lnet/selftest/rpc.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/selftest/rpc.c
  33  *
  34  * Author: Isaac Huang <isaac@clusterfs.com>
  35  *
  36  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  37  * - percpt data for service to improve smp performance
  38  * - code cleanup
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include "selftest.h"
  44
  45 enum srpc_state {
  46         SRPC_STATE_NONE,
  47         SRPC_STATE_NI_INIT,
  48         SRPC_STATE_EQ_INIT,
  49         SRPC_STATE_RUNNING,
  50         SRPC_STATE_STOPPING,
  51 };
  52
  53 static struct smoketest_rpc {
  54         spinlock_t       rpc_glock;     /* global lock */
  55         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  56         struct lnet_handle_eq    rpc_lnet_eq;   /* _the_ LNet event queue */
  57         enum srpc_state          rpc_state;
  58         struct srpc_counters     rpc_counters;
  59         __u64                    rpc_matchbits; /* matchbits counter */
  60 } srpc_data;
  61
  62 static inline int
  63 srpc_serv_portal(int svc_id)
  64 {
  65         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  66                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  67 }
  68
  69 /* forward ref's */
  70 static int srpc_handle_rpc(struct swi_workitem *wi);
  71
  72 void srpc_get_counters(struct srpc_counters *cnt)
  73 {
  74         spin_lock(&srpc_data.rpc_glock);
  75         *cnt = srpc_data.rpc_counters;
  76         spin_unlock(&srpc_data.rpc_glock);
  77 }
  78
  79 void srpc_set_counters(const struct srpc_counters *cnt)
  80 {
  81         spin_lock(&srpc_data.rpc_glock);
  82         srpc_data.rpc_counters = *cnt;
  83         spin_unlock(&srpc_data.rpc_glock);
  84 }
  85
  86 static int
  87 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
  88                    int nob)
  89 {
  90         LASSERT(off < PAGE_SIZE);
  91         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  92
  93         bk->bk_iovs[i].kiov_offset = off;
  94         bk->bk_iovs[i].kiov_page   = pg;
  95         bk->bk_iovs[i].kiov_len    = nob;
  96         return nob;
  97 }
  98
  99 void
 100 srpc_free_bulk(struct srpc_bulk *bk)
 101 {
 102         int i;
 103         struct page *pg;
 104
 105         LASSERT(bk != NULL);
 106
 107         for (i = 0; i < bk->bk_niov; i++) {
 108                 pg = bk->bk_iovs[i].kiov_page;
 109                 if (pg == NULL)
 110                         break;
 111
 112                 __free_page(pg);
 113         }
 114
 115         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
 116 }
 117
 118 struct srpc_bulk *
 119 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 120                 unsigned bulk_len, int sink)
 121 {
 122         struct srpc_bulk *bk;
 123         int i;
 124
 125         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 126
 127         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 128                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 129         if (bk == NULL) {
 130                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 131                 return NULL;
 132         }
 133
 134         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 135         bk->bk_sink   = sink;
 136         bk->bk_len    = bulk_len;
 137         bk->bk_niov   = bulk_npg;
 138
 139         for (i = 0; i < bulk_npg; i++) {
 140                 struct page *pg;
 141                 int nob;
 142
 143                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 144                 if (pg == NULL) {
 145                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 146                         srpc_free_bulk(bk);
 147                         return NULL;
 148                 }
 149
 150                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 151                       bulk_off;
 152
 153                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 154                 bulk_len -= nob;
 155                 bulk_off = 0;
 156         }
 157
 158         return bk;
 159 }
 160
 161 static inline __u64
 162 srpc_next_id (void)
 163 {
 164         __u64 id;
 165
 166         spin_lock(&srpc_data.rpc_glock);
 167         id = srpc_data.rpc_matchbits++;
 168         spin_unlock(&srpc_data.rpc_glock);
 169         return id;
 170 }
 171
 172 static void
 173 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 174                      struct srpc_service_cd *scd,
 175                      struct srpc_buffer *buffer)
 176 {
 177         memset(rpc, 0, sizeof(*rpc));
 178         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 179                           srpc_serv_is_framework(scd->scd_svc) ?
 180                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 181
 182         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 183
 184         rpc->srpc_scd      = scd;
 185         rpc->srpc_reqstbuf = buffer;
 186         rpc->srpc_peer     = buffer->buf_peer;
 187         rpc->srpc_self     = buffer->buf_self;
 188         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 189 }
 190
 191 static void
 192 srpc_service_fini(struct srpc_service *svc)
 193 {
 194         struct srpc_service_cd *scd;
 195         struct srpc_server_rpc *rpc;
 196         struct srpc_buffer *buf;
 197         struct list_head *q;
 198         int i;
 199
 200         if (svc->sv_cpt_data == NULL)
 201                 return;
 202
 203         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 204                 while (1) {
 205                         if (!list_empty(&scd->scd_buf_posted))
 206                                 q = &scd->scd_buf_posted;
 207                         else if (!list_empty(&scd->scd_buf_blocked))
 208                                 q = &scd->scd_buf_blocked;
 209                         else
 210                                 break;
 211
 212                         while (!list_empty(q)) {
 213                                 buf = list_entry(q->next,
 214                                                  struct srpc_buffer,
 215                                                  buf_list);
 216                                 list_del(&buf->buf_list);
 217                                 LIBCFS_FREE(buf, sizeof(*buf));
 218                         }
 219                 }
 220
 221                 LASSERT(list_empty(&scd->scd_rpc_active));
 222
 223                 while (!list_empty(&scd->scd_rpc_free)) {
 224                         rpc = list_entry(scd->scd_rpc_free.next,
 225                                          struct srpc_server_rpc,
 226                                          srpc_list);
 227                         list_del(&rpc->srpc_list);
 228                         LIBCFS_FREE(rpc, sizeof(*rpc));
 229                 }
 230         }
 231
 232         cfs_percpt_free(svc->sv_cpt_data);
 233         svc->sv_cpt_data = NULL;
 234 }
 235
 236 static int
 237 srpc_service_nrpcs(struct srpc_service *svc)
 238 {
 239         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 240
 241         return srpc_serv_is_framework(svc) ?
