&scd->scd_lock);
 509         return rc;
 510 }
 511
 512 int
 513 srpc_add_buffer(struct swi_workitem *wi)
 514 {
 515         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 516                                                    scd_buf_wi);
 517         struct srpc_buffer *buf;
 518         int rc = 0;
 519
 520         /* it's called by workitem scheduler threads, these threads
 521          * should have been set CPT affinity, so buffers will be posted
 522          * on CPT local list of Portal */
 523         spin_lock(&scd->scd_lock);
 524
 525         while (scd->scd_buf_adjust > 0 &&
 526                !scd->scd_svc->sv_shuttingdown) {
 527                 scd->scd_buf_adjust--; /* consume it */
 528                 scd->scd_buf_posting++;
 529
 530                 spin_unlock(&scd->scd_lock);
 531
 532                 LIBCFS_ALLOC(buf, sizeof(*buf));
 533                 if (buf == NULL) {
 534                         CERROR("Failed to add new buf to service: %s\n",
 535                                scd->scd_svc->sv_name);
 536                         spin_lock(&scd->scd_lock);
 537                         rc = -ENOMEM;
 538                         break;
 539                 }
 540
 541                 spin_lock(&scd->scd_lock);
 542                 if (scd->scd_svc->sv_shuttingdown) {
 543                         spin_unlock(&scd->scd_lock);
 544                         LIBCFS_FREE(buf, sizeof(*buf));
 545
 546                         spin_lock(&scd->scd_lock);
 547                         rc = -ESHUTDOWN;
 548                         break;
 549                 }
 550
 551                 rc = srpc_service_post_buffer(scd, buf);
 552                 if (rc != 0)
 553                         break; /* buf has been freed inside */
 554
 555                 LASSERT(scd->scd_buf_posting > 0);
 556                 scd->scd_buf_posting--;
 557                 scd->scd_buf_total++;
 558                 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
 559         }
 560
 561         if (rc != 0) {
 562                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 563                 scd->scd_buf_err = rc;
 564
 565                 LASSERT(scd->scd_buf_posting > 0);
 566                 scd->scd_buf_posting--;
 567         }
 568
 569         spin_unlock(&scd->scd_lock);
 570         return 0;
 571 }
 572
 573 int
 574 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 575 {
 576         struct srpc_service_cd *scd;
 577         int rc = 0;
 578         int i;
 579
 580         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 581
 582         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 583                 spin_lock(&scd->scd_lock);
 584
 585                 scd->scd_buf_err = 0;
 586                 scd->scd_buf_err_stamp = 0;
 587                 scd->scd_buf_posting = 0;
 588                 scd->scd_buf_adjust = nbuffer;
 589                 /* start to post buffers */
 590                 swi_schedule_workitem(&scd->scd_buf_wi);
 591                 spin_unlock(&scd->scd_lock);
 592
 593                 /* framework service only post buffer for one partition  */
 594                 if (srpc_serv_is_framework(sv))
 595                         break;
 596         }
 597
 598         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 599                 spin_lock(&scd->scd_lock);
 600                 /*
 601                  * NB: srpc_service_add_buffers() can be called inside
 602                  * thread context of lst_sched_serial, and we don't normally
 603                  * allow to sleep inside thread context of WI scheduler
 604                  * because it will block current scheduler thread from doing
 605                  * anything else, even worse, it could deadlock if it's
 606                  * waiting on result from another WI of the same scheduler.
 607                  * However, it's safe at here because scd_buf_wi is scheduled
 608                  * by thread in a different WI scheduler (lst_sched_test),
 609                  * so we don't have any risk of deadlock, though this could
 610                  * block all WIs pending on lst_sched_serial for a moment
 611                  * which is not good but not fatal.
