lnet/selftest/rpc.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/selftest/rpc.c
  37  *
  38  * Author: Isaac Huang <isaac@clusterfs.com>
  39  *
  40  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  41  * - percpt data for service to improve smp performance
  42  * - code cleanup
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_LNET
  46
  47 #include "selftest.h"
  48
  49 typedef enum {
  50         SRPC_STATE_NONE,
  51         SRPC_STATE_NI_INIT,
  52         SRPC_STATE_EQ_INIT,
  53         SRPC_STATE_RUNNING,
  54         SRPC_STATE_STOPPING,
  55 } srpc_state_t;
  56
  57 static struct smoketest_rpc {
  58         spinlock_t       rpc_glock;     /* global lock */
  59         srpc_service_t  *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  60         lnet_handle_eq_t rpc_lnet_eq;   /* _the_ LNet event queue */
  61         srpc_state_t     rpc_state;
  62         srpc_counters_t  rpc_counters;
  63         __u64            rpc_matchbits; /* matchbits counter */
  64 } srpc_data;
  65
  66 static inline int
  67 srpc_serv_portal(int svc_id)
  68 {
  69         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  70                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  71 }
  72
  73 /* forward ref's */
  74 int srpc_handle_rpc (swi_workitem_t *wi);
  75
  76 void srpc_get_counters (srpc_counters_t *cnt)
  77 {
  78         spin_lock(&srpc_data.rpc_glock);
  79         *cnt = srpc_data.rpc_counters;
  80         spin_unlock(&srpc_data.rpc_glock);
  81 }
  82
  83 void srpc_set_counters (const srpc_counters_t *cnt)
  84 {
  85         spin_lock(&srpc_data.rpc_glock);
  86         srpc_data.rpc_counters = *cnt;
  87         spin_unlock(&srpc_data.rpc_glock);
  88 }
  89
  90 static int
  91 srpc_add_bulk_page(srpc_bulk_t *bk, struct page *pg, int i, int nob)
  92 {
  93         nob = min(nob, (int)PAGE_CACHE_SIZE);
  94
  95         LASSERT(nob > 0);
  96         LASSERT(i >= 0 && i < bk->bk_niov);
  97
  98         bk->bk_iovs[i].kiov_offset = 0;
  99         bk->bk_iovs[i].kiov_page   = pg;
 100         bk->bk_iovs[i].kiov_len    = nob;
 101         return nob;
 102 }
 103
 104 void
 105 srpc_free_bulk (srpc_bulk_t *bk)
 106 {
 107         int         i;
 108         struct page *pg;
 109
 110         LASSERT (bk != NULL);
 111
 112         for (i = 0; i < bk->bk_niov; i++) {
 113                 pg = bk->bk_iovs[i].kiov_page;
 114                 if (pg == NULL) break;
 115
 116                 __free_page(pg);
 117         }
 118
 119         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 120         return;
 121 }
 122
 123 srpc_bulk_t *
 124 srpc_alloc_bulk(int cpt, unsigned bulk_npg, unsigned bulk_len, int sink)
 125 {
 126         srpc_bulk_t  *bk;
 127         int           i;
 128
 129         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 130
 131         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 132                          offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 133         if (bk == NULL) {
 134                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 135                 return NULL;
 136         }
 137
 138         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 139         bk->bk_sink   = sink;
 140         bk->bk_len    = bulk_len;
 141         bk->bk_niov   = bulk_npg;
 142
 143         for (i = 0; i < bulk_npg; i++) {
 144                 struct page *pg;
 145                 int         nob;
 146
 147                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 148                 if (pg == NULL) {
 149                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 150                         srpc_free_bulk(bk);
 151                         return NULL;
 152                 }
 153
 154                 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
 155                 bulk_len -= nob;
 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, rpc, 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         LNetInvalidateHandle(&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, scd,
 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 (srpc_service_t *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, lnet_process_id_t peer,
 356                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 357 {
 358         int              rc;
 359         lnet_md_t        md;
 360         lnet_handle_me_t meh;
 361
 362         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 363                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 364         if (rc != 0) {
 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.eq_handle = srpc_data.rpc_lnet_eq;
 376
 377         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 378         if (rc != 0) {
 379                 CERROR ("LNetMDAttach failed: %d\n", rc);
 380                 LASSERT (rc == -ENOMEM);
 381
 382                 rc = LNetMEUnlink(meh);
 383                 LASSERT (rc == 0);
 384                 return -ENOMEM;
 385         }
 386
 387         CDEBUG (D_NET,
 388                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\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, lnet_process_id_t peer, lnet_nid_t self,
 396                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 397 {
 398         int       rc;
