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 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * lnet/selftest/rpc.c
  35  *
  36  * Author: Isaac Huang <isaac@clusterfs.com>
  37  *
  38  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  39  * - percpt data for service to improve smp performance
  40  * - code cleanup
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44
  45 #include "selftest.h"
  46
  47 typedef enum {
  48         SRPC_STATE_NONE,
  49         SRPC_STATE_NI_INIT,
  50         SRPC_STATE_EQ_INIT,
  51         SRPC_STATE_RUNNING,
  52         SRPC_STATE_STOPPING,
  53 } srpc_state_t;
  54
  55 struct smoketest_rpc {
  56         cfs_spinlock_t    rpc_glock;     /* global lock */
  57         srpc_service_t   *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  58         lnet_handle_eq_t  rpc_lnet_eq;   /* _the_ LNet event queue */
  59         srpc_state_t      rpc_state;
  60         srpc_counters_t   rpc_counters;
  61         __u64             rpc_matchbits; /* matchbits counter */
  62 } srpc_data;
  63
  64 static inline int
  65 srpc_serv_portal(int svc_id)
  66 {
  67         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  68                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  69 }
  70
  71 /* forward ref's */
  72 int srpc_handle_rpc (swi_workitem_t *wi);
  73
  74 void srpc_get_counters (srpc_counters_t *cnt)
  75 {
  76         cfs_spin_lock(&srpc_data.rpc_glock);
  77         *cnt = srpc_data.rpc_counters;
  78         cfs_spin_unlock(&srpc_data.rpc_glock);
  79 }
  80
  81 void srpc_set_counters (const srpc_counters_t *cnt)
  82 {
  83         cfs_spin_lock(&srpc_data.rpc_glock);
  84         srpc_data.rpc_counters = *cnt;
  85         cfs_spin_unlock(&srpc_data.rpc_glock);
  86 }
  87
  88 void
  89 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
  90 {
  91         LASSERT (i >= 0 && i < bk->bk_niov);
  92
  93 #ifdef __KERNEL__
  94         bk->bk_iovs[i].kiov_offset = 0;
  95         bk->bk_iovs[i].kiov_page   = pg;
  96         bk->bk_iovs[i].kiov_len    = CFS_PAGE_SIZE;
  97 #else
  98         LASSERT (bk->bk_pages != NULL);
  99
 100         bk->bk_pages[i] = pg;
 101         bk->bk_iovs[i].iov_len  = CFS_PAGE_SIZE;
 102         bk->bk_iovs[i].iov_base = cfs_page_address(pg);
 103 #endif
 104         return;
 105 }
 106
 107 void
 108 srpc_free_bulk (srpc_bulk_t *bk)
 109 {
 110         int         i;
 111         cfs_page_t *pg;
 112
 113         LASSERT (bk != NULL);
 114 #ifndef __KERNEL__
 115         LASSERT (bk->bk_pages != NULL);
 116 #endif
 117
 118         for (i = 0; i < bk->bk_niov; i++) {
 119 #ifdef __KERNEL__
 120                 pg = bk->bk_iovs[i].kiov_page;
 121 #else
 122                 pg = bk->bk_pages[i];
 123 #endif
 124                 if (pg == NULL) break;
 125
 126                 cfs_free_page(pg);
 127         }
 128
 129 #ifndef __KERNEL__
 130         LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
 131 #endif
 132         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 133         return;
 134 }
 135
 136 srpc_bulk_t *
 137 srpc_alloc_bulk(int cpt, int npages, int sink)
 138 {
 139         srpc_bulk_t  *bk;
 140         cfs_page_t  **pages;
 141         int           i;
 142
 143         LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
 144
 145         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 146                          offsetof(srpc_bulk_t, bk_iovs[npages]));
 147         if (bk == NULL) {
 148                 CERROR ("Can't allocate descriptor for %d pages\n", npages);
 149                 return NULL;
 150         }
 151
 152         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
 153         bk->bk_sink = sink;
 154         bk->bk_niov = npages;
 155         bk->bk_len  = npages * CFS_PAGE_SIZE;
 156 #ifndef __KERNEL__
 157         LIBCFS_CPT_ALLOC(pages, lnet_cpt_table(), cpt,
 158                          sizeof(cfs_page_t *) * npages);
 159         if (pages == NULL) {
 160                 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 161                 CERROR ("Can't allocate page array for %d pages\n", npages);
 162                 return NULL;
 163         }
 164
 165         memset(pages, 0, sizeof(cfs_page_t *) * npages);
 166         bk->bk_pages = pages;
 167 #else
 168         UNUSED (pages);
 169 #endif
 170
 171         for (i = 0; i < npages; i++) {
 172                 cfs_page_t *pg;
 173
 174                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, CFS_ALLOC_STD);
 175                 if (pg == NULL) {
 176                         CERROR ("Can't allocate page %d of %d\n", i, npages);
 177                         srpc_free_bulk(bk);
 178                         return NULL;
 179                 }
 180
 181                 srpc_add_bulk_page(bk, pg, i);
 182         }
 183
 184         return bk;
 185 }
 186
 187 static inline __u64
 188 srpc_next_id (void)
 189 {
 190         __u64 id;
 191
 192         cfs_spin_lock(&srpc_data.rpc_glock);
 193         id = srpc_data.rpc_matchbits++;
 194         cfs_spin_unlock(&srpc_data.rpc_glock);
 195         return id;
 196 }
 197
 198 void
 199 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 200                      struct srpc_service_cd *scd,
 201                      struct srpc_buffer *buffer)
 202 {
 203         memset(rpc, 0, sizeof(*rpc));
 204         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
 205                           srpc_serv_is_framework(scd->scd_svc) ?
 206                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 207
 208         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 209
 210         rpc->srpc_scd      = scd;
 211         rpc->srpc_reqstbuf = buffer;
 212         rpc->srpc_peer     = buffer->buf_peer;
 213         rpc->srpc_self     = buffer->buf_self;
 214         LNetInvalidateHandle(&rpc->srpc_replymdh);
 215 }
 216
 217 static void
 218 srpc_service_fini(struct srpc_service *svc)
 219 {
 220         struct srpc_service_cd  *scd;
 221         struct srpc_server_rpc  *rpc;
 222         struct srpc_buffer      *buf;
 223         cfs_list_t              *q;
 224         int                     i;
 225
 226         if (svc->sv_cpt_data == NULL)
 227                 return;
 228
 229         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 230                 while (1) {
 231                         if (!cfs_list_empty(&scd->scd_buf_posted))
 232                                 q = &scd->scd_buf_posted;
 233                         else if (!cfs_list_empty(&scd->scd_buf_blocked))
 234                                 q = &scd->scd_buf_blocked;
 235                         else
 236                                 break;
 237
 238                         while (!cfs_list_empty(q)) {
 239                                 buf = cfs_list_entry(q->next,
 240                                                      struct srpc_buffer,
 241                                                      buf_list);
 242                                 cfs_list_del(&buf->buf_list);
 243                                 LIBCFS_FREE(buf, sizeof(*buf));
 244                         }
 245                 }
 246
 247                 LASSERT(cfs_list_empty(&scd->scd_rpc_active));
 248
 249                 while (!cfs_list_empty(&scd->scd_rpc_free)) {
 250                         rpc = cfs_list_entry(scd->scd_rpc_free.next,
 251                                              struct srpc_server_rpc,
 252                                              srpc_list);
 253                         cfs_list_del(&rpc->srpc_list);
 254                         LIBCFS_FREE(rpc, sizeof(*rpc));
 255                 }
 256         }
 257
 258         cfs_percpt_free(svc->sv_cpt_data);
 259         svc->sv_cpt_data = NULL;
 260 }
 261
 262 static int
 263 srpc_service_nrpcs(struct srpc_service *svc)
 264 {
 265         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 266
 267         return srpc_serv_is_framework(svc) ?