 242                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 243 }
 244
 245 int srpc_add_buffer(struct swi_workitem *wi);
 246
 247 static int
 248 srpc_service_init(struct srpc_service *svc)
 249 {
 250         struct srpc_service_cd *scd;
 251         struct srpc_server_rpc *rpc;
 252         int nrpcs;
 253         int i;
 254         int j;
 255
 256         svc->sv_shuttingdown = 0;
 257
 258         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 259                                             sizeof(struct srpc_service_cd));
 260         if (svc->sv_cpt_data == NULL)
 261                 return -ENOMEM;
 262
 263         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 264                         1 : cfs_cpt_number(lnet_cpt_table());
 265         nrpcs = srpc_service_nrpcs(svc);
 266
 267         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 268                 scd->scd_cpt = i;
 269                 scd->scd_svc = svc;
 270                 spin_lock_init(&scd->scd_lock);
 271                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 272                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 273                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 274                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 275
 276                 scd->scd_ev.ev_data = scd;
 277                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 278
 279                 /* NB: don't use lst_sched_serial for adding buffer,
 280                  * see details in srpc_service_add_buffers() */
 281                 swi_init_workitem(&scd->scd_buf_wi,
 282                                   srpc_add_buffer, lst_sched_test[i]);
 283
 284                 if (i != 0 && srpc_serv_is_framework(svc)) {
 285                         /* NB: framework service only needs srpc_service_cd for
 286                          * one partition, but we allocate for all to make
 287                          * it easier to implement, it will waste a little
 288                          * memory but nobody should care about this */
 289                         continue;
 290                 }
 291
 292                 for (j = 0; j < nrpcs; j++) {
 293                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 294                                          i, sizeof(*rpc));
 295                         if (rpc == NULL) {
 296                                 srpc_service_fini(svc);
 297                                 return -ENOMEM;
 298                         }
 299                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 300                 }
 301         }
 302
 303         return 0;
 304 }
 305
 306 int
 307 srpc_add_service(struct srpc_service *sv)
 308 {
 309         int id = sv->sv_id;
 310
 311         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 312
 313         if (srpc_service_init(sv) != 0)
 314                 return -ENOMEM;
 315
 316         spin_lock(&srpc_data.rpc_glock);
 317
 318         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 319
 320         if (srpc_data.rpc_services[id] != NULL) {
 321                 spin_unlock(&srpc_data.rpc_glock);
 322                 goto failed;
 323         }
 324
 325         srpc_data.rpc_services[id] = sv;
 326         spin_unlock(&srpc_data.rpc_glock);
 327
 328         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 329         return 0;
 330
 331 failed:
 332         srpc_service_fini(sv);
 333         return -EBUSY;
 334 }
 335
 336 int
 337 srpc_remove_service(struct srpc_service *sv)
 338 {
 339         int id = sv->sv_id;
 340
 341         spin_lock(&srpc_data.rpc_glock);
 342
 343         if (srpc_data.rpc_services[id] != sv) {
 344                 spin_unlock(&srpc_data.rpc_glock);
 345                 return -ENOENT;
 346         }
 347
 348         srpc_data.rpc_services[id] = NULL;
 349         spin_unlock(&srpc_data.rpc_glock);
 350         return 0;
 351 }
 352
 353 static int
 354 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 355                        int len, int options, struct lnet_process_id peer,
 356                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 357 {
 358         int rc;
 359         struct lnet_md md;
 360         struct lnet_me *me;
 361
 362         me = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 363                           local ? LNET_INS_LOCAL : LNET_INS_AFTER);
 364         if (IS_ERR(me)) {
 365                 rc = PTR_ERR(me);
 366                 CERROR("LNetMEAttach failed: %d\n", rc);
 367                 LASSERT(rc == -ENOMEM);
 368                 return -ENOMEM;
 369         }
 370
 371         md.threshold = 1;
 372         md.user_ptr  = ev;
 373         md.start     = buf;
 374         md.length    = len;
 375         md.options   = options;
 376         md.eq_handle = srpc_data.rpc_lnet_eq;
 377
 378         rc = LNetMDAttach(me, md, LNET_UNLINK, mdh);
 379         if (rc != 0) {
 380                 CERROR("LNetMDAttach failed: %d\n", rc);
 381                 LASSERT(rc == -ENOMEM);
 382
 383                 LNetMEUnlink(me);
 384                 return -ENOMEM;
 385         }
 386
 387         CDEBUG(D_NET,
 388                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 389                libcfs_id2str(peer), portal, matchbits);
 390         return 0;
 391 }
 392
 393 static int
 394 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 395                       int options, struct lnet_process_id peer,
 396                       lnet_nid_t self, struct lnet_handle_md *mdh,
 397                       struct srpc_event *ev)
 398 {
 399         int rc;
 400         struct lnet_md md;
 401
 402         md.user_ptr  = ev;
 403         md.start     = buf;
 404         md.length    = len;
 405         md.eq_handle = srpc_data.rpc_lnet_eq;
 406         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 407         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 408
 409         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 410         if (rc != 0) {
 411                 CERROR("LNetMDBind failed: %d\n", rc);
 412                 LASSERT(rc == -ENOMEM);
 413                 return -ENOMEM;
 414         }
 415
 416         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 417          * they're only meaningful for MDs attached to an ME (i.e. passive
 418          * buffers...
 419          */
 420         if ((options & LNET_MD_OP_PUT) != 0) {
 421                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 422                              portal, matchbits, 0, 0);
 423         } else {
 424                 LASSERT((options & LNET_MD_OP_GET) != 0);
 425
 426                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0, false);
 427         }
 428
 429         if (rc != 0) {
 430                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 431                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 432                        libcfs_id2str(peer), portal, matchbits, rc);
 433
 434                 /* The forthcoming unlink event will complete this operation
 435                  * with failure, so fall through and return success here.
