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