 268                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 269 }
 270
 271 int srpc_add_buffer(struct swi_workitem *wi);
 272
 273 static int
 274 srpc_service_init(struct srpc_service *svc)
 275 {
 276         struct srpc_service_cd  *scd;
 277         struct srpc_server_rpc  *rpc;
 278         int                     nrpcs;
 279         int                     i;
 280         int                     j;
 281
 282         svc->sv_shuttingdown = 0;
 283
 284         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 285                                             sizeof(struct srpc_service_cd));
 286         if (svc->sv_cpt_data == NULL)
 287                 return -ENOMEM;
 288
 289         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 290                         1 : cfs_cpt_number(lnet_cpt_table());
 291         nrpcs = srpc_service_nrpcs(svc);
 292
 293         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 294                 scd->scd_cpt = i;
 295                 scd->scd_svc = svc;
 296                 cfs_spin_lock_init(&scd->scd_lock);
 297                 CFS_INIT_LIST_HEAD(&scd->scd_rpc_free);
 298                 CFS_INIT_LIST_HEAD(&scd->scd_rpc_active);
 299                 CFS_INIT_LIST_HEAD(&scd->scd_buf_posted);
 300                 CFS_INIT_LIST_HEAD(&scd->scd_buf_blocked);
 301
 302                 scd->scd_ev.ev_data = scd;
 303                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 304
 305                 /* NB: don't use lst_sched_serial for adding buffer,
 306                  * see details in srpc_service_add_buffers() */
 307                 swi_init_workitem(&scd->scd_buf_wi, scd,
 308                                   srpc_add_buffer, lst_sched_test[i]);
 309
 310                 if (i != 0 && srpc_serv_is_framework(svc)) {
 311                         /* NB: framework service only needs srpc_service_cd for
 312                          * one partition, but we allocate for all to make
 313                          * it easier to implement, it will waste a little
 314                          * memory but nobody should care about this */
 315                         continue;
 316                 }
 317
 318                 for (j = 0; j < nrpcs; j++) {
 319                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 320                                          i, sizeof(*rpc));
 321                         if (rpc == NULL) {
 322                                 srpc_service_fini(svc);
 323                                 return -ENOMEM;
 324                         }
 325                         cfs_list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 326                 }
 327         }
 328
 329         return 0;
 330 }
 331
 332 int
 333 srpc_add_service(struct srpc_service *sv)
 334 {
 335         int id = sv->sv_id;
 336
 337         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 338
 339         if (srpc_service_init(sv) != 0)
 340                 return -ENOMEM;
 341
 342         cfs_spin_lock(&srpc_data.rpc_glock);
 343
 344         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 345
 346         if (srpc_data.rpc_services[id] != NULL) {
 347                 cfs_spin_unlock(&srpc_data.rpc_glock);
 348                 goto failed;
 349         }
 350
 351         srpc_data.rpc_services[id] = sv;
 352         cfs_spin_unlock(&srpc_data.rpc_glock);
 353
 354         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 355         return 0;
 356
 357  failed:
 358         srpc_service_fini(sv);
 359         return -EBUSY;
 360 }
 361
 362 int
 363 srpc_remove_service (srpc_service_t *sv)
 364 {
 365         int id = sv->sv_id;
 366
 367         cfs_spin_lock(&srpc_data.rpc_glock);
 368
 369         if (srpc_data.rpc_services[id] != sv) {
 370                 cfs_spin_unlock(&srpc_data.rpc_glock);
 371                 return -ENOENT;
 372         }
 373
 374         srpc_data.rpc_services[id] = NULL;
 375         cfs_spin_unlock(&srpc_data.rpc_glock);
 376         return 0;
 377 }
 378
 379 int
 380 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 381                        int len, int options, lnet_process_id_t peer,
 382                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 383 {
 384         int              rc;
 385         lnet_md_t        md;
 386         lnet_handle_me_t meh;
 387
 388         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 389                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 390         if (rc != 0) {
 391                 CERROR ("LNetMEAttach failed: %d\n", rc);
 392                 LASSERT (rc == -ENOMEM);
 393                 return -ENOMEM;
 394         }
 395
 396         md.threshold = 1;
 397         md.user_ptr  = ev;
 398         md.start     = buf;
 399         md.length    = len;
 400         md.options   = options;
 401         md.eq_handle = srpc_data.rpc_lnet_eq;
 402
 403         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 404         if (rc != 0) {
 405                 CERROR ("LNetMDAttach failed: %d\n", rc);
 406                 LASSERT (rc == -ENOMEM);
 407
 408                 rc = LNetMEUnlink(meh);
 409                 LASSERT (rc == 0);
 410                 return -ENOMEM;
 411         }
 412
 413         CDEBUG (D_NET,
 414                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 415                 libcfs_id2str(peer), portal, matchbits);
 416         return 0;
 417 }
 418
 419 int
 420 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 421                       int options, lnet_process_id_t peer, lnet_nid_t self,
 422                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 423 {
 424         int       rc;
 425         lnet_md_t md;
 426
 427         md.user_ptr  = ev;
 428         md.start     = buf;
 429         md.length    = len;
 430         md.eq_handle = srpc_data.rpc_lnet_eq;
 431         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 432         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 433
 434         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 435         if (rc != 0) {
 436                 CERROR ("LNetMDBind failed: %d\n", rc);
 437                 LASSERT (rc == -ENOMEM);
 438                 return -ENOMEM;
 439         }
 440
 441         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 442          * they're only meaningful for MDs attached to an ME (i.e. passive
 443          * buffers... */
 444         if ((options & LNET_MD_OP_PUT) != 0) {
 445                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 446                              portal, matchbits, 0, 0);
 447         } else {
 448                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 449
 450                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 451         }
 452
 453         if (rc != 0) {
 454                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 455                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 456                         libcfs_id2str(peer), portal, matchbits, rc);
 457
 458                 /* The forthcoming unlink event will complete this operation
 459                  * with failure, so fall through and return success here.