 436                  */
 437                 rc = LNetMDUnlink(*mdh);
 438                 LASSERT(rc == 0);
 439         } else {
 440                 CDEBUG(D_NET,
 441                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 442                        libcfs_id2str(peer), portal, matchbits);
 443         }
 444         return 0;
 445 }
 446
 447 static int
 448 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 449                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 450 {
 451         struct lnet_process_id any = {0};
 452
 453         any.nid = LNET_NID_ANY;
 454         any.pid = LNET_PID_ANY;
 455
 456         return srpc_post_passive_rdma(srpc_serv_portal(service),
 457                                       local, service, buf, len,
 458                                       LNET_MD_OP_PUT, any, mdh, ev);
 459 }
 460
 461 static int
 462 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 463 __must_hold(&scd->scd_lock)
 464 {
 465         struct srpc_service *sv = scd->scd_svc;
 466         struct srpc_msg *msg = &buf->buf_msg;
 467         int rc;
 468
 469         LNetInvalidateMDHandle(&buf->buf_mdh);
 470         list_add(&buf->buf_list, &scd->scd_buf_posted);
 471         scd->scd_buf_nposted++;
 472         spin_unlock(&scd->scd_lock);
 473
 474         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 475                                       !srpc_serv_is_framework(sv),
 476                                       msg, sizeof(*msg), &buf->buf_mdh,
 477                                       &scd->scd_ev);
 478
 479         /* At this point, a RPC (new or delayed) may have arrived in
 480          * msg and its event handler has been called. So we must add
 481          * buf to scd_buf_posted _before_ dropping scd_lock */
 482
 483         spin_lock(&scd->scd_lock);
 484
 485         if (rc == 0) {
 486                 if (!sv->sv_shuttingdown)
 487                         return 0;
 488
 489                 spin_unlock(&scd->scd_lock);
 490                 /* srpc_shutdown_service might have tried to unlink me
 491                  * when my buf_mdh was still invalid */
 492                 LNetMDUnlink(buf->buf_mdh);
 493                 spin_lock(&scd->scd_lock);
 494                 return 0;
 495         }
 496
 497         scd->scd_buf_nposted--;
 498         if (sv->sv_shuttingdown)
 499                 return rc; /* don't allow to change scd_buf_posted */
 500
 501         list_del(&buf->buf_list);
 502         spin_unlock(&scd->scd_lock);
 503
 504         LIBCFS_FREE(buf, sizeof(*buf));
 505
 506         spin_lock(&scd->scd_lock);
 507         return rc;
 508 }
 509
 510 int
 511 srpc_add_buffer(struct swi_workitem *wi)
 512 {
 513         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 514                                                    scd_buf_wi);
 515         struct srpc_buffer *buf;
 516         int rc = 0;
 517
 518         /* it's called by workitem scheduler threads, these threads
 519          * should have been set CPT affinity, so buffers will be posted
 520          * on CPT local list of Portal */
 521         spin_lock(&scd->scd_lock);
 522
 523         while (scd->scd_buf_adjust > 0 &&
 524                !scd->scd_svc->sv_shuttingdown) {
 525                 scd->scd_buf_adjust--; /* consume it */
 526                 scd->scd_buf_posting++;
 527
 528                 spin_unlock(&scd->scd_lock);
 529
 530                 LIBCFS_ALLOC(buf, sizeof(*buf));
 531                 if (buf == NULL) {
 532                         CERROR("Failed to add new buf to service: %s\n",
 533                                scd->scd_svc->sv_name);
 534                         spin_lock(&scd->scd_lock);
 535                         rc = -ENOMEM;
 536                         break;
 537                 }
 538
 539                 spin_lock(&scd->scd_lock);
 540                 if (scd->scd_svc->sv_shuttingdown) {
 541                         spin_unlock(&scd->scd_lock);
 542                         LIBCFS_FREE(buf, sizeof(*buf));
 543
 544                         spin_lock(&scd->scd_lock);
 545                         rc = -ESHUTDOWN;
 546                         break;
 547                 }
 548
 549                 rc = srpc_service_post_buffer(scd, buf);
 550                 if (rc != 0)
 551                         break; /* buf has been freed inside */
 552
 553                 LASSERT(scd->scd_buf_posting > 0);
 554                 scd->scd_buf_posting--;
 555                 scd->scd_buf_total++;
 556                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 557         }
 558
 559         if (rc != 0) {
 560                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 561                 scd->scd_buf_err = rc;
 562
 563                 LASSERT(scd->scd_buf_posting > 0);
 564                 scd->scd_buf_posting--;
 565         }
 566
 567         spin_unlock(&scd->scd_lock);
 568         return 0;
 569 }
 570
 571 int
 572 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 573 {
 574         struct srpc_service_cd *scd;
 575         int rc = 0;
 576         int i;
 577
 578         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 579
 580         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 581                 spin_lock(&scd->scd_lock);
 582
 583                 scd->scd_buf_err = 0;
 584                 scd->scd_buf_err_stamp = 0;
 585                 scd->scd_buf_posting = 0;
 586                 scd->scd_buf_adjust = nbuffer;
 587                 /* start to post buffers */
 588                 swi_schedule_workitem(&scd->scd_buf_wi);
 589                 spin_unlock(&scd->scd_lock);
 590
 591                 /* framework service only post buffer for one partition  */
 592                 if (srpc_serv_is_framework(sv))
 593                         break;
 594         }
 595
 596         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 597                 spin_lock(&scd->scd_lock);
 598                 /*
 599                  * NB: srpc_service_add_buffers() can be called inside
 600                  * thread context of lst_sched_serial, and we don't normally
 601                  * allow to sleep inside thread context of WI scheduler
 602                  * because it will block current scheduler thread from doing
 603                  * anything else, even worse, it could deadlock if it's
 604                  * waiting on result from another WI of the same scheduler.
 605                  * However, it's safe at here because scd_buf_wi is scheduled
 606                  * by thread in a different WI scheduler (lst_sched_test),
 607                  * so we don't have any risk of deadlock, though this could
 608                  * block all WIs pending on lst_sched_serial for a moment
 609                  * which is not good but not fatal.