 612                  */
 613                 lst_wait_until(scd->scd_buf_err != 0 ||
 614                                (scd->scd_buf_adjust == 0 &&
 615                                 scd->scd_buf_posting == 0),
 616                                scd->scd_lock, "waiting for adding buffer\n");
 617
 618                 if (scd->scd_buf_err != 0 && rc == 0)
 619                         rc = scd->scd_buf_err;
 620
 621                 spin_unlock(&scd->scd_lock);
 622         }
 623
 624         return rc;
 625 }
 626
 627 void
 628 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 629 {
 630         struct srpc_service_cd *scd;
 631         int num;
 632         int i;
 633
 634         LASSERT(!sv->sv_shuttingdown);
 635
 636         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 637                 spin_lock(&scd->scd_lock);
 638
 639                 num = scd->scd_buf_total + scd->scd_buf_posting;
 640                 scd->scd_buf_adjust -= min(nbuffer, num);
 641
 642                 spin_unlock(&scd->scd_lock);
 643         }
 644 }
 645
 646 /* returns 1 if sv has finished, otherwise 0 */
 647 int
 648 srpc_finish_service(struct srpc_service *sv)
 649 {
 650         struct srpc_service_cd *scd;
 651         struct srpc_server_rpc *rpc;
 652         int i;
 653
 654         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 655
 656         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 657                 spin_lock(&scd->scd_lock);
 658                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 659                         spin_unlock(&scd->scd_lock);
 660                         return 0;
 661                 }
 662
 663                 if (scd->scd_buf_nposted > 0) {
 664                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 665                                scd->scd_buf_nposted);
 666                         spin_unlock(&scd->scd_lock);
 667                         return 0;
 668                 }
 669
 670                 if (list_empty(&scd->scd_rpc_active)) {
 671                         spin_unlock(&scd->scd_lock);
 672                         continue;
 673                 }
 674
 675                 rpc = list_entry(scd->scd_rpc_active.next,
 676                                  struct srpc_server_rpc, srpc_list);
 677                 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",
 678                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 679                         swi_state2str(rpc->srpc_wi.swi_state),
 680                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 681                         rpc->srpc_wi.swi_workitem.wi_running,
 682                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 683                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 684                 spin_unlock(&scd->scd_lock);
 685                 return 0;
 686         }
 687
 688         /* no lock needed from now on */
 689         srpc_service_fini(sv);
 690         return 1;
 691 }
 692
 693 /* called with sv->sv_lock held */
 694 static void
 695 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 696                             struct srpc_buffer *buf)
 697 __must_hold(&scd->scd_lock)
 698 {
 699         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 700                 if (srpc_service_post_buffer(scd, buf) != 0) {
 701                         CWARN("Failed to post %s buffer\n",
 702                               scd->scd_svc->sv_name);
 703                 }
 704                 return;
 705         }
 706
 707         /* service is shutting down, or we want to recycle some buffers */
 708         scd->scd_buf_total--;
 709
 710         if (scd->scd_buf_adjust < 0) {
 711                 scd->scd_buf_adjust++;
 712                 if (scd->scd_buf_adjust < 0 &&
 713                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 714                         CDEBUG(D_INFO,
 715                                "Try to recyle %d buffers but nothing left\n",
 716                                scd->scd_buf_adjust);
 717                         scd->scd_buf_adjust = 0;
 718                 }
 719         }
 720
 721         spin_unlock(&scd->scd_lock);
 722         LIBCFS_FREE(buf, sizeof(*buf));
 723         spin_lock(&scd->scd_lock);
 724 }
 725
 726 void
 727 srpc_abort_service(struct srpc_service *sv)
 728 {
 729         struct srpc_service_cd *scd;
 730         struct srpc_server_rpc *rpc;
 731         int i;
 732
 733         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 734                sv->sv_id, sv->sv_name);
 735
 736         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 737                 spin_lock(&scd->scd_lock);
 738
 739                 /* schedule in-flight RPCs to notice the abort, NB:
 740                  * racing with incoming RPCs; complete fix should make test
 741                  * RPCs carry session ID in its headers