 399         lnet_md_t md;
 400
 401         md.user_ptr  = ev;
 402         md.start     = buf;
 403         md.length    = len;
 404         md.eq_handle = srpc_data.rpc_lnet_eq;
 405         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 406         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 407
 408         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 409         if (rc != 0) {
 410                 CERROR ("LNetMDBind failed: %d\n", rc);
 411                 LASSERT (rc == -ENOMEM);
 412                 return -ENOMEM;
 413         }
 414
 415         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 416          * they're only meaningful for MDs attached to an ME (i.e. passive
 417          * buffers... */
 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);
 425         }
 426
 427         if (rc != 0) {
 428                 CERROR ("LNet%s(%s, %d, "LPD64") 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 "LPX64"\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                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 448 {
 449         lnet_process_id_t 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         LNetInvalidateHandle(&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 = wi->swi_workitem.wi_data;
 512         struct srpc_buffer      *buf;
 513         int                     rc = 0;
 514
 515         /* it's called by workitem scheduler threads, these threads
 516          * should have been set CPT affinity, so buffers will be posted
 517          * on CPT local list of Portal */
 518         spin_lock(&scd->scd_lock);
 519
 520         while (scd->scd_buf_adjust > 0 &&
 521                !scd->scd_svc->sv_shuttingdown) {
 522                 scd->scd_buf_adjust--; /* consume it */
 523                 scd->scd_buf_posting++;
 524
 525                 spin_unlock(&scd->scd_lock);
 526
 527                 LIBCFS_ALLOC(buf, sizeof(*buf));
 528                 if (buf == NULL) {
 529                         CERROR("Failed to add new buf to service: %s\n",
 530                                scd->scd_svc->sv_name);
 531                         spin_lock(&scd->scd_lock);
 532                         rc = -ENOMEM;
 533                         break;
 534                 }
 535
 536                 spin_lock(&scd->scd_lock);
 537                 if (scd->scd_svc->sv_shuttingdown) {
 538                         spin_unlock(&scd->scd_lock);
 539                         LIBCFS_FREE(buf, sizeof(*buf));
 540
 541                         spin_lock(&scd->scd_lock);
 542                         rc = -ESHUTDOWN;
 543                         break;
 544                 }
 545
 546                 rc = srpc_service_post_buffer(scd, buf);
 547                 if (rc != 0)
 548                         break; /* buf has been freed inside */
 549
 550                 LASSERT(scd->scd_buf_posting > 0);
 551                 scd->scd_buf_posting--;
 552                 scd->scd_buf_total++;
 553                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 554         }
 555
 556         if (rc != 0) {
 557                 scd->scd_buf_err_stamp = cfs_time_current_sec();
 558                 scd->scd_buf_err = rc;
 559
 560                 LASSERT(scd->scd_buf_posting > 0);
 561                 scd->scd_buf_posting--;
 562         }
 563
 564         spin_unlock(&scd->scd_lock);
 565         return 0;
 566 }
 567
 568 int
 569 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 570 {
 571         struct srpc_service_cd  *scd;
 572         int                     rc = 0;
 573         int                     i;
 574
 575         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 576
 577         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 578                 spin_lock(&scd->scd_lock);
 579
 580                 scd->scd_buf_err = 0;
 581                 scd->scd_buf_err_stamp = 0;
 582                 scd->scd_buf_posting = 0;
 583                 scd->scd_buf_adjust = nbuffer;
 584                 /* start to post buffers */
 585                 swi_schedule_workitem(&scd->scd_buf_wi);
 586                 spin_unlock(&scd->scd_lock);
 587
 588                 /* framework service only post buffer for one partition  */
 589                 if (srpc_serv_is_framework(sv))
 590                         break;
 591         }
 592
 593         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 594                 spin_lock(&scd->scd_lock);
 595                 /*
 596                  * NB: srpc_service_add_buffers() can be called inside
 597                  * thread context of lst_sched_serial, and we don't normally
 598                  * allow to sleep inside thread context of WI scheduler
 599                  * because it will block current scheduler thread from doing
 600                  * anything else, even worse, it could deadlock if it's
 601                  * waiting on result from another WI of the same scheduler.
 602                  * However, it's safe at here because scd_buf_wi is scheduled
 603                  * by thread in a different WI scheduler (lst_sched_test),
 604                  * so we don't have any risk of deadlock, though this could
 605                  * block all WIs pending on lst_sched_serial for a moment
 606                  * which is not good but not fatal.