 460                  */
 461                 rc = LNetMDUnlink(*mdh);
 462                 LASSERT (rc == 0);
 463         } else {
 464                 CDEBUG (D_NET,
 465                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 466                         libcfs_id2str(peer), portal, matchbits);
 467         }
 468         return 0;
 469 }
 470
 471 int
 472 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 473                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 474 {
 475         return srpc_post_active_rdma(srpc_serv_portal(service), service,
 476                                      buf, len, LNET_MD_OP_PUT, peer,
 477                                      LNET_NID_ANY, mdh, ev);
 478 }
 479
 480 int
 481 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 482                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 483 {
 484         lnet_process_id_t any = {0};
 485
 486         any.nid = LNET_NID_ANY;
 487         any.pid = LNET_PID_ANY;
 488
 489         return srpc_post_passive_rdma(srpc_serv_portal(service),
 490                                       local, service, buf, len,
 491                                       LNET_MD_OP_PUT, any, mdh, ev);
 492 }
 493
 494 int
 495 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 496 {
 497         struct srpc_service     *sv = scd->scd_svc;
 498         struct srpc_msg         *msg = &buf->buf_msg;
 499         int                     rc;
 500
 501         LNetInvalidateHandle(&buf->buf_mdh);
 502         cfs_list_add(&buf->buf_list, &scd->scd_buf_posted);
 503         scd->scd_buf_nposted++;
 504         cfs_spin_unlock(&scd->scd_lock);
 505
 506         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 507                                       !srpc_serv_is_framework(sv),
 508                                       msg, sizeof(*msg), &buf->buf_mdh,
 509                                       &scd->scd_ev);
 510
 511         /* At this point, a RPC (new or delayed) may have arrived in
 512          * msg and its event handler has been called. So we must add
 513          * buf to scd_buf_posted _before_ dropping scd_lock */
 514
 515         cfs_spin_lock(&scd->scd_lock);
 516
 517         if (rc == 0) {
 518                 if (!sv->sv_shuttingdown)
 519                         return 0;
 520
 521                 cfs_spin_unlock(&scd->scd_lock);
 522                 /* srpc_shutdown_service might have tried to unlink me
 523                  * when my buf_mdh was still invalid */
 524                 LNetMDUnlink(buf->buf_mdh);
 525                 cfs_spin_lock(&scd->scd_lock);
 526                 return 0;
 527         }
 528
 529         scd->scd_buf_nposted--;
 530         if (sv->sv_shuttingdown)
 531                 return rc; /* don't allow to change scd_buf_posted */
 532
 533         cfs_list_del(&buf->buf_list);
 534         cfs_spin_unlock(&scd->scd_lock);
 535
 536         LIBCFS_FREE(buf, sizeof(*buf));
 537
 538         cfs_spin_lock(&scd->scd_lock);
 539         return rc;
 540 }
 541
 542 int
 543 srpc_add_buffer(struct swi_workitem *wi)
 544 {
 545         struct srpc_service_cd  *scd = wi->swi_workitem.wi_data;
 546         struct srpc_buffer      *buf;
 547         int                     rc = 0;
 548
 549         /* it's called by workitem scheduler threads, these threads
 550          * should have been set CPT affinity, so buffers will be posted
 551          * on CPT local list of Portal */
 552         cfs_spin_lock(&scd->scd_lock);
 553
 554         while (scd->scd_buf_adjust > 0 &&
 555                !scd->scd_svc->sv_shuttingdown) {
 556                 scd->scd_buf_adjust--; /* consume it */
 557                 scd->scd_buf_posting++;
 558
 559                 cfs_spin_unlock(&scd->scd_lock);
 560
 561                 LIBCFS_ALLOC(buf, sizeof(*buf));
 562                 if (buf == NULL) {
 563                         CERROR("Failed to add new buf to service: %s\n",
 564                                scd->scd_svc->sv_name);
 565                         cfs_spin_lock(&scd->scd_lock);
 566                         rc = -ENOMEM;
 567                         break;
 568                 }
 569
 570                 cfs_spin_lock(&scd->scd_lock);
 571                 if (scd->scd_svc->sv_shuttingdown) {
 572                         cfs_spin_unlock(&scd->scd_lock);
 573                         LIBCFS_FREE(buf, sizeof(*buf));
 574
 575                         cfs_spin_lock(&scd->scd_lock);
 576                         rc = -ESHUTDOWN;
 577                         break;
 578                 }
 579
 580                 rc = srpc_service_post_buffer(scd, buf);
 581                 if (rc != 0)
 582                         break; /* buf has been freed inside */
 583
 584                 LASSERT(scd->scd_buf_posting > 0);
 585                 scd->scd_buf_posting--;
 586                 scd->scd_buf_total++;
 587                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 588         }
 589
 590         if (rc != 0) {
 591                 scd->scd_buf_err_stamp = cfs_time_current_sec();
 592                 scd->scd_buf_err = rc;
 593
 594                 LASSERT(scd->scd_buf_posting > 0);
 595                 scd->scd_buf_posting--;
 596         }
 597
 598         cfs_spin_unlock(&scd->scd_lock);
 599         return 0;
 600 }
 601
 602 int
 603 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 604 {
 605         struct srpc_service_cd  *scd;
 606         int                     rc = 0;
 607         int                     i;
 608
 609         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 610
 611         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 612                 cfs_spin_lock(&scd->scd_lock);
 613
 614                 scd->scd_buf_err = 0;
 615                 scd->scd_buf_err_stamp = 0;
 616                 scd->scd_buf_posting = 0;
 617                 scd->scd_buf_adjust = nbuffer;
 618                 /* start to post buffers */
 619                 swi_schedule_workitem(&scd->scd_buf_wi);
 620                 cfs_spin_unlock(&scd->scd_lock);
 621
 622                 /* framework service only post buffer for one partition  */
 623                 if (srpc_serv_is_framework(sv))
 624                         break;
 625         }
 626
 627         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 628                 cfs_spin_lock(&scd->scd_lock);
 629                 /*
 630                  * NB: srpc_service_add_buffers() can be called inside
 631                  * thread context of lst_sched_serial, and we don't normally
 632                  * allow to sleep inside thread context of WI scheduler
 633                  * because it will block current scheduler thread from doing
 634                  * anything else, even worse, it could deadlock if it's
 635                  * waiting on result from another WI of the same scheduler.
 636                  * However, it's safe at here because scd_buf_wi is scheduled
 637                  * by thread in a different WI scheduler (lst_sched_test),
 638                  * so we don't have any risk of deadlock, though this could
 639                  * block all WIs pending on lst_sched_serial for a moment
 640                  * which is not good but not fatal.