 742                  */
 743                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 744                         rpc->srpc_aborted = 1;
 745                         swi_schedule_workitem(&rpc->srpc_wi);
 746                 }
 747
 748                 spin_unlock(&scd->scd_lock);
 749         }
 750 }
 751
 752 void
 753 srpc_shutdown_service(struct srpc_service *sv)
 754 {
 755         struct srpc_service_cd *scd;
 756         struct srpc_server_rpc *rpc;
 757         struct srpc_buffer *buf;
 758         int i;
 759
 760         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 761                sv->sv_id, sv->sv_name);
 762
 763         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 764                 spin_lock(&scd->scd_lock);
 765
 766         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 767
 768         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 769                 spin_unlock(&scd->scd_lock);
 770
 771         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 772                 spin_lock(&scd->scd_lock);
 773
 774                 /* schedule in-flight RPCs to notice the shutdown */
 775                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 776                         swi_schedule_workitem(&rpc->srpc_wi);
 777
 778                 spin_unlock(&scd->scd_lock);
 779
 780                 /* OK to traverse scd_buf_posted without lock, since no one
 781                  * touches scd_buf_posted now
 782                  */
 783                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 784                         LNetMDUnlink(buf->buf_mdh);
 785         }
 786 }
 787
 788 static int
 789 srpc_send_request(struct srpc_client_rpc *rpc)
 790 {
 791         struct srpc_event *ev = &rpc->crpc_reqstev;
 792         int rc;
 793
 794         ev->ev_fired = 0;
 795         ev->ev_data  = rpc;
 796         ev->ev_type  = SRPC_REQUEST_SENT;
 797
 798         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 799                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 800                                    sizeof(struct srpc_msg), LNET_MD_OP_PUT,
 801                                    rpc->crpc_dest, LNET_NID_ANY,
 802                                    &rpc->crpc_reqstmdh, ev);
 803         if (rc != 0) {
 804                 LASSERT(rc == -ENOMEM);
 805                 ev->ev_fired = 1;  /* no more event expected */
 806         }
 807         return rc;
 808 }
 809
 810 static int
 811 srpc_prepare_reply(struct srpc_client_rpc *rpc)
 812 {
 813         struct srpc_event *ev = &rpc->crpc_replyev;
 814         u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 815         int rc;
 816
 817         ev->ev_fired = 0;
 818         ev->ev_data  = rpc;
 819         ev->ev_type  = SRPC_REPLY_RCVD;
 820
 821         *id = srpc_next_id();
 822
 823         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 824                                     &rpc->crpc_replymsg,
 825                                     sizeof(struct srpc_msg),
 826                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 827                                     &rpc->crpc_replymdh, ev);
 828         if (rc != 0) {
 829                 LASSERT(rc == -ENOMEM);
 830                 ev->ev_fired = 1;  /* no more event expected */
 831         }
 832         return rc;
 833 }
 834
 835 static int
 836 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 837 {
 838         struct srpc_bulk *bk = &rpc->crpc_bulk;
 839         struct srpc_event *ev = &rpc->crpc_bulkev;
 840         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 841         int rc;
 842         int opt;
 843
 844         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 845
 846         /* nothing to do */
 847         if (bk->bk_niov == 0)
 848                 return 0;
 849
 850         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 851         opt |= LNET_MD_KIOV;
 852
 853         ev->ev_fired = 0;
 854         ev->ev_data  = rpc;
 855         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 856
 857         *id = srpc_next_id();
 858
 859         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 860                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 861                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 862         if (rc != 0) {
 863                 LASSERT(rc == -ENOMEM);
 864                 ev->ev_fired = 1;  /* no more event expected */
 865         }
 866         return rc;
 867 }
 868
 869 static int
 870 srpc_do_bulk(struct srpc_server_rpc *rpc)
 871 {
 872         struct srpc_event *ev = &rpc->srpc_ev;
 873         struct srpc_bulk *bk = rpc->srpc_bulk;
 874         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 875         int rc;
 876         int opt;
 877
 878         LASSERT(bk != NULL);
 879