 607                  */
 608                 lst_wait_until(scd->scd_buf_err != 0 ||
 609                                (scd->scd_buf_adjust == 0 &&
 610                                 scd->scd_buf_posting == 0),
 611                                scd->scd_lock, "waiting for adding buffer\n");
 612
 613                 if (scd->scd_buf_err != 0 && rc == 0)
 614                         rc = scd->scd_buf_err;
 615
 616                 spin_unlock(&scd->scd_lock);
 617         }
 618
 619         return rc;
 620 }
 621
 622 void
 623 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 624 {
 625         struct srpc_service_cd  *scd;
 626         int                     num;
 627         int                     i;
 628
 629         LASSERT(!sv->sv_shuttingdown);
 630
 631         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 632                 spin_lock(&scd->scd_lock);
 633
 634                 num = scd->scd_buf_total + scd->scd_buf_posting;
 635                 scd->scd_buf_adjust -= min(nbuffer, num);
 636
 637                 spin_unlock(&scd->scd_lock);
 638         }
 639 }
 640
 641 /* returns 1 if sv has finished, otherwise 0 */
 642 int
 643 srpc_finish_service(struct srpc_service *sv)
 644 {
 645         struct srpc_service_cd  *scd;
 646         struct srpc_server_rpc  *rpc;
 647         int                     i;
 648
 649         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 650
 651         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 652                 spin_lock(&scd->scd_lock);
 653                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 654                         spin_unlock(&scd->scd_lock);
 655                         return 0;
 656                 }
 657
 658                 if (scd->scd_buf_nposted > 0) {
 659                         CDEBUG(D_NET, "waiting for %d posted buffers to "
 660                                "unlink\n", scd->scd_buf_nposted);
 661                         spin_unlock(&scd->scd_lock);
 662                         return 0;
 663                 }
 664
 665                 if (list_empty(&scd->scd_rpc_active)) {
 666                         spin_unlock(&scd->scd_lock);
 667                         continue;
 668                 }
 669
 670                 rpc = list_entry(scd->scd_rpc_active.next,
 671                                      struct srpc_server_rpc, srpc_list);
 672                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
 673                         "wi %s scheduled %d running %d, "
 674                         "ev fired %d type %d status %d lnet %d\n",
 675                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 676                         swi_state2str(rpc->srpc_wi.swi_state),
 677                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 678                         rpc->srpc_wi.swi_workitem.wi_running,
 679                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 680                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 681                 spin_unlock(&scd->scd_lock);
 682                 return 0;
 683         }
 684
 685         /* no lock needed from now on */
 686         srpc_service_fini(sv);
 687         return 1;
 688 }
 689
 690 /* called with sv->sv_lock held */
 691 static void
 692 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *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                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 739                         rpc->srpc_aborted = 1;
 740                         swi_schedule_workitem(&rpc->srpc_wi);
 741                 }
 742
 743                 spin_unlock(&scd->scd_lock);
 744         }
 745 }
 746
 747 void
 748 srpc_shutdown_service(srpc_service_t *sv)
 749 {
 750         struct srpc_service_cd  *scd;
 751         struct srpc_server_rpc  *rpc;
 752         srpc_buffer_t           *buf;
 753         int                     i;
 754
 755         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 756                sv->sv_id, sv->sv_name);
 757
 758         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 759                 spin_lock(&scd->scd_lock);
 760
 761         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 762
 763         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 764                 spin_unlock(&scd->scd_lock);
 765
 766         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 767                 spin_lock(&scd->scd_lock);
 768
 769                 /* schedule in-flight RPCs to notice the shutdown */
 770                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 771                         swi_schedule_workitem(&rpc->srpc_wi);
 772
 773                 spin_unlock(&scd->scd_lock);
 774
 775                 /* OK to traverse scd_buf_posted without lock, since no one
 776                  * touches scd_buf_posted now */
 777                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 778                         LNetMDUnlink(buf->buf_mdh);
 779         }
 780 }
 781
 782 static int
 783 srpc_send_request (srpc_client_rpc_t *rpc)
 784 {
 785         srpc_event_t *ev = &rpc->crpc_reqstev;
 786         int           rc;
 787
 788         ev->ev_fired = 0;
 789         ev->ev_data  = rpc;
 790         ev->ev_type  = SRPC_REQUEST_SENT;
 791
 792         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 793                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 794                                    sizeof(srpc_msg_t), LNET_MD_OP_PUT,
 795                                    rpc->crpc_dest, LNET_NID_ANY,
 796                                    &rpc->crpc_reqstmdh, ev);
 797         if (rc != 0) {
 798                 LASSERT (rc == -ENOMEM);
 799                 ev->ev_fired = 1;  /* no more event expected */
 800         }
 801         return rc;
 802 }
 803
 804 static int
 805 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 806 {
 807         srpc_event_t *ev = &rpc->crpc_replyev;
 808         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 809         int           rc;
 810
 811         ev->ev_fired = 0;
 812         ev->ev_data  = rpc;
 813         ev->ev_type  = SRPC_REPLY_RCVD;
 814
 815         *id = srpc_next_id();
 816