 641                  */
 642                 lst_wait_until(scd->scd_buf_err != 0 ||
 643                                (scd->scd_buf_adjust == 0 &&
 644                                 scd->scd_buf_posting == 0),
 645                                scd->scd_lock, "waiting for adding buffer\n");
 646
 647                 if (scd->scd_buf_err != 0 && rc == 0)
 648                         rc = scd->scd_buf_err;
 649
 650                 cfs_spin_unlock(&scd->scd_lock);
 651         }
 652
 653         return rc;
 654 }
 655
 656 void
 657 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 658 {
 659         struct srpc_service_cd  *scd;
 660         int                     num;
 661         int                     i;
 662
 663         LASSERT(!sv->sv_shuttingdown);
 664
 665         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 666                 cfs_spin_lock(&scd->scd_lock);
 667
 668                 num = scd->scd_buf_total + scd->scd_buf_posting;
 669                 scd->scd_buf_adjust -= min(nbuffer, num);
 670
 671                 cfs_spin_unlock(&scd->scd_lock);
 672         }
 673 }
 674
 675 /* returns 1 if sv has finished, otherwise 0 */
 676 int
 677 srpc_finish_service(struct srpc_service *sv)
 678 {
 679         struct srpc_service_cd  *scd;
 680         struct srpc_server_rpc  *rpc;
 681         int                     i;
 682
 683         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 684
 685         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 686                 cfs_spin_lock(&scd->scd_lock);
 687                 if (!swi_deschedule_workitem(&scd->scd_buf_wi))
 688                         return 0;
 689
 690                 if (scd->scd_buf_nposted > 0) {
 691                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink",
 692                                scd->scd_buf_nposted);
 693                         cfs_spin_unlock(&scd->scd_lock);
 694                         return 0;
 695                 }
 696
 697                 if (cfs_list_empty(&scd->scd_rpc_active)) {
 698                         cfs_spin_unlock(&scd->scd_lock);
 699                         continue;
 700                 }
 701
 702                 rpc = cfs_list_entry(scd->scd_rpc_active.next,
 703                                      struct srpc_server_rpc, srpc_list);
 704                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
 705                         "wi %s scheduled %d running %d, "
 706                         "ev fired %d type %d status %d lnet %d\n",
 707                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 708                         swi_state2str(rpc->srpc_wi.swi_state),
 709                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 710                         rpc->srpc_wi.swi_workitem.wi_running,
 711                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 712                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 713                 cfs_spin_unlock(&scd->scd_lock);
 714                 return 0;
 715         }
 716
 717         /* no lock needed from now on */
 718         srpc_service_fini(sv);
 719         return 1;
 720 }
 721
 722 /* called with sv->sv_lock held */
 723 void
 724 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
 725 {
 726         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 727                 if (srpc_service_post_buffer(scd, buf) != 0) {
 728                         CWARN("Failed to post %s buffer\n",
 729                               scd->scd_svc->sv_name);
 730                 }
 731                 return;
 732         }
 733
 734         /* service is shutting down, or we want to recycle some buffers */
 735         scd->scd_buf_total--;
 736
 737         if (scd->scd_buf_adjust < 0) {
 738                 scd->scd_buf_adjust++;
 739                 if (scd->scd_buf_adjust < 0 &&
 740                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 741                         CDEBUG(D_INFO,
 742                                "Try to recyle %d buffers but nothing left\n",
 743                                scd->scd_buf_adjust);
 744                         scd->scd_buf_adjust = 0;
 745                 }
 746         }
 747
 748         cfs_spin_unlock(&scd->scd_lock);
 749         LIBCFS_FREE(buf, sizeof(*buf));
 750         cfs_spin_lock(&scd->scd_lock);
 751 }
 752
 753 void
 754 srpc_abort_service(struct srpc_service *sv)
 755 {
 756         struct srpc_service_cd  *scd;
 757         struct srpc_server_rpc  *rpc;
 758         int                     i;
 759
 760         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 761                sv->sv_id, sv->sv_name);
 762
 763         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 764                 cfs_spin_lock(&scd->scd_lock);
 765
 766                 /* schedule in-flight RPCs to notice the abort, NB:
 767                  * racing with incoming RPCs; complete fix should make test
 768                  * RPCs carry session ID in its headers */
 769                 cfs_list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 770                         rpc->srpc_aborted = 1;
 771                         swi_schedule_workitem(&rpc->srpc_wi);
 772                 }
 773
 774                 cfs_spin_unlock(&scd->scd_lock);
 775         }
 776 }
 777
 778 void
 779 srpc_shutdown_service(srpc_service_t *sv)
 780 {
 781         struct srpc_service_cd  *scd;
 782         struct srpc_server_rpc  *rpc;
 783         srpc_buffer_t           *buf;
 784         int                     i;
 785
 786         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 787                sv->sv_id, sv->sv_name);
 788
 789         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 790                 cfs_spin_lock(&scd->scd_lock);
 791
 792         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 793
 794         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 795                 cfs_spin_unlock(&scd->scd_lock);
 796
 797         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 798                 cfs_spin_lock(&scd->scd_lock);
 799
 800                 /* schedule in-flight RPCs to notice the shutdown */
 801                 cfs_list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 802                         swi_schedule_workitem(&rpc->srpc_wi);
 803
 804                 cfs_spin_unlock(&scd->scd_lock);
 805
 806                 /* OK to traverse scd_buf_posted without lock, since no one
 807                  * touches scd_buf_posted now */
 808                 cfs_list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 809                         LNetMDUnlink(buf->buf_mdh);
 810         }
 811 }
 812
 813 int
 814 srpc_send_request (srpc_client_rpc_t *rpc)
 815 {
 816         srpc_event_t *ev = &rpc->crpc_reqstev;
 817         int           rc;
 818
 819         ev->ev_fired = 0;
 820         ev->ev_data  = rpc;
 821         ev->ev_type  = SRPC_REQUEST_SENT;
 822
 823         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 824                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 825                                      &rpc->crpc_reqstmdh, 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 int
 834 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 835 {
 836         srpc_event_t *ev = &rpc->crpc_replyev;
 837         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 838         int           rc;
 839
 840         ev->ev_fired = 0;
 841         ev->ev_data  = rpc;
 842         ev->ev_type  = SRPC_REPLY_RCVD;
 843
 844         *id = srpc_next_id();
 845
 846         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 847                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 848                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 849                                     &rpc->crpc_replymdh, ev);
 850         if (rc != 0) {
 851                 LASSERT (rc == -ENOMEM);
 852                 ev->ev_fired = 1;  /* no more event expected */
 853         }
 854         return rc;
 855 }
 856
 857 int
 858 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 859 {
 860         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 861         srpc_event_t *ev = &rpc->crpc_bulkev;
 862         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 863         int           rc;
 864         int           opt;
 865
 866         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 867
 868         if (bk->bk_niov == 0) return 0; /* nothing to do */
 869
 870         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 871 #ifdef __KERNEL__
 872         opt |= LNET_MD_KIOV;
 873 #else
 874         opt |= LNET_MD_IOVEC;
 875 #endif
 876
 877         ev->ev_fired = 0;
 878         ev->ev_data  = rpc;
 879         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 880
 881         *id = srpc_next_id();
 882
 883         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 884                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 885                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 886         if (rc != 0) {
 887                 LASSERT (rc == -ENOMEM);
 888                 ev->ev_fired = 1;  /* no more event expected */
 889         }
 890         return rc;