 880         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 881         opt |= LNET_MD_KIOV;
 882
 883         ev->ev_fired = 0;
 884         ev->ev_data  = rpc;
 885         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 886
 887         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 888                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 889                                    rpc->srpc_peer, rpc->srpc_self,
 890                                    &bk->bk_mdh, ev);
 891         if (rc != 0)
 892                 ev->ev_fired = 1;  /* no more event expected */
 893         return rc;
 894 }
 895
 896 /* only called from srpc_handle_rpc */
 897 static void
 898 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 899 {
 900         struct srpc_service_cd  *scd = rpc->srpc_scd;
 901         struct srpc_service     *sv  = scd->scd_svc;
 902         struct srpc_buffer *buffer;
 903
 904         LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 905
 906         rpc->srpc_status = status;
 907
 908         CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
 909                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 910                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 911                      swi_state2str(rpc->srpc_wi.swi_state), status);
 912
 913         if (status != 0) {
 914                 spin_lock(&srpc_data.rpc_glock);
 915                 srpc_data.rpc_counters.rpcs_dropped++;
 916                 spin_unlock(&srpc_data.rpc_glock);
 917         }
 918
 919         if (rpc->srpc_done != NULL)
 920                 (*rpc->srpc_done) (rpc);
 921         LASSERT(rpc->srpc_bulk == NULL);
 922
 923         spin_lock(&scd->scd_lock);
 924
 925         if (rpc->srpc_reqstbuf != NULL) {
 926                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 927                  * sv won't go away for scd_rpc_active must not be empty
 928                  */
 929                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 930                 rpc->srpc_reqstbuf = NULL;
 931         }
 932
 933         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 934
 935         /*
 936          * No one can schedule me now since:
 937          * - I'm not on scd_rpc_active.
 938          * - all LNet events have been fired.
 939          * Cancel pending schedules and prevent future schedule attempts:
 940          */
 941         LASSERT(rpc->srpc_ev.ev_fired);
 942         swi_exit_workitem(&rpc->srpc_wi);
 943
 944         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 945                 buffer = list_entry(scd->scd_buf_blocked.next,
 946                                     struct srpc_buffer, buf_list);
 947                 list_del(&buffer->buf_list);
 948
 949                 srpc_init_server_rpc(rpc, scd, buffer);
 950                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 951                 swi_schedule_workitem(&rpc->srpc_wi);
 952         } else {
 953                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 954         }
 955
 956         spin_unlock(&scd->scd_lock);
 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 }
1109
1110 /*
1111  * Called with rpc->crpc_lock held.
1112  *
1113  * Upon exit the RPC expiry timer is not queued and the handler is not
1114  * running on any CPU.
1115  */
1116 static void
1117 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1118 {
1119         /* timer not planted or already exploded */
1120         if (rpc->crpc_timeout == 0)
1121                 return;
1122
1123         /* timer successfully defused */
1124         if (stt_del_timer(&rpc->crpc_timer))
1125                 return;
1126
1127         /* timer detonated, wait for it to explode */
1128         while (rpc->crpc_timeout != 0) {
1129                 spin_unlock(&rpc->crpc_lock);
1130
1131                 schedule();
1132
1133                 spin_lock(&rpc->crpc_lock);
1134         }
1135 }
1136
1137 static void
1138 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1139 {
1140         struct swi_workitem *wi = &rpc->crpc_wi;
1141
1142         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1143
1144         spin_lock(&rpc->crpc_lock);
1145
1146         rpc->crpc_closed = 1;
1147         if (rpc->crpc_status == 0)
1148                 rpc->crpc_status = status;
1149
1150         srpc_del_client_rpc_timer(rpc);
1151
1152         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1153                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1154                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1155                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1156
1157         /*
1158          * No one can schedule me now since:
1159          * - RPC timer has been defused.
1160          * - all LNet events have been fired.
1161          * - crpc_closed has been set, preventing srpc_abort_rpc from
1162          *   scheduling me.