 817         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 818                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 819                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 820                                     &rpc->crpc_replymdh, ev);
 821         if (rc != 0) {
 822                 LASSERT (rc == -ENOMEM);
 823                 ev->ev_fired = 1;  /* no more event expected */
 824         }
 825         return rc;
 826 }
 827
 828 static int
 829 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 830 {
 831         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 832         srpc_event_t *ev = &rpc->crpc_bulkev;
 833         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 834         int           rc;
 835         int           opt;
 836
 837         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 838
 839         if (bk->bk_niov == 0) return 0; /* nothing to do */
 840
 841         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 842         opt |= LNET_MD_KIOV;
 843
 844         ev->ev_fired = 0;
 845         ev->ev_data  = rpc;
 846         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 847
 848         *id = srpc_next_id();
 849
 850         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 851                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 852                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 853         if (rc != 0) {
 854                 LASSERT (rc == -ENOMEM);
 855                 ev->ev_fired = 1;  /* no more event expected */
 856         }
 857         return rc;
 858 }
 859
 860 static int
 861 srpc_do_bulk (srpc_server_rpc_t *rpc)
 862 {
 863         srpc_event_t  *ev = &rpc->srpc_ev;
 864         srpc_bulk_t   *bk = rpc->srpc_bulk;
 865         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 866         int            rc;
 867         int            opt;
 868
 869         LASSERT (bk != NULL);
 870
 871         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 872         opt |= LNET_MD_KIOV;
 873
 874         ev->ev_fired = 0;
 875         ev->ev_data  = rpc;
 876         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 877
 878         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 879                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 880                                    rpc->srpc_peer, rpc->srpc_self,
 881                                    &bk->bk_mdh, ev);
 882         if (rc != 0)
 883                 ev->ev_fired = 1;  /* no more event expected */
 884         return rc;
 885 }
 886
 887 /* only called from srpc_handle_rpc */
 888 static void
 889 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
 890 {
 891         struct srpc_service_cd  *scd = rpc->srpc_scd;
 892         struct srpc_service     *sv  = scd->scd_svc;
 893         srpc_buffer_t           *buffer;
 894
 895         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 896
 897         rpc->srpc_status = status;
 898
 899         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 900                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 901                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 902                 swi_state2str(rpc->srpc_wi.swi_state), status);
 903
 904         if (status != 0) {
 905                 spin_lock(&srpc_data.rpc_glock);
 906                 srpc_data.rpc_counters.rpcs_dropped++;
 907                 spin_unlock(&srpc_data.rpc_glock);
 908         }
 909
 910         if (rpc->srpc_done != NULL)
 911                 (*rpc->srpc_done) (rpc);
 912         LASSERT(rpc->srpc_bulk == NULL);
 913
 914         spin_lock(&scd->scd_lock);
 915
 916         if (rpc->srpc_reqstbuf != NULL) {
 917                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 918                  * sv won't go away for scd_rpc_active must not be empty */
 919                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 920                 rpc->srpc_reqstbuf = NULL;
 921         }
 922
 923         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 924
 925         /*
 926          * No one can schedule me now since:
 927          * - I'm not on scd_rpc_active.
 928          * - all LNet events have been fired.
 929          * Cancel pending schedules and prevent future schedule attempts:
 930          */
 931         LASSERT(rpc->srpc_ev.ev_fired);
 932         swi_exit_workitem(&rpc->srpc_wi);
 933
 934         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 935                 buffer = list_entry(scd->scd_buf_blocked.next,
 936                                         srpc_buffer_t, buf_list);
 937                 list_del(&buffer->buf_list);
 938
 939                 srpc_init_server_rpc(rpc, scd, buffer);
 940                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 941                 swi_schedule_workitem(&rpc->srpc_wi);
 942         } else {
 943                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 944         }
 945
 946         spin_unlock(&scd->scd_lock);
 947         return;
 948 }
 949
 950 /* handles an incoming RPC */
 951 int
 952 srpc_handle_rpc(swi_workitem_t *wi)
 953 {
 954         struct srpc_server_rpc  *rpc = wi->swi_workitem.wi_data;
 955         struct srpc_service_cd  *scd = rpc->srpc_scd;
 956         struct srpc_service     *sv = scd->scd_svc;
 957         srpc_event_t            *ev = &rpc->srpc_ev;
 958         int                     rc = 0;
 959
 960         LASSERT(wi == &rpc->srpc_wi);
 961
 962         spin_lock(&scd->scd_lock);
 963
 964         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 965                 spin_unlock(&scd->scd_lock);
 966
 967                 if (rpc->srpc_bulk != NULL)
 968                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 969                 LNetMDUnlink(rpc->srpc_replymdh);
 970
 971                 if (ev->ev_fired) { /* no more event, OK to finish */
 972                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 973                         return 1;
 974                 }
 975                 return 0;
 976         }
 977
 978         spin_unlock(&scd->scd_lock);
 979
 980         switch (wi->swi_state) {
 981         default:
 982                 LBUG ();
 983         case SWI_STATE_NEWBORN: {