 891 }
 892
 893 int
 894 srpc_do_bulk (srpc_server_rpc_t *rpc)
 895 {
 896         srpc_event_t  *ev = &rpc->srpc_ev;
 897         srpc_bulk_t   *bk = rpc->srpc_bulk;
 898         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 899         int            rc;
 900         int            opt;
 901
 902         LASSERT (bk != NULL);
 903
 904         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 905 #ifdef __KERNEL__
 906         opt |= LNET_MD_KIOV;
 907 #else
 908         opt |= LNET_MD_IOVEC;
 909 #endif
 910
 911         ev->ev_fired = 0;
 912         ev->ev_data  = rpc;
 913         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 914
 915         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 916                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 917                                    rpc->srpc_peer, rpc->srpc_self,
 918                                    &bk->bk_mdh, ev);
 919         if (rc != 0)
 920                 ev->ev_fired = 1;  /* no more event expected */
 921         return rc;
 922 }
 923
 924 /* only called from srpc_handle_rpc */
 925 void
 926 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
 927 {
 928         struct srpc_service_cd  *scd = rpc->srpc_scd;
 929         struct srpc_service     *sv  = scd->scd_svc;
 930         srpc_buffer_t           *buffer;
 931
 932         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 933
 934         rpc->srpc_status = status;
 935
 936         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 937                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 938                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 939                 swi_state2str(rpc->srpc_wi.swi_state), status);
 940
 941         if (status != 0) {
 942                 cfs_spin_lock(&srpc_data.rpc_glock);
 943                 srpc_data.rpc_counters.rpcs_dropped++;
 944                 cfs_spin_unlock(&srpc_data.rpc_glock);
 945         }
 946
 947         if (rpc->srpc_done != NULL)
 948                 (*rpc->srpc_done) (rpc);
 949         LASSERT (rpc->srpc_bulk == NULL);
 950
 951         cfs_spin_lock(&scd->scd_lock);
 952
 953         if (rpc->srpc_reqstbuf != NULL) {
 954                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 955                  * sv won't go away for scd_rpc_active must not be empty */
 956                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 957                 rpc->srpc_reqstbuf = NULL;
 958         }
 959
 960         cfs_list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 961
 962         /*
 963          * No one can schedule me now since:
 964          * - I'm not on scd_rpc_active.
 965          * - all LNet events have been fired.
 966          * Cancel pending schedules and prevent future schedule attempts:
 967          */
 968         LASSERT(rpc->srpc_ev.ev_fired);
 969         swi_exit_workitem(&rpc->srpc_wi);
 970
 971         if (!sv->sv_shuttingdown && !cfs_list_empty(&scd->scd_buf_blocked)) {
 972                 buffer = cfs_list_entry(scd->scd_buf_blocked.next,
 973                                         srpc_buffer_t, buf_list);
 974                 cfs_list_del(&buffer->buf_list);
 975
 976                 srpc_init_server_rpc(rpc, scd, buffer);
 977                 cfs_list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 978                 swi_schedule_workitem(&rpc->srpc_wi);
 979         } else {
 980                 cfs_list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 981         }
 982
 983         cfs_spin_unlock(&scd->scd_lock);
 984         return;
 985 }
 986
 987 /* handles an incoming RPC */
 988 int
 989 srpc_handle_rpc(swi_workitem_t *wi)
 990 {
 991         struct srpc_server_rpc  *rpc = wi->swi_workitem.wi_data;
 992         struct srpc_service_cd  *scd = rpc->srpc_scd;
 993         struct srpc_service     *sv = scd->scd_svc;
 994         srpc_event_t            *ev = &rpc->srpc_ev;
 995         int                     rc = 0;
 996
 997         LASSERT(wi == &rpc->srpc_wi);
 998
 999         cfs_spin_lock(&scd->scd_lock);
1000
1001         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
1002                 cfs_spin_unlock(&scd->scd_lock);
1003
1004                 if (rpc->srpc_bulk != NULL)
1005                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
1006                 LNetMDUnlink(rpc->srpc_replymdh);
1007
1008                 if (ev->ev_fired) { /* no more event, OK to finish */
1009                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
1010                         return 1;
1011                 }
1012                 return 0;
1013         }
1014
1015         cfs_spin_unlock(&scd->scd_lock);
1016
1017         switch (wi->swi_state) {
1018         default:
1019                 LBUG ();
1020         case SWI_STATE_NEWBORN: {
1021                 srpc_msg_t           *msg;
1022                 srpc_generic_reply_t *reply;
1023
1024                 msg = &rpc->srpc_reqstbuf->buf_msg;
1025                 reply = &rpc->srpc_replymsg.msg_body.reply;
1026
1027                 if (msg->msg_magic == 0) {
1028                         /* moaned already in srpc_lnet_ev_handler */
1029                         rc = EBADMSG;
1030                 } else if (msg->msg_version != SRPC_MSG_VERSION &&
1031                            msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
1032                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
1033                                msg->msg_version, SRPC_MSG_VERSION,
1034                                libcfs_id2str(rpc->srpc_peer));
1035                         reply->status = EPROTO;
1036                 } else {
1037                         reply->status = 0;
1038                         rc = (*sv->sv_handler) (rpc);
1039                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
1040                 }
1041
1042                 if (rc != 0) {
1043                         srpc_server_rpc_done(rpc, rc);
1044                         return 1;
1045                 }
1046
1047                 wi->swi_state = SWI_STATE_BULK_STARTED;
1048
1049                 if (rpc->srpc_bulk != NULL) {
1050                         rc = srpc_do_bulk(rpc);
1051                         if (rc == 0)
1052                                 return 0; /* wait for bulk */
1053
1054                         LASSERT (ev->ev_fired);
1055                         ev->ev_status = rc;
1056                 }
1057         }
1058         case SWI_STATE_BULK_STARTED:
1059                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1060
1061                 if (rpc->srpc_bulk != NULL) {
1062                         rc = ev->ev_status;
1063
1064                         if (sv->sv_bulk_ready != NULL)
1065                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1066
1067                         if (rc != 0) {
1068                                 srpc_server_rpc_done(rpc, rc);
1069                                 return 1;
1070                         }
1071                 }
1072
1073                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1074                 rc = srpc_send_reply(rpc);
1075                 if (rc == 0)
1076                         return 0; /* wait for reply */
1077                 srpc_server_rpc_done(rpc, rc);
1078                 return 1;
1079
1080         case SWI_STATE_REPLY_SUBMITTED:
1081                 if (!ev->ev_fired) {
1082                         CERROR("RPC %p: bulk %p, service %d\n",
1083                                rpc, rpc->srpc_bulk, sv->sv_id);
1084                         CERROR("Event: status %d, type %d, lnet %d\n",
1085                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1086                         LASSERT (ev->ev_fired);
1087                 }
1088
1089                 wi->swi_state = SWI_STATE_DONE;
1090                 srpc_server_rpc_done(rpc, ev->ev_status);
1091                 return 1;
1092         }
1093
1094         return 0;
1095 }
1096
1097 void
1098 srpc_client_rpc_expired (void *data)
1099 {
1100         srpc_client_rpc_t *rpc = data;
1101
1102         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1103                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1104                rpc->crpc_timeout);
1105
1106         cfs_spin_lock(&rpc->crpc_lock);
1107
1108         rpc->crpc_timeout = 0;
1109         srpc_abort_rpc(rpc, -ETIMEDOUT);
1110
1111         cfs_spin_unlock(&rpc->crpc_lock);
1112
1113         cfs_spin_lock(&srpc_data.rpc_glock);
1114         srpc_data.rpc_counters.rpcs_expired++;
1115         cfs_spin_unlock(&srpc_data.rpc_glock);
1116         return;
1117 }
1118
1119 inline void
1120 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
1121 {
1122         stt_timer_t *timer = &rpc->crpc_timer;
1123
1124         if (rpc->crpc_timeout == 0) return;
1125
1126         CFS_INIT_LIST_HEAD(&timer->stt_list);
1127         timer->stt_data    = rpc;
1128         timer->stt_func    = srpc_client_rpc_expired;
1129         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1130                                           cfs_time_current_sec());
1131         stt_add_timer(timer);
1132         return;
1133 }
1134
1135 /*
1136  * Called with rpc->crpc_lock held.