1163          * Cancel pending schedules and prevent future schedule attempts:
1164          */
1165         LASSERT(!srpc_event_pending(rpc));
1166         swi_exit_workitem(wi);
1167
1168         spin_unlock(&rpc->crpc_lock);
1169
1170         (*rpc->crpc_done)(rpc);
1171 }
1172
1173 /* sends an outgoing RPC */
1174 int
1175 srpc_send_rpc(struct swi_workitem *wi)
1176 {
1177         int rc = 0;
1178         struct srpc_client_rpc *rpc;
1179         struct srpc_msg *reply;
1180         int do_bulk;
1181
1182         LASSERT(wi != NULL);
1183
1184         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1185
1186         LASSERT(rpc != NULL);
1187         LASSERT(wi == &rpc->crpc_wi);
1188
1189         reply = &rpc->crpc_replymsg;
1190         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1191
1192         spin_lock(&rpc->crpc_lock);
1193
1194         if (rpc->crpc_aborted) {
1195                 spin_unlock(&rpc->crpc_lock);
1196                 goto abort;
1197         }
1198
1199         spin_unlock(&rpc->crpc_lock);
1200
1201         switch (wi->swi_state) {
1202         default:
1203                 LBUG();
1204         case SWI_STATE_NEWBORN:
1205                 LASSERT(!srpc_event_pending(rpc));
1206
1207                 rc = srpc_prepare_reply(rpc);
1208                 if (rc != 0) {
1209                         srpc_client_rpc_done(rpc, rc);
1210                         return 1;
1211                 }
1212
1213                 rc = srpc_prepare_bulk(rpc);
1214                 if (rc != 0)
1215                         break;
1216
1217                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1218                 rc = srpc_send_request(rpc);
1219                 break;
1220
1221         case SWI_STATE_REQUEST_SUBMITTED:
1222                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1223                  * order; however, they're processed in a strict order:
1224                  * rqt, rpy, and bulk.
1225                  */
1226                 if (!rpc->crpc_reqstev.ev_fired)
1227                         break;
1228
1229                 rc = rpc->crpc_reqstev.ev_status;
1230                 if (rc != 0)
1231                         break;
1232
1233                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1234                 /* fallthrough */
1235         case SWI_STATE_REQUEST_SENT: {
1236                 enum srpc_msg_type type;
1237
1238                 type = srpc_service2reply(rpc->crpc_service);
1239
1240                 if (!rpc->crpc_replyev.ev_fired)
1241                         break;
1242
1243                 rc = rpc->crpc_replyev.ev_status;
1244                 if (rc != 0)
1245                         break;
1246
1247                 srpc_unpack_msg_hdr(reply);
1248                 if (reply->msg_type != type ||
1249                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1250                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1251                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1252                               libcfs_id2str(rpc->crpc_dest),
1253                               reply->msg_type, type,
1254                               reply->msg_magic, SRPC_MSG_MAGIC);
1255                         rc = -EBADMSG;
1256                         break;
1257                 }
1258
1259                 if (do_bulk && reply->msg_body.reply.status != 0) {
1260                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1261                               reply->msg_body.reply.status,
1262                               libcfs_id2str(rpc->crpc_dest));
1263                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1264                 }
1265
1266                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1267         }
1268         /* fallthrough */
1269         case SWI_STATE_REPLY_RECEIVED:
1270                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1271                         break;
1272
1273                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1274
1275                 /* Bulk buffer was unlinked due to remote error. Clear error
1276                  * since reply buffer still contains valid data.
1277                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1278                  * remote error.