 984                 srpc_msg_t           *msg;
 985                 srpc_generic_reply_t *reply;
 986
 987                 msg = &rpc->srpc_reqstbuf->buf_msg;
 988                 reply = &rpc->srpc_replymsg.msg_body.reply;
 989
 990                 if (msg->msg_magic == 0) {
 991                         /* moaned already in srpc_lnet_ev_handler */
 992                         srpc_server_rpc_done(rpc, EBADMSG);
 993                         return 1;
 994                 }
 995
 996                 srpc_unpack_msg_hdr(msg);
 997                 if (msg->msg_version != SRPC_MSG_VERSION) {
 998                         CWARN("Version mismatch: %u, %u expected, from %s\n",
 999                               msg->msg_version, SRPC_MSG_VERSION,
1000                               libcfs_id2str(rpc->srpc_peer));
1001                         reply->status = EPROTO;
1002                         /* drop through and send reply */
1003                 } else {
1004                         reply->status = 0;
1005                         rc = (*sv->sv_handler)(rpc);
1006                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1007                         if (rc != 0) {
1008                                 srpc_server_rpc_done(rpc, rc);
1009                                 return 1;
1010                         }
1011                 }
1012
1013                 wi->swi_state = SWI_STATE_BULK_STARTED;
1014
1015                 if (rpc->srpc_bulk != NULL) {
1016                         rc = srpc_do_bulk(rpc);
1017                         if (rc == 0)
1018                                 return 0; /* wait for bulk */
1019
1020                         LASSERT (ev->ev_fired);
1021                         ev->ev_status = rc;
1022                 }
1023         }
1024         case SWI_STATE_BULK_STARTED:
1025                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1026
1027                 if (rpc->srpc_bulk != NULL) {
1028                         rc = ev->ev_status;
1029
1030                         if (sv->sv_bulk_ready != NULL)
1031                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1032
1033                         if (rc != 0) {
1034                                 srpc_server_rpc_done(rpc, rc);
1035                                 return 1;
1036                         }
1037                 }
1038
1039                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1040                 rc = srpc_send_reply(rpc);
1041                 if (rc == 0)
1042                         return 0; /* wait for reply */
1043                 srpc_server_rpc_done(rpc, rc);
1044                 return 1;
1045
1046         case SWI_STATE_REPLY_SUBMITTED:
1047                 if (!ev->ev_fired) {
1048                         CERROR("RPC %p: bulk %p, service %d\n",
1049                                rpc, rpc->srpc_bulk, sv->sv_id);
1050                         CERROR("Event: status %d, type %d, lnet %d\n",
1051                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1052                         LASSERT (ev->ev_fired);
1053                 }
1054
1055                 wi->swi_state = SWI_STATE_DONE;
1056                 srpc_server_rpc_done(rpc, ev->ev_status);
1057                 return 1;
1058         }
1059
1060         return 0;
1061 }
1062
1063 static void
1064 srpc_client_rpc_expired (void *data)
1065 {
1066         srpc_client_rpc_t *rpc = data;
1067
1068         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1069                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1070                rpc->crpc_timeout);
1071
1072         spin_lock(&rpc->crpc_lock);
1073
1074         rpc->crpc_timeout = 0;
1075         srpc_abort_rpc(rpc, -ETIMEDOUT);
1076
1077         spin_unlock(&rpc->crpc_lock);
1078
1079         spin_lock(&srpc_data.rpc_glock);
1080         srpc_data.rpc_counters.rpcs_expired++;
1081         spin_unlock(&srpc_data.rpc_glock);
1082 }
1083
1084 static void
1085 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1086 {
1087         stt_timer_t *timer = &rpc->crpc_timer;
1088
1089         if (rpc->crpc_timeout == 0)
1090                 return;
1091
1092         INIT_LIST_HEAD(&timer->stt_list);
1093         timer->stt_data    = rpc;
1094         timer->stt_func    = srpc_client_rpc_expired;
1095         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1096                                           cfs_time_current_sec());
1097         stt_add_timer(timer);
1098         return;
1099 }
1100
1101 /*
1102  * Called with rpc->crpc_lock held.
1103  *
1104  * Upon exit the RPC expiry timer is not queued and the handler is not
1105  * running on any CPU. */
1106 static void
1107 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1108 {
1109         /* timer not planted or already exploded */
1110         if (rpc->crpc_timeout == 0)
1111                 return;
1112
1113         /* timer successfully defused */
1114         if (stt_del_timer(&rpc->crpc_timer))
1115                 return;
1116
1117         /* timer detonated, wait for it to explode */
1118         while (rpc->crpc_timeout != 0) {
1119                 spin_unlock(&rpc->crpc_lock);
1120
1121                 schedule();
1122
1123                 spin_lock(&rpc->crpc_lock);
1124         }
1125 }
1126
1127 static void
1128 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1129 {
1130         swi_workitem_t *wi = &rpc->crpc_wi;
1131
1132         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1133
1134         spin_lock(&rpc->crpc_lock);
1135
1136         rpc->crpc_closed = 1;
1137         if (rpc->crpc_status == 0)
1138                 rpc->crpc_status = status;
1139
1140         srpc_del_client_rpc_timer(rpc);
1141
1142         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1143                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1144                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1145                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1146
1147         /*
1148          * No one can schedule me now since:
1149          * - RPC timer has been defused.
1150          * - all LNet events have been fired.
1151          * - crpc_closed has been set, preventing srpc_abort_rpc from
1152          *   scheduling me.