 610                  */
 611                 lst_wait_until(scd->scd_buf_err != 0 ||
 612                                (scd->scd_buf_adjust == 0 &&
 613                                 scd->scd_buf_posting == 0),
 614                                scd->scd_lock, "waiting for adding buffer\n");
 615
 616                 if (scd->scd_buf_err != 0 && rc == 0)
 617                         rc = scd->scd_buf_err;
 618
 619                 spin_unlock(&scd->scd_lock);
 620         }
 621
 622         return rc;
 623 }
 624
 625 void
 626 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 627 {
 628         struct srpc_service_cd *scd;
 629         int num;
 630         int i;
 631
 632         LASSERT(!sv->sv_shuttingdown);
 633
 634         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 635                 spin_lock(&scd->scd_lock);
 636
 637                 num = scd->scd_buf_total + scd->scd_buf_posting;
 638                 scd->scd_buf_adjust -= min(nbuffer, num);
 639
 640                 spin_unlock(&scd->scd_lock);
 641         }
 642 }
 643
 644 /* returns 1 if sv has finished, otherwise 0 */
 645 int
 646 srpc_finish_service(struct srpc_service *sv)
 647 {
 648         struct srpc_service_cd *scd;
 649         struct srpc_server_rpc *rpc;
 650         int i;
 651
 652         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 653
 654         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 655                 spin_lock(&scd->scd_lock);
 656                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 657                         spin_unlock(&scd->scd_lock);
 658                         return 0;
 659                 }
 660
 661                 if (scd->scd_buf_nposted > 0) {
 662                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 663                                scd->scd_buf_nposted);
 664                         spin_unlock(&scd->scd_lock);
 665                         return 0;
 666                 }
 667
 668                 if (list_empty(&scd->scd_rpc_active)) {
 669                         spin_unlock(&scd->scd_lock);
 670                         continue;
 671                 }
 672
 673                 rpc = list_entry(scd->scd_rpc_active.next,
 674                                  struct srpc_server_rpc, srpc_list);
 675                 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",
 676                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 677                         swi_state2str(rpc->srpc_wi.swi_state),
 678                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 679                         rpc->srpc_wi.swi_workitem.wi_running,
 680                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 681                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 682                 spin_unlock(&scd->scd_lock);
 683                 return 0;
 684         }
 685
 686         /* no lock needed from now on */
 687         srpc_service_fini(sv);
 688         return 1;
 689 }
 690
 691 /* called with sv->sv_lock held */
 692 static void
 693 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 694                             struct srpc_buffer *buf)
 695 __must_hold(&scd->scd_lock)
 696 {
 697         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 698                 if (srpc_service_post_buffer(scd, buf) != 0) {
 699                         CWARN("Failed to post %s buffer\n",
 700                               scd->scd_svc->sv_name);
 701                 }
 702                 return;
 703         }
 704
 705         /* service is shutting down, or we want to recycle some buffers */
 706         scd->scd_buf_total--;
 707
 708         if (scd->scd_buf_adjust < 0) {
 709                 scd->scd_buf_adjust++;
 710                 if (scd->scd_buf_adjust < 0 &&
 711                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 712                         CDEBUG(D_INFO,
 713                                "Try to recyle %d buffers but nothing left\n",
 714                                scd->scd_buf_adjust);
 715                         scd->scd_buf_adjust = 0;
 716                 }
 717         }
 718
 719         spin_unlock(&scd->scd_lock);
 720         LIBCFS_FREE(buf, sizeof(*buf));
 721         spin_lock(&scd->scd_lock);
 722 }
 723
 724 void
 725 srpc_abort_service(struct srpc_service *sv)
 726 {
 727         struct srpc_service_cd *scd;
 728         struct srpc_server_rpc *rpc;
 729         int i;
 730
 731         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 732                sv->sv_id, sv->sv_name);
 733
 734         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 735                 spin_lock(&scd->scd_lock);
 736
 737                 /* schedule in-flight RPCs to notice the abort, NB:
 738                  * racing with incoming RPCs; complete fix should make test
 739                  * RPCs carry session ID in its headers
 740                  */
 741                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 742                         rpc->srpc_aborted = 1;
 743                         swi_schedule_workitem(&rpc->srpc_wi);
 744                 }
 745
 746                 spin_unlock(&scd->scd_lock);
 747         }
 748 }
 749
 750 void
 751 srpc_shutdown_service(struct srpc_service *sv)
 752 {
 753         struct srpc_service_cd *scd;
 754         struct srpc_server_rpc *rpc;
 755         struct srpc_buffer *buf;
 756         int i;
 757
 758         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 759                sv->sv_id, sv->sv_name);
 760
 761         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 762                 spin_lock(&scd->scd_lock);
 763
 764         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 765
 766         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 767                 spin_unlock(&scd->scd_lock);
 768
 769         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 770                 spin_lock(&scd->scd_lock);
 771
 772                 /* schedule in-flight RPCs to notice the shutdown */
 773                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 774                         swi_schedule_workitem(&rpc->srpc_wi);
 775
 776                 spin_unlock(&scd->scd_lock);
 777
 778                 /* OK to traverse scd_buf_posted without lock, since no one
 779                  * touches scd_buf_posted now
 780                  */
 781                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 782                         LNetMDUnlink(buf->buf_mdh);
 783         }
 784 }
 785
 786 static int
 787 srpc_send_request(struct srpc_client_rpc *rpc)
 788 {
 789         struct srpc_event *ev = &rpc->crpc_reqstev;
 790         int rc;
 791
 792         ev->ev_fired = 0;
 793         ev->ev_data  = rpc;
 794         ev->ev_type  = SRPC_REQUEST_SENT;
 795
 796         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 797                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 798                                    sizeof(struct srpc_msg), LNET_MD_OP_PUT,
 799                                    rpc->crpc_dest, LNET_NID_ANY,
 800                                    &rpc->crpc_reqstmdh, ev);
 801         if (rc != 0) {
 802                 LASSERT(rc == -ENOMEM);
 803                 ev->ev_fired = 1;  /* no more event expected */
 804         }
 805         return rc;
 806 }
 807
 808 static int
 809 srpc_prepare_reply(struct srpc_client_rpc *rpc)
 810 {
 811         struct srpc_event *ev = &rpc->crpc_replyev;
 812         u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 813         int rc;
 814
 815         ev->ev_fired = 0;
 816         ev->ev_data  = rpc;
 817         ev->ev_type  = SRPC_REPLY_RCVD;
 818
 819         *id = srpc_next_id();
 820
 821         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 822                                     &rpc->crpc_replymsg,
 823                                     sizeof(struct srpc_msg),
 824                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 825                                     &rpc->crpc_replymdh, ev);
 826         if (rc != 0) {
 827                 LASSERT(rc == -ENOMEM);
 828                 ev->ev_fired = 1;  /* no more event expected */
 829         }
 830         return rc;
 831 }
 832
 833 static int
 834 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 835 {
 836         struct srpc_bulk *bk = &rpc->crpc_bulk;
 837         struct srpc_event *ev = &rpc->crpc_bulkev;
 838         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 839         int rc;
 840         int opt;
 841
 842         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 843
 844         /* nothing to do */
 845         if (bk->bk_niov == 0)
 846                 return 0;
 847
 848         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 849         opt |= LNET_MD_KIOV;
 850
 851         ev->ev_fired = 0;
 852         ev->ev_data  = rpc;
 853         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 854
 855         *id = srpc_next_id();
 856
 857         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 858                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 859                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 860         if (rc != 0) {
 861                 LASSERT(rc == -ENOMEM);
 862                 ev->ev_fired = 1;  /* no more event expected */
 863         }
 864         return rc;
 865 }
 866
 867 static int
 868 srpc_do_bulk(struct srpc_server_rpc *rpc)
 869 {
 870         struct srpc_event *ev = &rpc->srpc_ev;
 871         struct srpc_bulk *bk = rpc->srpc_bulk;
 872         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 873         int rc;
 874         int opt;
 875
 876         LASSERT(bk != NULL);
 877
 878         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 879         opt |= LNET_MD_KIOV;
 880
 881         ev->ev_fired = 0;
 882         ev->ev_data  = rpc;
 883         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 884
 885         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 886                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 887                                    rpc->srpc_peer, rpc->srpc_self,
 888                                    &bk->bk_mdh, ev);
 889         if (rc != 0)
 890                 ev->ev_fired = 1;  /* no more event expected */
 891         return rc;
 892 }
 893
 894 /* only called from srpc_handle_rpc */
 895 static void
 896 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 897 {
 898         struct srpc_service_cd  *scd = rpc->srpc_scd;
 899         struct srpc_service     *sv  = scd->scd_svc;
 900         struct srpc_buffer *buffer;
 901
 902         LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 903
 904         rpc->srpc_status = status;
 905
 906         CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
 907                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 908                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 909                      swi_state2str(rpc->srpc_wi.swi_state), status);
 910
 911         if (status != 0) {
 912                 spin_lock(&srpc_data.rpc_glock);
 913                 srpc_data.rpc_counters.rpcs_dropped++;
 914                 spin_unlock(&srpc_data.rpc_glock);
 915         }
 916
 917         if (rpc->srpc_done != NULL)
 918                 (*rpc->srpc_done) (rpc);
 919         LASSERT(rpc->srpc_bulk == NULL);
 920
 921         spin_lock(&scd->scd_lock);
 922
 923         if (rpc->srpc_reqstbuf != NULL) {
 924                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 925                  * sv won't go away for scd_rpc_active must not be empty
 926                  */
 927                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 928                 rpc->srpc_reqstbuf = NULL;
 929         }
 930
 931         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 932
 933         /*
 934          * No one can schedule me now since:
 935          * - I'm not on scd_rpc_active.