1279                  */
1280                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1281                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1282                         rc = 0;
1283
1284                 wi->swi_state = SWI_STATE_DONE;
1285                 srpc_client_rpc_done(rpc, rc);
1286                 return 1;
1287         }
1288
1289         if (rc != 0) {
1290                 spin_lock(&rpc->crpc_lock);
1291                 srpc_abort_rpc(rpc, rc);
1292                 spin_unlock(&rpc->crpc_lock);
1293         }
1294
1295 abort:
1296         if (rpc->crpc_aborted) {
1297                 LNetMDUnlink(rpc->crpc_reqstmdh);
1298                 LNetMDUnlink(rpc->crpc_replymdh);
1299                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1300
1301                 if (!srpc_event_pending(rpc)) {
1302                         srpc_client_rpc_done(rpc, -EINTR);
1303                         return 1;
1304                 }
1305         }
1306         return 0;
1307 }
1308
1309 struct srpc_client_rpc *
1310 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1311                        int nbulkiov, int bulklen,
1312                        void (*rpc_done)(struct srpc_client_rpc *),
1313                        void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1314 {
1315         struct srpc_client_rpc *rpc;
1316
1317         LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1318                                    crpc_bulk.bk_iovs[nbulkiov]));
1319         if (rpc == NULL)
1320                 return NULL;
1321
1322         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1323                              bulklen, rpc_done, rpc_fini, priv);
1324         return rpc;
1325 }
1326
1327 /* called with rpc->crpc_lock held */
1328 void
1329 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1330 {
1331         LASSERT(why != 0);
1332
1333         if (rpc->crpc_aborted || /* already aborted */
1334             rpc->crpc_closed)    /* callback imminent */
1335                 return;
1336
1337         CDEBUG(D_NET,
1338                "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1339                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1340                swi_state2str(rpc->crpc_wi.swi_state), why);
1341
1342         rpc->crpc_aborted = 1;
1343         rpc->crpc_status  = why;
1344         swi_schedule_workitem(&rpc->crpc_wi);
1345 }
1346
1347 /* called with rpc->crpc_lock held */
1348 void
1349 srpc_post_rpc(struct srpc_client_rpc *rpc)
1350 {
1351         LASSERT(!rpc->crpc_aborted);
1352         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1353
1354         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1355                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1356                rpc->crpc_timeout);
1357
1358         srpc_add_client_rpc_timer(rpc);
1359         swi_schedule_workitem(&rpc->crpc_wi);
1360 }
1361
1362
1363 int
1364 srpc_send_reply(struct srpc_server_rpc *rpc)
1365 {
1366         struct srpc_event *ev = &rpc->srpc_ev;
1367         struct srpc_msg *msg = &rpc->srpc_replymsg;
1368         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1369         struct srpc_service_cd *scd = rpc->srpc_scd;
1370         struct srpc_service *sv = scd->scd_svc;
1371         __u64 rpyid;
1372         int rc;
1373
1374         LASSERT(buffer != NULL);
1375         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1376
1377         spin_lock(&scd->scd_lock);
1378
1379         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1380                 /* Repost buffer before replying since test client
1381                  * might send me another RPC once it gets the reply
1382                  */
1383                 if (srpc_service_post_buffer(scd, buffer) != 0)
1384                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1385                 rpc->srpc_reqstbuf = NULL;
1386         }
1387
1388         spin_unlock(&scd->scd_lock);
1389
1390         ev->ev_fired = 0;
1391         ev->ev_data  = rpc;
1392         ev->ev_type  = SRPC_REPLY_SENT;
1393
1394         msg->msg_magic   = SRPC_MSG_MAGIC;
1395         msg->msg_version = SRPC_MSG_VERSION;
1396         msg->msg_type    = srpc_service2reply(sv->sv_id);
1397
1398         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1399                                    sizeof(*msg), LNET_MD_OP_PUT,
1400                                    rpc->srpc_peer, rpc->srpc_self,
1401                                    &rpc->srpc_replymdh, ev);
1402         if (rc != 0)
1403                 ev->ev_fired = 1;  /* no more event expected */
1404         return rc;
1405 }
1406
1407 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1408 static void
1409 srpc_lnet_ev_handler(struct lnet_event *ev)
1410 {
1411         struct srpc_service_cd *scd;
1412         struct srpc_event *rpcev = ev->md_user_ptr;
1413         struct srpc_client_rpc *crpc;
1414         struct srpc_server_rpc *srpc;
1415         struct srpc_buffer *buffer;
1416         struct srpc_service *sv;
1417         struct srpc_msg *msg;
1418         enum srpc_msg_type type;
1419
1420         LASSERT(!in_interrupt());
1421
1422         if (ev->status != 0) {