1153          * Cancel pending schedules and prevent future schedule attempts:
1154          */
1155         LASSERT (!srpc_event_pending(rpc));
1156         swi_exit_workitem(wi);
1157
1158         spin_unlock(&rpc->crpc_lock);
1159
1160         (*rpc->crpc_done)(rpc);
1161         return;
1162 }
1163
1164 /* sends an outgoing RPC */
1165 int
1166 srpc_send_rpc (swi_workitem_t *wi)
1167 {
1168         int                rc = 0;
1169         srpc_client_rpc_t *rpc;
1170         srpc_msg_t        *reply;
1171         int                do_bulk;
1172
1173         LASSERT(wi != NULL);
1174
1175         rpc = wi->swi_workitem.wi_data;
1176
1177         LASSERT (rpc != NULL);
1178         LASSERT (wi == &rpc->crpc_wi);
1179
1180         reply = &rpc->crpc_replymsg;
1181         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1182
1183         spin_lock(&rpc->crpc_lock);
1184
1185         if (rpc->crpc_aborted) {
1186                 spin_unlock(&rpc->crpc_lock);
1187                 goto abort;
1188         }
1189
1190         spin_unlock(&rpc->crpc_lock);
1191
1192         switch (wi->swi_state) {
1193         default:
1194                 LBUG ();
1195         case SWI_STATE_NEWBORN:
1196                 LASSERT (!srpc_event_pending(rpc));
1197
1198                 rc = srpc_prepare_reply(rpc);
1199                 if (rc != 0) {
1200                         srpc_client_rpc_done(rpc, rc);
1201                         return 1;
1202                 }
1203
1204                 rc = srpc_prepare_bulk(rpc);
1205                 if (rc != 0) break;
1206
1207                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1208                 rc = srpc_send_request(rpc);
1209                 break;
1210
1211         case SWI_STATE_REQUEST_SUBMITTED:
1212                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1213                  * order; however, they're processed in a strict order:
1214                  * rqt, rpy, and bulk. */
1215                 if (!rpc->crpc_reqstev.ev_fired) break;
1216
1217                 rc = rpc->crpc_reqstev.ev_status;
1218                 if (rc != 0) break;
1219
1220                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1221                 /* perhaps more events, fall thru */
1222         case SWI_STATE_REQUEST_SENT: {
1223                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1224
1225                 if (!rpc->crpc_replyev.ev_fired) break;
1226
1227                 rc = rpc->crpc_replyev.ev_status;
1228                 if (rc != 0) break;
1229
1230                 srpc_unpack_msg_hdr(reply);
1231                 if (reply->msg_type != type ||
1232                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1233                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1234                         CWARN ("Bad message from %s: type %u (%d expected),"
1235                                " magic %u (%d expected).\n",
1236                                libcfs_id2str(rpc->crpc_dest),
1237                                reply->msg_type, type,
1238                                reply->msg_magic, SRPC_MSG_MAGIC);
1239                         rc = -EBADMSG;
1240                         break;
1241                 }
1242
1243                 if (do_bulk && reply->msg_body.reply.status != 0) {
1244                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1245                                "case peer didn't initiate bulk transfer\n",
1246                                reply->msg_body.reply.status,
1247                                libcfs_id2str(rpc->crpc_dest));
1248                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1249                 }
1250
1251                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1252         }
1253         case SWI_STATE_REPLY_RECEIVED:
1254                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1255
1256                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1257
1258                 /* Bulk buffer was unlinked due to remote error. Clear error
1259                  * since reply buffer still contains valid data.
1260                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1261                  * remote error. */
1262                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1263                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1264                         rc = 0;
1265
1266                 wi->swi_state = SWI_STATE_DONE;
1267                 srpc_client_rpc_done(rpc, rc);
1268                 return 1;
1269         }
1270
1271         if (rc != 0) {
1272                 spin_lock(&rpc->crpc_lock);
1273                 srpc_abort_rpc(rpc, rc);
1274                 spin_unlock(&rpc->crpc_lock);
1275         }
1276
1277 abort:
1278         if (rpc->crpc_aborted) {
1279                 LNetMDUnlink(rpc->crpc_reqstmdh);
1280                 LNetMDUnlink(rpc->crpc_replymdh);
1281                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1282
1283                 if (!srpc_event_pending(rpc)) {
1284                         srpc_client_rpc_done(rpc, -EINTR);
1285                         return 1;
1286                 }
1287         }
1288         return 0;
1289 }
1290
1291 srpc_client_rpc_t *
1292 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1293                         int nbulkiov, int bulklen,
1294                         void (*rpc_done)(srpc_client_rpc_t *),
1295                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1296 {
1297         srpc_client_rpc_t *rpc;
1298
1299         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1300                                    crpc_bulk.bk_iovs[nbulkiov]));
1301         if (rpc == NULL)
1302                 return NULL;
1303
1304         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1305                              bulklen, rpc_done, rpc_fini, priv);
1306         return rpc;
1307 }
1308
1309 /* called with rpc->crpc_lock held */
1310 void
1311 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1312 {
1313         LASSERT (why != 0);
1314
1315         if (rpc->crpc_aborted || /* already aborted */
1316             rpc->crpc_closed)    /* callback imminent */
1317                 return;
1318
1319         CDEBUG (D_NET,
1320                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1321                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1322                 swi_state2str(rpc->crpc_wi.swi_state), why);
1323
1324         rpc->crpc_aborted = 1;
1325         rpc->crpc_status  = why;
1326         swi_schedule_workitem(&rpc->crpc_wi);
1327         return;
1328 }
1329
1330 /* called with rpc->crpc_lock held */
1331 void
1332 srpc_post_rpc (srpc_client_rpc_t *rpc)