 936          * - all LNet events have been fired.
 937          * Cancel pending schedules and prevent future schedule attempts:
 938          */
 939         LASSERT(rpc->srpc_ev.ev_fired);
 940         swi_exit_workitem(&rpc->srpc_wi);
 941
 942         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 943                 buffer = list_entry(scd->scd_buf_blocked.next,
 944                                     struct srpc_buffer, buf_list);
 945                 list_del(&buffer->buf_list);
 946
 947                 srpc_init_server_rpc(rpc, scd, buffer);
 948                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 949                 swi_schedule_workitem(&rpc->srpc_wi);
 950         } else {
 951                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 952         }
 953
 954         spin_unlock(&scd->scd_lock);
 955 }
 956
 957 /* handles an incoming RPC */
 958 static int srpc_handle_rpc(struct swi_workitem *wi)
 959 {
 960         struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc,
 961                                                    srpc_wi);
 962         struct srpc_service_cd *scd = rpc->srpc_scd;
 963         struct srpc_service *sv = scd->scd_svc;
 964         struct srpc_event *ev = &rpc->srpc_ev;
 965         int rc = 0;
 966
 967         LASSERT(wi == &rpc->srpc_wi);
 968
 969         spin_lock(&scd->scd_lock);
 970
 971         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 972                 spin_unlock(&scd->scd_lock);
 973
 974                 if (rpc->srpc_bulk != NULL)
 975                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 976                 LNetMDUnlink(rpc->srpc_replymdh);
 977
 978                 if (ev->ev_fired) { /* no more event, OK to finish */
 979                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 980                         return 1;
 981                 }
 982                 return 0;
 983         }
 984
 985         spin_unlock(&scd->scd_lock);
 986
 987         switch (wi->swi_state) {
 988         default:
 989                 LBUG();
 990                 /* fallthrough */
 991         case SWI_STATE_NEWBORN: {
 992                 struct srpc_msg *msg;
 993                 struct srpc_generic_reply *reply;
 994
 995                 msg = &rpc->srpc_reqstbuf->buf_msg;
 996                 reply = &rpc->srpc_replymsg.msg_body.reply;
 997
 998                 if (msg->msg_magic == 0) {
 999                         /* moaned already in srpc_lnet_ev_handler */
1000                         srpc_server_rpc_done(rpc, EBADMSG);
1001                         return 1;
1002                 }
1003
1004                 srpc_unpack_msg_hdr(msg);
1005                 if (msg->msg_version != SRPC_MSG_VERSION) {
1006                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1007                               msg->msg_version, SRPC_MSG_VERSION,
1008                               libcfs_id2str(rpc->srpc_peer));
1009                         reply->status = EPROTO;
1010                         /* drop through and send reply */
1011                 } else {
1012                         reply->status = 0;
1013                         rc = (*sv->sv_handler)(rpc);
1014                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1015                         if (rc != 0) {
1016                                 srpc_server_rpc_done(rpc, rc);
1017                                 return 1;
1018                         }
1019                 }
1020
1021                 wi->swi_state = SWI_STATE_BULK_STARTED;
1022
1023                 if (rpc->srpc_bulk != NULL) {
1024                         rc = srpc_do_bulk(rpc);
1025                         if (rc == 0)
1026                                 return 0; /* wait for bulk */
1027
1028                         LASSERT(ev->ev_fired);
1029                         ev->ev_status = rc;
1030                 }
1031         }
1032         /* fallthrough */
1033         case SWI_STATE_BULK_STARTED:
1034                 LASSERT(rpc->srpc_bulk == NULL || ev->ev_fired);
1035
1036                 if (rpc->srpc_bulk != NULL) {
1037                         rc = ev->ev_status;
1038
1039                         if (sv->sv_bulk_ready != NULL)
1040                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1041
1042                         if (rc != 0) {
1043                                 srpc_server_rpc_done(rpc, rc);
1044                                 return 1;
1045                         }
1046                 }
1047
1048                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1049                 rc = srpc_send_reply(rpc);
1050                 if (rc == 0)
1051                         return 0; /* wait for reply */
1052                 srpc_server_rpc_done(rpc, rc);
1053                 return 1;
1054
1055         case SWI_STATE_REPLY_SUBMITTED:
1056                 if (!ev->ev_fired) {
1057                         CERROR("RPC %p: bulk %p, service %d\n",
1058                                rpc, rpc->srpc_bulk, sv->sv_id);
1059                         CERROR("Event: status %d, type %d, lnet %d\n",
1060                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1061                         LASSERT(ev->ev_fired);
1062                 }
1063
1064                 wi->swi_state = SWI_STATE_DONE;
1065                 srpc_server_rpc_done(rpc, ev->ev_status);
1066                 return 1;
1067         }
1068
1069         return 0;
1070 }
1071
1072 static void
1073 srpc_client_rpc_expired (void *data)
1074 {
1075         struct srpc_client_rpc *rpc = data;
1076
1077         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1078               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1079               rpc->crpc_timeout);
1080
1081         spin_lock(&rpc->crpc_lock);
1082
1083         rpc->crpc_timeout = 0;
1084         srpc_abort_rpc(rpc, -ETIMEDOUT);
1085
1086         spin_unlock(&rpc->crpc_lock);
1087
1088         spin_lock(&srpc_data.rpc_glock);
1089         srpc_data.rpc_counters.rpcs_expired++;
1090         spin_unlock(&srpc_data.rpc_glock);
1091 }
1092
1093 static void
1094 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1095 {
1096         struct stt_timer *timer = &rpc->crpc_timer;
1097
1098         if (rpc->crpc_timeout == 0)
1099                 return;
1100
1101         INIT_LIST_HEAD(&timer->stt_list);
1102         timer->stt_data    = rpc;
1103         timer->stt_func    = srpc_client_rpc_expired;
1104         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1105         stt_add_timer(timer);
1106 }
1107
1108 /*
1109  * Called with rpc->crpc_lock held.