1423                 __u32 errors;
1424
1425                 spin_lock(&srpc_data.rpc_glock);
1426                 if (ev->status != -ECANCELED) /* cancellation is not error */
1427                         srpc_data.rpc_counters.errors++;
1428                 errors = srpc_data.rpc_counters.errors;
1429                 spin_unlock(&srpc_data.rpc_glock);
1430
1431                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1432                         ev->status, ev->type, errors);
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                 /* fallthrough */
1443         case SRPC_REQUEST_SENT:
1444                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1445                         spin_lock(&srpc_data.rpc_glock);
1446                         srpc_data.rpc_counters.rpcs_sent++;
1447                         spin_unlock(&srpc_data.rpc_glock);
1448                 }
1449                 /* fallthrough */
1450         case SRPC_REPLY_RCVD:
1451         case SRPC_BULK_REQ_RCVD:
1452                 crpc = rpcev->ev_data;
1453
1454                 if (rpcev != &crpc->crpc_reqstev &&
1455                     rpcev != &crpc->crpc_replyev &&
1456                     rpcev != &crpc->crpc_bulkev) {
1457                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1458                                rpcev, crpc, &crpc->crpc_reqstev,
1459                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1460                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1461                                rpcev->ev_status, rpcev->ev_type,
1462                                rpcev->ev_lnet);
1463                         LBUG();
1464                 }
1465
1466                 spin_lock(&crpc->crpc_lock);
1467
1468                 LASSERT(rpcev->ev_fired == 0);
1469                 rpcev->ev_fired  = 1;
1470                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1471                                    -EINTR : ev->status;
1472                 swi_schedule_workitem(&crpc->crpc_wi);
1473
1474                 spin_unlock(&crpc->crpc_lock);
1475                 break;
1476
1477         case SRPC_REQUEST_RCVD:
1478                 scd = rpcev->ev_data;
1479                 sv = scd->scd_svc;
1480
1481                 LASSERT(rpcev == &scd->scd_ev);
1482
1483                 spin_lock(&scd->scd_lock);
1484
1485                 LASSERT(ev->unlinked);
1486                 LASSERT(ev->type == LNET_EVENT_PUT ||
1487                         ev->type == LNET_EVENT_UNLINK);
1488                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1489                         sv->sv_shuttingdown);
1490
1491                 buffer = container_of(ev->md_start, struct srpc_buffer,
1492                                       buf_msg);
1493                 buffer->buf_peer = ev->source;
1494                 buffer->buf_self = ev->target.nid;
1495
1496                 LASSERT(scd->scd_buf_nposted > 0);
1497                 scd->scd_buf_nposted--;
1498
1499                 if (sv->sv_shuttingdown) {
1500                         /* Leave buffer on scd->scd_buf_nposted since
1501                          * srpc_finish_service needs to traverse it.
1502                          */
1503                         spin_unlock(&scd->scd_lock);
1504                         break;
1505                 }
1506
1507                 if (scd->scd_buf_err_stamp != 0 &&
1508                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1509                         /* re-enable adding buffer */
1510                         scd->scd_buf_err_stamp = 0;
1511                         scd->scd_buf_err = 0;
1512                 }
1513
1514                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1515                     scd->scd_buf_adjust == 0 &&
1516                     scd->scd_buf_nposted < scd->scd_buf_low) {
1517                         scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1518                                                   SFW_TEST_WI_MIN);
1519                         swi_schedule_workitem(&scd->scd_buf_wi);
1520                 }
1521
1522                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1523                 msg = &buffer->buf_msg;
1524                 type = srpc_service2request(sv->sv_id);
1525
1526                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1527                     (msg->msg_type != type &&
1528                      msg->msg_type != __swab32(type)) ||
1529                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1530                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1531                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1532                                sv->sv_name, libcfs_id2str(ev->initiator),
1533                                ev->status, ev->mlength,
1534                                msg->msg_type, msg->msg_magic);
1535
1536                         /* NB can't call srpc_service_recycle_buffer here since
1537                          * it may call LNetM[DE]Attach. The invalid magic tells
1538                          * srpc_handle_rpc to drop this RPC
1539                          */
1540                         msg->msg_magic = 0;