1333 {
1334         LASSERT (!rpc->crpc_aborted);
1335         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1336
1337         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1338                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1339                 rpc->crpc_timeout);
1340
1341         srpc_add_client_rpc_timer(rpc);
1342         swi_schedule_workitem(&rpc->crpc_wi);
1343         return;
1344 }
1345
1346
1347 int
1348 srpc_send_reply(struct srpc_server_rpc *rpc)
1349 {
1350         srpc_event_t            *ev = &rpc->srpc_ev;
1351         struct srpc_msg         *msg = &rpc->srpc_replymsg;
1352         struct srpc_buffer      *buffer = rpc->srpc_reqstbuf;
1353         struct srpc_service_cd  *scd = rpc->srpc_scd;
1354         struct srpc_service     *sv = scd->scd_svc;
1355         __u64                   rpyid;
1356         int                     rc;
1357
1358         LASSERT(buffer != NULL);
1359         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1360
1361         spin_lock(&scd->scd_lock);
1362
1363         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1364                 /* Repost buffer before replying since test client
1365                  * might send me another RPC once it gets the reply */
1366                 if (srpc_service_post_buffer(scd, buffer) != 0)
1367                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1368                 rpc->srpc_reqstbuf = NULL;
1369         }
1370
1371         spin_unlock(&scd->scd_lock);
1372
1373         ev->ev_fired = 0;
1374         ev->ev_data  = rpc;
1375         ev->ev_type  = SRPC_REPLY_SENT;
1376
1377         msg->msg_magic   = SRPC_MSG_MAGIC;
1378         msg->msg_version = SRPC_MSG_VERSION;
1379         msg->msg_type    = srpc_service2reply(sv->sv_id);
1380
1381         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1382                                    sizeof(*msg), LNET_MD_OP_PUT,
1383                                    rpc->srpc_peer, rpc->srpc_self,
1384                                    &rpc->srpc_replymdh, ev);
1385         if (rc != 0)
1386                 ev->ev_fired = 1;  /* no more event expected */
1387         return rc;
1388 }
1389
1390 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1391 static void
1392 srpc_lnet_ev_handler(lnet_event_t *ev)
1393 {
1394         struct srpc_service_cd  *scd;
1395         srpc_event_t      *rpcev = ev->md.user_ptr;
1396         srpc_client_rpc_t *crpc;
1397         srpc_server_rpc_t *srpc;
1398         srpc_buffer_t     *buffer;
1399         srpc_service_t    *sv;
1400         srpc_msg_t        *msg;
1401         srpc_msg_type_t    type;
1402
1403         LASSERT (!in_interrupt());
1404
1405         if (ev->status != 0) {
1406                 __u32 errors;
1407
1408                 spin_lock(&srpc_data.rpc_glock);
1409                 if (ev->status != -ECANCELED) /* cancellation is not error */
1410                         srpc_data.rpc_counters.errors++;
1411                 errors = srpc_data.rpc_counters.errors;
1412                 spin_unlock(&srpc_data.rpc_glock);
1413
1414                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1415                         ev->status, ev->type, errors);
1416         }
1417
1418         rpcev->ev_lnet = ev->type;
1419
1420         switch (rpcev->ev_type) {
1421         default:
1422                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1423                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1424                 LBUG ();
1425         case SRPC_REQUEST_SENT:
1426                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1427                         spin_lock(&srpc_data.rpc_glock);
1428                         srpc_data.rpc_counters.rpcs_sent++;
1429                         spin_unlock(&srpc_data.rpc_glock);
1430                 }
1431         case SRPC_REPLY_RCVD:
1432         case SRPC_BULK_REQ_RCVD:
1433                 crpc = rpcev->ev_data;
1434
1435                 if (rpcev != &crpc->crpc_reqstev &&
1436                     rpcev != &crpc->crpc_replyev &&
1437                     rpcev != &crpc->crpc_bulkev) {
1438                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1439                                rpcev, crpc, &crpc->crpc_reqstev,
1440                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1441                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1442                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1443                         LBUG ();
1444                 }
1445
1446                 spin_lock(&crpc->crpc_lock);
1447
1448                 LASSERT(rpcev->ev_fired == 0);
1449                 rpcev->ev_fired  = 1;
1450                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1451                                                 -EINTR : ev->status;
1452                 swi_schedule_workitem(&crpc->crpc_wi);
1453
1454                 spin_unlock(&crpc->crpc_lock);
1455                 break;
1456
1457         case SRPC_REQUEST_RCVD:
1458                 scd = rpcev->ev_data;
1459                 sv = scd->scd_svc;
1460
1461                 LASSERT(rpcev == &scd->scd_ev);
1462
1463                 spin_lock(&scd->scd_lock);
1464
1465                 LASSERT (ev->unlinked);
1466                 LASSERT (ev->type == LNET_EVENT_PUT ||
1467                          ev->type == LNET_EVENT_UNLINK);
1468                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1469                          sv->sv_shuttingdown);
1470
1471                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1472                 buffer->buf_peer = ev->initiator;
1473                 buffer->buf_self = ev->target.nid;
1474
1475                 LASSERT(scd->scd_buf_nposted > 0);
1476                 scd->scd_buf_nposted--;
1477
1478                 if (sv->sv_shuttingdown) {
1479                         /* Leave buffer on scd->scd_buf_nposted since
1480                          * srpc_finish_service needs to traverse it. */
1481                         spin_unlock(&scd->scd_lock);
1482                         break;
1483                 }
1484
1485                 if (scd->scd_buf_err_stamp != 0 &&
1486                     scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1487                         /* re-enable adding buffer */
1488                         scd->scd_buf_err_stamp = 0;
1489                         scd->scd_buf_err = 0;
1490                 }
1491
1492                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1493                     scd->scd_buf_adjust == 0 &&
1494                     scd->scd_buf_nposted < scd->scd_buf_low) {