1110  *
1111  * Upon exit the RPC expiry timer is not queued and the handler is not
1112  * running on any CPU.
1113  */
1114 static void
1115 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1116 {
1117         /* timer not planted or already exploded */
1118         if (rpc->crpc_timeout == 0)
1119                 return;
1120
1121         /* timer successfully defused */
1122         if (stt_del_timer(&rpc->crpc_timer))
1123                 return;
1124
1125         /* timer detonated, wait for it to explode */
1126         while (rpc->crpc_timeout != 0) {
1127                 spin_unlock(&rpc->crpc_lock);
1128
1129                 schedule();
1130
1131                 spin_lock(&rpc->crpc_lock);
1132         }
1133 }
1134
1135 static void
1136 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1137 {
1138         struct swi_workitem *wi = &rpc->crpc_wi;
1139
1140         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1141
1142         spin_lock(&rpc->crpc_lock);
1143
1144         rpc->crpc_closed = 1;
1145         if (rpc->crpc_status == 0)
1146                 rpc->crpc_status = status;
1147
1148         srpc_del_client_rpc_timer(rpc);
1149
1150         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1151                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1152                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1153                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1154
1155         /*
1156          * No one can schedule me now since:
1157          * - RPC timer has been defused.
1158          * - all LNet events have been fired.
1159          * - crpc_closed has been set, preventing srpc_abort_rpc from
1160          *   scheduling me.
1161          * Cancel pending schedules and prevent future schedule attempts:
1162          */
1163         LASSERT(!srpc_event_pending(rpc));
1164         swi_exit_workitem(wi);
1165
1166         spin_unlock(&rpc->crpc_lock);
1167
1168         (*rpc->crpc_done)(rpc);
1169 }
1170
1171 /* sends an outgoing RPC */
1172 int
1173 srpc_send_rpc(struct swi_workitem *wi)
1174 {
1175         int rc = 0;
1176         struct srpc_client_rpc *rpc;
1177         struct srpc_msg *reply;
1178         int do_bulk;
1179
1180         LASSERT(wi != NULL);
1181
1182         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1183
1184         LASSERT(rpc != NULL);
1185         LASSERT(wi == &rpc->crpc_wi);
1186
1187         reply = &rpc->crpc_replymsg;
1188         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1189
1190         spin_lock(&rpc->crpc_lock);
1191
1192         if (rpc->crpc_aborted) {
1193                 spin_unlock(&rpc->crpc_lock);
1194                 goto abort;
1195         }
1196
1197         spin_unlock(&rpc->crpc_lock);
1198
1199         switch (wi->swi_state) {
1200         default:
1201                 LBUG();
1202         case SWI_STATE_NEWBORN:
1203                 LASSERT(!srpc_event_pending(rpc));
1204
1205                 rc = srpc_prepare_reply(rpc);
1206                 if (rc != 0) {
1207                         srpc_client_rpc_done(rpc, rc);
1208                         return 1;
1209                 }
1210
1211                 rc = srpc_prepare_bulk(rpc);
1212                 if (rc != 0)
1213                         break;
1214
1215                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1216                 rc = srpc_send_request(rpc);
1217                 break;
1218
1219         case SWI_STATE_REQUEST_SUBMITTED:
1220                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1221                  * order; however, they're processed in a strict order:
1222                  * rqt, rpy, and bulk.
1223                  */
1224                 if (!rpc->crpc_reqstev.ev_fired)
1225                         break;
1226
1227                 rc = rpc->crpc_reqstev.ev_status;
1228                 if (rc != 0)
1229                         break;
1230
1231                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1232                 /* fallthrough */
1233         case SWI_STATE_REQUEST_SENT: {
1234                 enum srpc_msg_type type;
1235
1236                 type = srpc_service2reply(rpc->crpc_service);
1237
1238                 if (!rpc->crpc_replyev.ev_fired)
1239                         break;
1240
1241                 rc = rpc->crpc_replyev.ev_status;
1242                 if (rc != 0)
1243                         break;
1244
1245                 srpc_unpack_msg_hdr(reply);
1246                 if (reply->msg_type != type ||
1247                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1248                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1249                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1250                               libcfs_id2str(rpc->crpc_dest),
1251                               reply->msg_type, type,
1252                               reply->msg_magic, SRPC_MSG_MAGIC);
1253                         rc = -EBADMSG;
1254                         break;
1255                 }
1256
1257                 if (do_bulk && reply->msg_body.reply.status != 0) {
1258                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1259                               reply->msg_body.reply.status,
1260                               libcfs_id2str(rpc->crpc_dest));
1261                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1262                 }
1263
1264                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1265         }
1266         /* fallthrough */
1267         case SWI_STATE_REPLY_RECEIVED:
1268                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1269                         break;
1270
1271                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1272
1273                 /* Bulk buffer was unlinked due to remote error. Clear error
1274                  * since reply buffer still contains valid data.
1275                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1276                  * remote error.