1137  *
1138  * Upon exit the RPC expiry timer is not queued and the handler is not
1139  * running on any CPU. */
1140 void
1141 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1142 {
1143         /* timer not planted or already exploded */
1144         if (rpc->crpc_timeout == 0) return;
1145
1146         /* timer sucessfully defused */
1147         if (stt_del_timer(&rpc->crpc_timer)) return;
1148
1149 #ifdef __KERNEL__
1150         /* timer detonated, wait for it to explode */
1151         while (rpc->crpc_timeout != 0) {
1152                 cfs_spin_unlock(&rpc->crpc_lock);
1153
1154                 cfs_schedule();
1155
1156                 cfs_spin_lock(&rpc->crpc_lock);
1157         }
1158 #else
1159         LBUG(); /* impossible in single-threaded runtime */
1160 #endif
1161         return;
1162 }
1163
1164 void
1165 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1166 {
1167         swi_workitem_t *wi = &rpc->crpc_wi;
1168
1169         LASSERT (status != 0 || wi->swi_state == SWI_STATE_DONE);
1170
1171         cfs_spin_lock(&rpc->crpc_lock);
1172
1173         rpc->crpc_closed = 1;
1174         if (rpc->crpc_status == 0)
1175                 rpc->crpc_status = status;
1176
1177         srpc_del_client_rpc_timer(rpc);
1178
1179         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1180                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1181                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1182                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1183
1184         /*
1185          * No one can schedule me now since:
1186          * - RPC timer has been defused.
1187          * - all LNet events have been fired.
1188          * - crpc_closed has been set, preventing srpc_abort_rpc from
1189          *   scheduling me.
1190          * Cancel pending schedules and prevent future schedule attempts:
1191          */
1192         LASSERT (!srpc_event_pending(rpc));
1193         swi_exit_workitem(wi);
1194
1195         cfs_spin_unlock(&rpc->crpc_lock);
1196
1197         (*rpc->crpc_done)(rpc);
1198         return;
1199 }
1200
1201 /* sends an outgoing RPC */
1202 int
1203 srpc_send_rpc (swi_workitem_t *wi)
1204 {
1205         int                rc = 0;
1206         srpc_client_rpc_t *rpc = wi->swi_workitem.wi_data;
1207         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1208         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1209
1210         LASSERT (rpc != NULL);
1211         LASSERT (wi == &rpc->crpc_wi);
1212
1213         cfs_spin_lock(&rpc->crpc_lock);
1214
1215         if (rpc->crpc_aborted) {
1216                 cfs_spin_unlock(&rpc->crpc_lock);
1217                 goto abort;
1218         }
1219
1220         cfs_spin_unlock(&rpc->crpc_lock);
1221
1222         switch (wi->swi_state) {
1223         default:
1224                 LBUG ();
1225         case SWI_STATE_NEWBORN:
1226                 LASSERT (!srpc_event_pending(rpc));
1227
1228                 rc = srpc_prepare_reply(rpc);
1229                 if (rc != 0) {
1230                         srpc_client_rpc_done(rpc, rc);
1231                         return 1;
1232                 }
1233
1234                 rc = srpc_prepare_bulk(rpc);
1235                 if (rc != 0) break;
1236
1237                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1238                 rc = srpc_send_request(rpc);
1239                 break;
1240
1241         case SWI_STATE_REQUEST_SUBMITTED:
1242                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1243                  * order; however, they're processed in a strict order:
1244                  * rqt, rpy, and bulk. */
1245                 if (!rpc->crpc_reqstev.ev_fired) break;
1246
1247                 rc = rpc->crpc_reqstev.ev_status;
1248                 if (rc != 0) break;
1249
1250                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1251                 /* perhaps more events, fall thru */
1252         case SWI_STATE_REQUEST_SENT: {
1253                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1254
1255                 if (!rpc->crpc_replyev.ev_fired) break;
1256
1257                 rc = rpc->crpc_replyev.ev_status;
1258                 if (rc != 0) break;
1259
1260                 if ((reply->msg_type != type &&
1261                      reply->msg_type != __swab32(type)) ||
1262                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1263                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1264                         CWARN ("Bad message from %s: type %u (%d expected),"
1265                                " magic %u (%d expected).\n",
1266                                libcfs_id2str(rpc->crpc_dest),
1267                                reply->msg_type, type,
1268                                reply->msg_magic, SRPC_MSG_MAGIC);
1269                         rc = -EBADMSG;
1270                         break;
1271                 }
1272
1273                 if (do_bulk && reply->msg_body.reply.status != 0) {
1274                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1275                                "case peer didn't initiate bulk transfer\n",
1276                                reply->msg_body.reply.status,
1277                                libcfs_id2str(rpc->crpc_dest));
1278                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1279                 }
1280
1281                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1282         }
1283         case SWI_STATE_REPLY_RECEIVED:
1284                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1285
1286                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1287
1288                 /* Bulk buffer was unlinked due to remote error. Clear error
1289                  * since reply buffer still contains valid data.