1495                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1496                                                   SFW_TEST_WI_MIN);
1497                         swi_schedule_workitem(&scd->scd_buf_wi);
1498                 }
1499
1500                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1501                 msg = &buffer->buf_msg;
1502                 type = srpc_service2request(sv->sv_id);
1503
1504                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1505                     (msg->msg_type != type &&
1506                      msg->msg_type != __swab32(type)) ||
1507                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1508                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1509                         CERROR ("Dropping RPC (%s) from %s: "
1510                                 "status %d mlength %d type %u magic %u.\n",
1511                                 sv->sv_name, libcfs_id2str(ev->initiator),
1512                                 ev->status, ev->mlength,
1513                                 msg->msg_type, msg->msg_magic);
1514
1515                         /* NB can't call srpc_service_recycle_buffer here since
1516                          * it may call LNetM[DE]Attach. The invalid magic tells
1517                          * srpc_handle_rpc to drop this RPC */
1518                         msg->msg_magic = 0;
1519                 }
1520
1521                 if (!list_empty(&scd->scd_rpc_free)) {
1522                         srpc = list_entry(scd->scd_rpc_free.next,
1523                                           struct srpc_server_rpc,
1524                                           srpc_list);
1525                         list_del(&srpc->srpc_list);
1526
1527                         srpc_init_server_rpc(srpc, scd, buffer);
1528                         list_add_tail(&srpc->srpc_list,
1529                                       &scd->scd_rpc_active);
1530                         swi_schedule_workitem(&srpc->srpc_wi);
1531                 } else {
1532                         list_add_tail(&buffer->buf_list,
1533                                       &scd->scd_buf_blocked);
1534                 }
1535
1536                 spin_unlock(&scd->scd_lock);
1537
1538                 spin_lock(&srpc_data.rpc_glock);
1539                 srpc_data.rpc_counters.rpcs_rcvd++;
1540                 spin_unlock(&srpc_data.rpc_glock);
1541                 break;
1542
1543         case SRPC_BULK_GET_RPLD:
1544                 LASSERT (ev->type == LNET_EVENT_SEND ||
1545                          ev->type == LNET_EVENT_REPLY ||
1546                          ev->type == LNET_EVENT_UNLINK);
1547
1548                 if (!ev->unlinked)
1549                         break; /* wait for final event */
1550
1551         case SRPC_BULK_PUT_SENT:
1552                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1553                         spin_lock(&srpc_data.rpc_glock);
1554
1555                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1556                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1557                         else
1558                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1559
1560                         spin_unlock(&srpc_data.rpc_glock);
1561                 }
1562         case SRPC_REPLY_SENT:
1563                 srpc = rpcev->ev_data;
1564                 scd  = srpc->srpc_scd;
1565
1566                 LASSERT(rpcev == &srpc->srpc_ev);
1567
1568                 spin_lock(&scd->scd_lock);
1569
1570                 rpcev->ev_fired  = 1;
1571                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1572                                    -EINTR : ev->status;
1573                 swi_schedule_workitem(&srpc->srpc_wi);
1574
1575                 spin_unlock(&scd->scd_lock);
1576                 break;
1577         }
1578 }
1579
1580
1581 int
1582 srpc_startup (void)
1583 {
1584         int rc;
1585
1586         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1587         spin_lock_init(&srpc_data.rpc_glock);
1588
1589         /* 1 second pause to avoid timestamp reuse */
1590         set_current_state(TASK_UNINTERRUPTIBLE);
1591         schedule_timeout(cfs_time_seconds(1));
1592         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1593
1594         srpc_data.rpc_state = SRPC_STATE_NONE;
1595
1596         rc = LNetNIInit(LNET_PID_LUSTRE);
1597         if (rc < 0) {
1598                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1599                 return rc;
1600         }
1601
1602         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1603
1604         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1605         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1606         if (rc != 0) {
1607                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1608                 goto bail;
1609         }
1610
1611         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1612         LASSERT(rc == 0);
1613         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1614         LASSERT(rc == 0);
1615
1616         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1617
1618         rc = stt_startup();
1619
1620 bail:
1621         if (rc != 0)
1622                 srpc_shutdown();
1623         else
1624                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1625
1626         return rc;
1627 }
1628
1629 void
1630 srpc_shutdown (void)
1631 {
1632         int i;
1633         int rc;
1634         int state;
1635
1636         state = srpc_data.rpc_state;
1637         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1638
1639         switch (state) {
1640         default:
1641                 LBUG ();
1642         case SRPC_STATE_RUNNING:
1643                 spin_lock(&srpc_data.rpc_glock);
1644
1645                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1646                         srpc_service_t *sv = srpc_data.rpc_services[i];
1647
1648                         LASSERTF (sv == NULL,
1649                                   "service not empty: id %d, name %s\n",
1650                                   i, sv->sv_name);
1651                 }
1652
1653                 spin_unlock(&srpc_data.rpc_glock);
1654
1655                 stt_shutdown();
1656
1657         case SRPC_STATE_EQ_INIT:
1658                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1659                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1660                 LASSERT (rc == 0);
1661                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1662                 LASSERT (rc == 0); /* the EQ should have no user by now */
1663
1664         case SRPC_STATE_NI_INIT:
1665                 LNetNIFini();
1666         }
1667
1668         return;
1669 }