1277                  */
1278                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1279                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1280                         rc = 0;
1281
1282                 wi->swi_state = SWI_STATE_DONE;
1283                 srpc_client_rpc_done(rpc, rc);
1284                 return 1;
1285         }
1286
1287         if (rc != 0) {
1288                 spin_lock(&rpc->crpc_lock);
1289                 srpc_abort_rpc(rpc, rc);
1290                 spin_unlock(&rpc->crpc_lock);
1291         }
1292
1293 abort:
1294         if (rpc->crpc_aborted) {
1295                 LNetMDUnlink(rpc->crpc_reqstmdh);
1296                 LNetMDUnlink(rpc->crpc_replymdh);
1297                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1298
1299                 if (!srpc_event_pending(rpc)) {
1300                         srpc_client_rpc_done(rpc, -EINTR);
1301                         return 1;
1302                 }
1303         }
1304         return 0;
1305 }
1306
1307 struct srpc_client_rpc *
1308 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1309                        int nbulkiov, int bulklen,
1310                        void (*rpc_done)(struct srpc_client_rpc *),
1311                        void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1312 {
1313         struct srpc_client_rpc *rpc;
1314
1315         LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1316                                    crpc_bulk.bk_iovs[nbulkiov]));
1317         if (rpc == NULL)
1318                 return NULL;
1319
1320         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1321                              bulklen, rpc_done, rpc_fini, priv);
1322         return rpc;
1323 }
1324
1325 /* called with rpc->crpc_lock held */
1326 void
1327 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1328 {
1329         LASSERT(why != 0);
1330
1331         if (rpc->crpc_aborted || /* already aborted */
1332             rpc->crpc_closed)    /* callback imminent */
1333                 return;
1334
1335         CDEBUG(D_NET,
1336                "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1337                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1338                swi_state2str(rpc->crpc_wi.swi_state), why);
1339
1340         rpc->crpc_aborted = 1;
1341         rpc->crpc_status  = why;
1342         swi_schedule_workitem(&rpc->crpc_wi);
1343 }
1344
1345 /* called with rpc->crpc_lock held */
1346 void
1347 srpc_post_rpc(struct srpc_client_rpc *rpc)
1348 {
1349         LASSERT(!rpc->crpc_aborted);
1350         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1351
1352         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1353                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1354                rpc->crpc_timeout);
1355
1356         srpc_add_client_rpc_timer(rpc);
1357         swi_schedule_workitem(&rpc->crpc_wi);
1358 }
1359
1360
1361 int
1362 srpc_send_reply(struct srpc_server_rpc *rpc)
1363 {
1364         struct srpc_event *ev = &rpc->srpc_ev;
1365         struct srpc_msg *msg = &rpc->srpc_replymsg;
1366         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1367         struct srpc_service_cd *scd = rpc->srpc_scd;
1368         struct srpc_service *sv = scd->scd_svc;
1369         __u64 rpyid;
1370         int rc;
1371
1372         LASSERT(buffer != NULL);
1373         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1374
1375         spin_lock(&scd->scd_lock);
1376
1377         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1378                 /* Repost buffer before replying since test client
1379                  * might send me another RPC once it gets the reply
1380                  */
1381                 if (srpc_service_post_buffer(scd, buffer) != 0)
1382                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1383                 rpc->srpc_reqstbuf = NULL;
1384         }
1385
1386         spin_unlock(&scd->scd_lock);
1387
1388         ev->ev_fired = 0;
1389         ev->ev_data  = rpc;
1390         ev->ev_type  = SRPC_REPLY_SENT;
1391
1392         msg->msg_magic   = SRPC_MSG_MAGIC;
1393         msg->msg_version = SRPC_MSG_VERSION;
1394         msg->msg_type    = srpc_service2reply(sv->sv_id);
1395
1396         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1397                                    sizeof(*msg), LNET_MD_OP_PUT,
1398                                    rpc->srpc_peer, rpc->srpc_self,
1399                                    &rpc->srpc_replymdh, ev);
1400         if (rc != 0)
1401                 ev->ev_fired = 1;  /* no more event expected */
1402         return rc;
1403 }
1404
1405 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1406 static void
1407 srpc_lnet_ev_handler(struct lnet_event *ev)
1408 {
1409         struct srpc_service_cd *scd;
1410         struct srpc_event *rpcev = ev->md.user_ptr;
1411         struct srpc_client_rpc *crpc;
1412         struct srpc_server_rpc *srpc;
1413         struct srpc_buffer *buffer;
1414         struct srpc_service *sv;
1415         struct srpc_msg *msg;
1416         enum srpc_msg_type type;
1417
1418         LASSERT(!in_interrupt());
1419
1420         if (ev->status != 0) {
1421                 __u32 errors;
1422
1423                 spin_lock(&srpc_data.rpc_glock);
1424                 if (ev->status != -ECANCELED) /* cancellation is not error */
1425                         srpc_data.rpc_counters.errors++;
1426                 errors = srpc_data.rpc_counters.errors;
1427                 spin_unlock(&srpc_data.rpc_glock);
1428
1429                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1430                         ev->status, ev->type, errors);
1431         }
1432
1433         rpcev->ev_lnet = ev->type;
1434
1435         switch (rpcev->ev_type) {
1436         default:
1437                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1438                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1439                 LBUG();
1440                 /* fallthrough */
1441         case SRPC_REQUEST_SENT:
1442                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1443                         spin_lock(&srpc_data.rpc_glock);
1444                         srpc_data.rpc_counters.rpcs_sent++;
1445                         spin_unlock(&srpc_data.rpc_glock);
1446                 }
1447                 /* fallthrough */
1448         case SRPC_REPLY_RCVD:
1449         case SRPC_BULK_REQ_RCVD:
1450                 crpc = rpcev->ev_data;
1451
1452                 if (rpcev != &crpc->crpc_reqstev &&
1453                     rpcev != &crpc->crpc_replyev &&
1454                     rpcev != &crpc->crpc_bulkev) {
1455                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1456                                rpcev, crpc, &crpc->crpc_reqstev,
1457                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1458                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1459                                rpcev->ev_status, rpcev->ev_type,
1460                                rpcev->ev_lnet);
1461                         LBUG();
1462                 }
1463
1464                 spin_lock(&crpc->crpc_lock);
1465
1466                 LASSERT(rpcev->ev_fired == 0);
1467                 rpcev->ev_fired  = 1;
1468                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1469                                    -EINTR : ev->status;
1470                 swi_schedule_workitem(&crpc->crpc_wi);
1471
1472                 spin_unlock(&crpc->crpc_lock);
1473                 break;
1474
1475         case SRPC_REQUEST_RCVD:
1476                 scd = rpcev->ev_data;
1477                 sv = scd->scd_svc;
1478
1479                 LASSERT(rpcev == &scd->scd_ev);
1480
1481                 spin_lock(&scd->scd_lock);
1482
1483                 LASSERT(ev->unlinked);
1484                 LASSERT(ev->type == LNET_EVENT_PUT ||
1485                         ev->type == LNET_EVENT_UNLINK);
1486                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1487                         sv->sv_shuttingdown);
1488
1489                 buffer = container_of(ev->md.start, struct srpc_buffer,
1490                                       buf_msg);
1491                 buffer->buf_peer = ev->source;
1492                 buffer->buf_self = ev->target.nid;
1493
1494                 LASSERT(scd->scd_buf_nposted > 0);
1495                 scd->scd_buf_nposted--;
1496
1497                 if (sv->sv_shuttingdown) {
1498                         /* Leave buffer on scd->scd_buf_nposted since
1499                          * srpc_finish_service needs to traverse it.