1290                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1291                  * remote error. */
1292                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1293                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1294                         rc = 0;
1295
1296                 wi->swi_state = SWI_STATE_DONE;
1297                 srpc_client_rpc_done(rpc, rc);
1298                 return 1;
1299         }
1300
1301         if (rc != 0) {
1302                 cfs_spin_lock(&rpc->crpc_lock);
1303                 srpc_abort_rpc(rpc, rc);
1304                 cfs_spin_unlock(&rpc->crpc_lock);
1305         }
1306
1307 abort:
1308         if (rpc->crpc_aborted) {
1309                 LNetMDUnlink(rpc->crpc_reqstmdh);
1310                 LNetMDUnlink(rpc->crpc_replymdh);
1311                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1312
1313                 if (!srpc_event_pending(rpc)) {
1314                         srpc_client_rpc_done(rpc, -EINTR);
1315                         return 1;
1316                 }
1317         }
1318         return 0;
1319 }
1320
1321 srpc_client_rpc_t *
1322 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1323                         int nbulkiov, int bulklen,
1324                         void (*rpc_done)(srpc_client_rpc_t *),
1325                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1326 {
1327         srpc_client_rpc_t *rpc;
1328
1329         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1330                                    crpc_bulk.bk_iovs[nbulkiov]));
1331         if (rpc == NULL)
1332                 return NULL;
1333
1334         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1335                              bulklen, rpc_done, rpc_fini, priv);
1336         return rpc;
1337 }
1338
1339 /* called with rpc->crpc_lock held */
1340 void
1341 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1342 {
1343         LASSERT (why != 0);
1344
1345         if (rpc->crpc_aborted || /* already aborted */
1346             rpc->crpc_closed)    /* callback imminent */
1347                 return;
1348
1349         CDEBUG (D_NET,
1350                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1351                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1352                 swi_state2str(rpc->crpc_wi.swi_state), why);
1353
1354         rpc->crpc_aborted = 1;
1355         rpc->crpc_status  = why;
1356         swi_schedule_workitem(&rpc->crpc_wi);
1357         return;
1358 }
1359
1360 /* called with rpc->crpc_lock held */
1361 void
1362 srpc_post_rpc (srpc_client_rpc_t *rpc)
1363 {
1364         LASSERT (!rpc->crpc_aborted);
1365         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1366         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1367
1368         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1369                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1370                 rpc->crpc_timeout);
1371
1372         srpc_add_client_rpc_timer(rpc);
1373         swi_schedule_workitem(&rpc->crpc_wi);
1374         return;
1375 }
1376
1377
1378 int
1379 srpc_send_reply(struct srpc_server_rpc *rpc)
1380 {
1381         srpc_event_t            *ev = &rpc->srpc_ev;
1382         struct srpc_msg         *msg = &rpc->srpc_replymsg;
1383         struct srpc_buffer      *buffer = rpc->srpc_reqstbuf;
1384         struct srpc_service_cd  *scd = rpc->srpc_scd;
1385         struct srpc_service     *sv = scd->scd_svc;
1386         __u64                   rpyid;
1387         int                     rc;
1388
1389         LASSERT(buffer != NULL);
1390         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1391
1392         cfs_spin_lock(&scd->scd_lock);
1393
1394         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1395                 /* Repost buffer before replying since test client
1396                  * might send me another RPC once it gets the reply */
1397                 if (srpc_service_post_buffer(scd, buffer) != 0)
1398                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1399                 rpc->srpc_reqstbuf = NULL;
1400         }
1401
1402         cfs_spin_unlock(&scd->scd_lock);
1403
1404         ev->ev_fired = 0;
1405         ev->ev_data  = rpc;
1406         ev->ev_type  = SRPC_REPLY_SENT;
1407
1408         msg->msg_magic   = SRPC_MSG_MAGIC;
1409         msg->msg_version = SRPC_MSG_VERSION;
1410         msg->msg_type    = srpc_service2reply(sv->sv_id);
1411
1412         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1413                                    sizeof(*msg), LNET_MD_OP_PUT,
1414                                    rpc->srpc_peer, rpc->srpc_self,
1415                                    &rpc->srpc_replymdh, ev);
1416         if (rc != 0)
1417                 ev->ev_fired = 1;  /* no more event expected */
1418         return rc;
1419 }
1420
1421 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1422 void
1423 srpc_lnet_ev_handler(lnet_event_t *ev)
1424 {
1425         struct srpc_service_cd  *scd;
1426         srpc_event_t      *rpcev = ev->md.user_ptr;
1427         srpc_client_rpc_t *crpc;
1428         srpc_server_rpc_t *srpc;
1429         srpc_buffer_t     *buffer;
1430         srpc_service_t    *sv;
1431         srpc_msg_t        *msg;
1432         srpc_msg_type_t    type;
1433
1434         LASSERT (!cfs_in_interrupt());
1435
1436         if (ev->status != 0) {
1437                 cfs_spin_lock(&srpc_data.rpc_glock);
1438                 srpc_data.rpc_counters.errors++;
1439                 cfs_spin_unlock(&srpc_data.rpc_glock);
1440         }
1441
1442         rpcev->ev_lnet = ev->type;
1443
1444         switch (rpcev->ev_type) {
1445         default:
1446                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1447                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1448                 LBUG ();
1449         case SRPC_REQUEST_SENT:
1450                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1451                         cfs_spin_lock(&srpc_data.rpc_glock);
1452                         srpc_data.rpc_counters.rpcs_sent++;
1453                         cfs_spin_unlock(&srpc_data.rpc_glock);
1454                 }
1455         case SRPC_REPLY_RCVD:
1456         case SRPC_BULK_REQ_RCVD:
1457                 crpc = rpcev->ev_data;
1458
1459                 if (rpcev != &crpc->crpc_reqstev &&
1460                     rpcev != &crpc->crpc_replyev &&
1461                     rpcev != &crpc->crpc_bulkev) {
1462                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1463                                rpcev, crpc, &crpc->crpc_reqstev,
1464                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1465                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1466                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1467                         LBUG ();
1468                 }
1469
1470                 cfs_spin_lock(&crpc->crpc_lock);
1471
1472                 LASSERT (rpcev->ev_fired == 0);
1473                 rpcev->ev_fired  = 1;
1474                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1475                                                 -EINTR : ev->status;
1476                 swi_schedule_workitem(&crpc->crpc_wi);
1477
1478                 cfs_spin_unlock(&crpc->crpc_lock);
1479                 break;
1480
1481         case SRPC_REQUEST_RCVD:
1482                 scd = rpcev->ev_data;
1483                 sv = scd->scd_svc;
1484
1485                 LASSERT(rpcev == &scd->scd_ev);
1486
1487                 cfs_spin_lock(&scd->scd_lock);
1488
1489                 LASSERT (ev->unlinked);
1490                 LASSERT (ev->type == LNET_EVENT_PUT ||
1491                          ev->type == LNET_EVENT_UNLINK);
1492                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1493                          sv->sv_shuttingdown);
1494
1495                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1496                 buffer->buf_peer = ev->initiator;
1497                 buffer->buf_self = ev->target.nid;
1498
1499                 LASSERT(scd->scd_buf_nposted > 0);
1500                 scd->scd_buf_nposted--;
1501
1502                 if (sv->sv_shuttingdown) {
1503                         /* Leave buffer on scd->scd_buf_nposted since
1504                          * srpc_finish_service needs to traverse it. */
1505                         cfs_spin_unlock(&scd->scd_lock);