1541                 }
1542
1543                 if (!list_empty(&scd->scd_rpc_free)) {
1544                         srpc = list_entry(scd->scd_rpc_free.next,
1545                                           struct srpc_server_rpc,
1546                                           srpc_list);
1547                         list_del(&srpc->srpc_list);
1548
1549                         srpc_init_server_rpc(srpc, scd, buffer);
1550                         list_add_tail(&srpc->srpc_list,
1551                                       &scd->scd_rpc_active);
1552                         swi_schedule_workitem(&srpc->srpc_wi);
1553                 } else {
1554                         list_add_tail(&buffer->buf_list,
1555                                       &scd->scd_buf_blocked);
1556                 }
1557
1558                 spin_unlock(&scd->scd_lock);
1559
1560                 spin_lock(&srpc_data.rpc_glock);
1561                 srpc_data.rpc_counters.rpcs_rcvd++;
1562                 spin_unlock(&srpc_data.rpc_glock);
1563                 break;
1564
1565         case SRPC_BULK_GET_RPLD:
1566                 LASSERT(ev->type == LNET_EVENT_SEND ||
1567                         ev->type == LNET_EVENT_REPLY ||
1568                         ev->type == LNET_EVENT_UNLINK);
1569
1570                 if (!ev->unlinked)
1571                         break; /* wait for final event */
1572                 /* fallthrough */
1573         case SRPC_BULK_PUT_SENT:
1574                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1575                         spin_lock(&srpc_data.rpc_glock);
1576
1577                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1578                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1579                         else
1580                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1581
1582                         spin_unlock(&srpc_data.rpc_glock);
1583                 }
1584                 /* fallthrough */
1585         case SRPC_REPLY_SENT:
1586                 srpc = rpcev->ev_data;
1587                 scd  = srpc->srpc_scd;
1588
1589                 LASSERT(rpcev == &srpc->srpc_ev);
1590
1591                 spin_lock(&scd->scd_lock);
1592
1593                 rpcev->ev_fired  = 1;
1594                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1595                                    -EINTR : ev->status;
1596                 swi_schedule_workitem(&srpc->srpc_wi);
1597
1598                 spin_unlock(&scd->scd_lock);
1599                 break;
1600         }
1601 }
1602
1603
1604 int
1605 srpc_startup (void)
1606 {
1607         int rc;
1608
1609         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1610         spin_lock_init(&srpc_data.rpc_glock);
1611
1612         /* 1 second pause to avoid timestamp reuse */
1613         schedule_timeout_uninterruptible(cfs_time_seconds(1));
1614         srpc_data.rpc_matchbits = ((__u64) ktime_get_real_seconds()) << 48;
1615
1616         srpc_data.rpc_state = SRPC_STATE_NONE;
1617
1618         rc = LNetNIInit(LNET_PID_LUSTRE);
1619         if (rc < 0) {
1620                 CERROR("LNetNIInit() has failed: %d\n", rc);
1621                 return rc;
1622         }
1623
1624         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1625
1626         srpc_data.rpc_lnet_handler = srpc_lnet_ev_handler;
1627
1628         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1629         LASSERT(rc == 0);
1630         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1631         LASSERT(rc == 0);
1632
1633         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1634
1635         rc = stt_startup();
1636
1637         if (rc != 0)
1638                 srpc_shutdown();
1639         else
1640                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1641
1642         return rc;
1643 }
1644
1645 void
1646 srpc_shutdown (void)
1647 {
1648         int i;
1649         int rc;
1650         int state;
1651
1652         state = srpc_data.rpc_state;
1653         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1654
1655         switch (state) {
1656         default:
1657                 LBUG();
1658                 /* fallthrough */
1659         case SRPC_STATE_RUNNING:
1660                 spin_lock(&srpc_data.rpc_glock);
1661
1662                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1663                         struct srpc_service *sv = srpc_data.rpc_services[i];
1664
1665                         LASSERTF(sv == NULL,
1666                                  "service not empty: id %d, name %s\n",
1667                                  i, sv->sv_name);
1668                 }
1669
1670                 spin_unlock(&srpc_data.rpc_glock);
1671
1672                 stt_shutdown();
1673                 /* fallthrough */
1674
1675         case SRPC_STATE_EQ_INIT:
1676                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1677                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1678                 LASSERT(rc == 0);
1679                 lnet_assert_handler_unused(srpc_data.rpc_lnet_handler);
1680                 /* fallthrough */
1681
1682         case SRPC_STATE_NI_INIT:
1683                 LNetNIFini();
1684         }
1685 }