1500                          */
1501                         spin_unlock(&scd->scd_lock);
1502                         break;
1503                 }
1504
1505                 if (scd->scd_buf_err_stamp != 0 &&
1506                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1507                         /* re-enable adding buffer */
1508                         scd->scd_buf_err_stamp = 0;
1509                         scd->scd_buf_err = 0;
1510                 }
1511
1512                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1513                     scd->scd_buf_adjust == 0 &&
1514                     scd->scd_buf_nposted < scd->scd_buf_low) {
1515                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1516                                                   SFW_TEST_WI_MIN);
1517                         swi_schedule_workitem(&scd->scd_buf_wi);
1518                 }
1519
1520                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1521                 msg = &buffer->buf_msg;
1522                 type = srpc_service2request(sv->sv_id);
1523
1524                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1525                     (msg->msg_type != type &&
1526                      msg->msg_type != __swab32(type)) ||
1527                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1528                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1529                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1530                                sv->sv_name, libcfs_id2str(ev->initiator),
1531                                ev->status, ev->mlength,
1532                                msg->msg_type, msg->msg_magic);
1533
1534                         /* NB can't call srpc_service_recycle_buffer here since
1535                          * it may call LNetM[DE]Attach. The invalid magic tells
1536                          * srpc_handle_rpc to drop this RPC
1537                          */
1538                         msg->msg_magic = 0;
1539                 }
1540
1541                 if (!list_empty(&scd->scd_rpc_free)) {
1542                         srpc = list_entry(scd->scd_rpc_free.next,
1543                                           struct srpc_server_rpc,
1544                                           srpc_list);
1545                         list_del(&srpc->srpc_list);
1546
1547                         srpc_init_server_rpc(srpc, scd, buffer);
1548                         list_add_tail(&srpc->srpc_list,
1549                                       &scd->scd_rpc_active);
1550                         swi_schedule_workitem(&srpc->srpc_wi);
1551                 } else {
1552                         list_add_tail(&buffer->buf_list,
1553                                       &scd->scd_buf_blocked);
1554                 }
1555
1556                 spin_unlock(&scd->scd_lock);
1557
1558                 spin_lock(&srpc_data.rpc_glock);
1559                 srpc_data.rpc_counters.rpcs_rcvd++;
1560                 spin_unlock(&srpc_data.rpc_glock);
1561                 break;
1562
1563         case SRPC_BULK_GET_RPLD:
1564                 LASSERT(ev->type == LNET_EVENT_SEND ||
1565                         ev->type == LNET_EVENT_REPLY ||
1566                         ev->type == LNET_EVENT_UNLINK);
1567
1568                 if (!ev->unlinked)
1569                         break; /* wait for final event */
1570                 /* fallthrough */
1571         case SRPC_BULK_PUT_SENT:
1572                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1573                         spin_lock(&srpc_data.rpc_glock);
1574
1575                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1576                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1577                         else
1578                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1579
1580                         spin_unlock(&srpc_data.rpc_glock);
1581                 }
1582                 /* fallthrough */
1583         case SRPC_REPLY_SENT:
1584                 srpc = rpcev->ev_data;
1585                 scd  = srpc->srpc_scd;
1586
1587                 LASSERT(rpcev == &srpc->srpc_ev);
1588
1589                 spin_lock(&scd->scd_lock);
1590
1591                 rpcev->ev_fired  = 1;
1592                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1593                                    -EINTR : ev->status;
1594                 swi_schedule_workitem(&srpc->srpc_wi);
1595
1596                 spin_unlock(&scd->scd_lock);
1597                 break;
1598         }
1599 }
1600
1601
1602 int
1603 srpc_startup (void)
1604 {
1605         int rc;
1606
1607         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1608         spin_lock_init(&srpc_data.rpc_glock);
1609
1610         /* 1 second pause to avoid timestamp reuse */
1611         set_current_state(TASK_UNINTERRUPTIBLE);
1612         schedule_timeout(cfs_time_seconds(1));
1613         srpc_data.rpc_matchbits = ((__u64) ktime_get_real_seconds()) << 48;
1614
1615         srpc_data.rpc_state = SRPC_STATE_NONE;
1616
1617         rc = LNetNIInit(LNET_PID_LUSTRE);
1618         if (rc < 0) {
1619                 CERROR("LNetNIInit() has failed: %d\n", rc);
1620                 return rc;
1621         }
1622
1623         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1624
1625         LNetInvalidateEQHandle(&srpc_data.rpc_lnet_eq);
1626         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1627         if (rc != 0) {
1628                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1629                 goto bail;
1630         }
1631
1632         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1633         LASSERT(rc == 0);
1634         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1635         LASSERT(rc == 0);
1636
1637         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1638
1639         rc = stt_startup();
1640
1641 bail:
1642         if (rc != 0)
1643                 srpc_shutdown();
1644         else
1645                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1646
1647         return rc;
1648 }
1649
1650 void
1651 srpc_shutdown (void)
1652 {
1653         int i;
1654         int rc;
1655         int state;
1656
1657         state = srpc_data.rpc_state;
1658         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1659
1660         switch (state) {
1661         default:
1662                 LBUG();
1663                 /* fallthrough */
1664         case SRPC_STATE_RUNNING:
1665                 spin_lock(&srpc_data.rpc_glock);
1666
1667                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1668                         struct srpc_service *sv = srpc_data.rpc_services[i];
1669
1670                         LASSERTF(sv == NULL,
1671                                  "service not empty: id %d, name %s\n",
1672                                  i, sv->sv_name);
1673                 }
1674
1675                 spin_unlock(&srpc_data.rpc_glock);
1676
1677                 stt_shutdown();
1678                 /* fallthrough */
1679
1680         case SRPC_STATE_EQ_INIT:
1681                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1682                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1683                 LASSERT(rc == 0);
1684                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1685                 LASSERT(rc == 0); /* the EQ should have no user by now */
1686                 /* fallthrough */
1687
1688         case SRPC_STATE_NI_INIT:
1689                 LNetNIFini();
1690         }
1691 }