1506                         break;
1507                 }
1508
1509                 if (scd->scd_buf_err_stamp != 0 &&
1510                     scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1511                         /* re-enable adding buffer */
1512                         scd->scd_buf_err_stamp = 0;
1513                         scd->scd_buf_err = 0;
1514                 }
1515
1516                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1517                     scd->scd_buf_adjust == 0 &&
1518                     scd->scd_buf_nposted < scd->scd_buf_low) {
1519                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1520                                                   SFW_TEST_WI_MIN);
1521                         swi_schedule_workitem(&scd->scd_buf_wi);
1522                 }
1523
1524                 cfs_list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1525                 msg = &buffer->buf_msg;
1526                 type = srpc_service2request(sv->sv_id);
1527
1528                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1529                     (msg->msg_type != type &&
1530                      msg->msg_type != __swab32(type)) ||
1531                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1532                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1533                         CERROR ("Dropping RPC (%s) from %s: "
1534                                 "status %d mlength %d type %u magic %u.\n",
1535                                 sv->sv_name, libcfs_id2str(ev->initiator),
1536                                 ev->status, ev->mlength,
1537                                 msg->msg_type, msg->msg_magic);
1538
1539                         /* NB can't call srpc_service_recycle_buffer here since
1540                          * it may call LNetM[DE]Attach. The invalid magic tells
1541                          * srpc_handle_rpc to drop this RPC */
1542                         msg->msg_magic = 0;
1543                 }
1544
1545                 if (!cfs_list_empty(&scd->scd_rpc_free)) {
1546                         srpc = cfs_list_entry(scd->scd_rpc_free.next,
1547                                               struct srpc_server_rpc,
1548                                               srpc_list);
1549                         cfs_list_del(&srpc->srpc_list);
1550
1551                         srpc_init_server_rpc(srpc, scd, buffer);
1552                         cfs_list_add_tail(&srpc->srpc_list,
1553                                           &scd->scd_rpc_active);
1554                         swi_schedule_workitem(&srpc->srpc_wi);
1555                 } else {
1556                         cfs_list_add_tail(&buffer->buf_list,
1557                                           &scd->scd_buf_blocked);
1558                 }
1559
1560                 cfs_spin_unlock(&scd->scd_lock);
1561
1562                 cfs_spin_lock(&srpc_data.rpc_glock);
1563                 srpc_data.rpc_counters.rpcs_rcvd++;
1564                 cfs_spin_unlock(&srpc_data.rpc_glock);
1565                 break;
1566
1567         case SRPC_BULK_GET_RPLD:
1568                 LASSERT (ev->type == LNET_EVENT_SEND ||
1569                          ev->type == LNET_EVENT_REPLY ||
1570                          ev->type == LNET_EVENT_UNLINK);
1571
1572                 if (!ev->unlinked)
1573                         break; /* wait for final event */
1574
1575         case SRPC_BULK_PUT_SENT:
1576                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1577                         cfs_spin_lock(&srpc_data.rpc_glock);
1578
1579                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1580                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1581                         else
1582                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1583
1584                         cfs_spin_unlock(&srpc_data.rpc_glock);
1585                 }
1586         case SRPC_REPLY_SENT:
1587                 srpc = rpcev->ev_data;
1588                 scd  = srpc->srpc_scd;
1589
1590                 LASSERT(rpcev == &srpc->srpc_ev);
1591
1592                 cfs_spin_lock(&scd->scd_lock);
1593
1594                 rpcev->ev_fired  = 1;
1595                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1596                                    -EINTR : ev->status;
1597                 swi_schedule_workitem(&srpc->srpc_wi);
1598
1599                 cfs_spin_unlock(&scd->scd_lock);
1600                 break;
1601         }
1602 }
1603
1604 #ifndef __KERNEL__
1605
1606 int
1607 srpc_check_event (int timeout)
1608 {
1609         lnet_event_t ev;
1610         int          rc;
1611         int          i;
1612
1613         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1614                         timeout * 1000, &ev, &i);
1615         if (rc == 0) return 0;
1616
1617         LASSERT (rc == -EOVERFLOW || rc == 1);
1618
1619         /* We can't affort to miss any events... */
1620         if (rc == -EOVERFLOW) {
1621                 CERROR ("Dropped an event!!!\n");
1622                 abort();
1623         }
1624
1625         srpc_lnet_ev_handler(&ev);
1626         return 1;
1627 }
1628
1629 #endif
1630
1631 int
1632 srpc_startup (void)
1633 {
1634         int rc;
1635
1636         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1637         cfs_spin_lock_init(&srpc_data.rpc_glock);
1638
1639         /* 1 second pause to avoid timestamp reuse */
1640         cfs_pause(cfs_time_seconds(1));
1641         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1642
1643         srpc_data.rpc_state = SRPC_STATE_NONE;
1644
1645 #ifdef __KERNEL__
1646         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1647 #else
1648         if (the_lnet.ln_server_mode_flag)
1649                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1650         else
1651                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1652 #endif
1653         if (rc < 0) {
1654                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1655                 return rc;
1656         }
1657
1658         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1659
1660         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1661 #ifdef __KERNEL__
1662         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1663 #else
1664         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1665 #endif
1666         if (rc != 0) {
1667                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1668                 goto bail;
1669         }
1670
1671         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1672         LASSERT(rc == 0);
1673         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1674         LASSERT(rc == 0);
1675
1676         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1677
1678         rc = stt_startup();
1679
1680 bail:
1681         if (rc != 0)
1682                 srpc_shutdown();
1683         else
1684                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1685
1686         return rc;
1687 }
1688
1689 void
1690 srpc_shutdown (void)
1691 {
1692         int i;
1693         int rc;
1694         int state;
1695
1696         state = srpc_data.rpc_state;
1697         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1698
1699         switch (state) {
1700         default:
1701                 LBUG ();
1702         case SRPC_STATE_RUNNING:
1703                 cfs_spin_lock(&srpc_data.rpc_glock);
1704
1705                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1706                         srpc_service_t *sv = srpc_data.rpc_services[i];
1707
1708                         LASSERTF (sv == NULL,
1709                                   "service not empty: id %d, name %s\n",
1710                                   i, sv->sv_name);
1711                 }
1712
1713                 cfs_spin_unlock(&srpc_data.rpc_glock);
1714
1715                 stt_shutdown();
1716
1717         case SRPC_STATE_EQ_INIT:
1718                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1719                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1720                 LASSERT (rc == 0);
1721                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1722                 LASSERT (rc == 0); /* the EQ should have no user by now */
1723
1724         case SRPC_STATE_NI_INIT:
1725                 LNetNIFini();
1726         }
1727
1728         return;
1729 }