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, 2014, 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 #ifdef __KERNEL__
  99         bk->bk_iovs[i].kiov_offset = 0;
 100         bk->bk_iovs[i].kiov_page   = pg;
 101         bk->bk_iovs[i].kiov_len    = nob;
 102 #else
 103         LASSERT(bk->bk_pages != NULL);
 104
 105         bk->bk_pages[i] = pg;
 106         bk->bk_iovs[i].iov_len  = nob;
 107         bk->bk_iovs[i].iov_base = page_address(pg);
 108 #endif
 109         return nob;
 110 }
 111
 112 void
 113 srpc_free_bulk (srpc_bulk_t *bk)
 114 {
 115         int         i;
 116         struct page *pg;
 117
 118         LASSERT (bk != NULL);
 119 #ifndef __KERNEL__
 120         LASSERT (bk->bk_pages != NULL);
 121 #endif
 122
 123         for (i = 0; i < bk->bk_niov; i++) {
 124 #ifdef __KERNEL__
 125                 pg = bk->bk_iovs[i].kiov_page;
 126 #else
 127                 pg = bk->bk_pages[i];
 128 #endif
 129                 if (pg == NULL) break;
 130
 131                 __free_page(pg);
 132         }
 133
 134 #ifndef __KERNEL__
 135         LIBCFS_FREE(bk->bk_pages, sizeof(struct page *) * bk->bk_niov);
 136 #endif
 137         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 138         return;
 139 }
 140
 141 srpc_bulk_t *
 142 srpc_alloc_bulk(int cpt, unsigned bulk_npg, unsigned bulk_len, int sink)
 143 {
 144         srpc_bulk_t  *bk;
 145         int           i;
 146
 147         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 148
 149         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 150                          offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 151         if (bk == NULL) {
 152                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 153                 return NULL;
 154         }
 155
 156         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 157         bk->bk_sink   = sink;
 158         bk->bk_len    = bulk_len;
 159         bk->bk_niov   = bulk_npg;
 160 #ifndef __KERNEL__
 161         {
 162                 struct page  **pages;
 163
 164                 LIBCFS_CPT_ALLOC(pages, lnet_cpt_table(), cpt,
 165                                  sizeof(struct page *) * bulk_npg);
 166                 if (pages == NULL) {
 167                         LIBCFS_FREE(bk, offsetof(srpc_bulk_t,
 168                                     bk_iovs[bulk_npg]));
 169                         CERROR("Can't allocate page array for %d pages\n",
 170                                 bulk_npg);
 171                         return NULL;
 172                 }
 173
 174                 memset(pages, 0, sizeof(struct page *) * bulk_npg);
 175                 bk->bk_pages = pages;
 176         }
 177 #endif
 178
 179         for (i = 0; i < bulk_npg; i++) {
 180                 struct page *pg;
 181                 int         nob;
 182
 183                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_IOFS);
 184                 if (pg == NULL) {
 185                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 186                         srpc_free_bulk(bk);
 187                         return NULL;
 188                 }
 189
 190                 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
 191                 bulk_len -= nob;
 192         }
 193
 194         return bk;
 195 }
 196
 197 static inline __u64
 198 srpc_next_id (void)
 199 {
 200         __u64 id;
 201
 202         spin_lock(&srpc_data.rpc_glock);
 203         id = srpc_data.rpc_matchbits++;
 204         spin_unlock(&srpc_data.rpc_glock);
 205         return id;
 206 }
 207
 208 static void
 209 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 210                      struct srpc_service_cd *scd,
 211                      struct srpc_buffer *buffer)
 212 {
 213         memset(rpc, 0, sizeof(*rpc));
 214         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
 215                           srpc_serv_is_framework(scd->scd_svc) ?
 216                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 217
 218         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 219
 220         rpc->srpc_scd      = scd;
 221         rpc->srpc_reqstbuf = buffer;
 222         rpc->srpc_peer     = buffer->buf_peer;
 223         rpc->srpc_self     = buffer->buf_self;
 224         LNetInvalidateHandle(&rpc->srpc_replymdh);
 225 }
 226
 227 static void
 228 srpc_service_fini(struct srpc_service *svc)
 229 {
 230         struct srpc_service_cd  *scd;
 231         struct srpc_server_rpc  *rpc;
 232         struct srpc_buffer      *buf;
 233         struct list_head                *q;
 234         int                     i;
 235
 236         if (svc->sv_cpt_data == NULL)
 237                 return;
 238
 239         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 240                 while (1) {
 241                         if (!list_empty(&scd->scd_buf_posted))
 242                                 q = &scd->scd_buf_posted;
 243                         else if (!list_empty(&scd->scd_buf_blocked))
 244                                 q = &scd->scd_buf_blocked;
 245                         else
 246                                 break;
 247
 248                         while (!list_empty(q)) {
 249                                 buf = list_entry(q->next,
 250                                                      struct srpc_buffer,
 251                                                      buf_list);
 252                                 list_del(&buf->buf_list);
 253                                 LIBCFS_FREE(buf, sizeof(*buf));
 254                         }
 255                 }
 256
 257                 LASSERT(list_empty(&scd->scd_rpc_active));
 258
 259                 while (!list_empty(&scd->scd_rpc_free)) {
 260                         rpc = list_entry(scd->scd_rpc_free.next,
 261                                              struct srpc_server_rpc,
 262                                              srpc_list);
 263                         list_del(&rpc->srpc_list);
 264                         LIBCFS_FREE(rpc, sizeof(*rpc));
 265                 }
 266         }
 267
 268         cfs_percpt_free(svc->sv_cpt_data);
 269         svc->sv_cpt_data = NULL;
 270 }
 271
 272 static int
 273 srpc_service_nrpcs(struct srpc_service *svc)
 274 {
 275         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 276
 277         return srpc_serv_is_framework(svc) ?
 278                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 279 }
 280
 281 int srpc_add_buffer(struct swi_workitem *wi);
 282
 283 static int
 284 srpc_service_init(struct srpc_service *svc)
 285 {
 286         struct srpc_service_cd  *scd;
 287         struct srpc_server_rpc  *rpc;
 288         int                     nrpcs;
 289         int                     i;
 290         int                     j;
 291
 292         svc->sv_shuttingdown = 0;
 293
 294         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 295                                             sizeof(struct srpc_service_cd));
 296         if (svc->sv_cpt_data == NULL)
 297                 return -ENOMEM;
 298
 299         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 300                         1 : cfs_cpt_number(lnet_cpt_table());
 301         nrpcs = srpc_service_nrpcs(svc);
 302
 303         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 304                 scd->scd_cpt = i;
 305                 scd->scd_svc = svc;
 306                 spin_lock_init(&scd->scd_lock);
 307                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 308                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 309                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 310                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 311
 312                 scd->scd_ev.ev_data = scd;
 313                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 314
 315                 /* NB: don't use lst_sched_serial for adding buffer,
 316                  * see details in srpc_service_add_buffers() */
 317                 swi_init_workitem(&scd->scd_buf_wi, scd,
 318                                   srpc_add_buffer, lst_sched_test[i]);
 319
 320                 if (i != 0 && srpc_serv_is_framework(svc)) {
 321                         /* NB: framework service only needs srpc_service_cd for
 322                          * one partition, but we allocate for all to make
 323                          * it easier to implement, it will waste a little
 324                          * memory but nobody should care about this */
 325                         continue;
 326                 }
 327
 328                 for (j = 0; j < nrpcs; j++) {
 329                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 330                                          i, sizeof(*rpc));
 331                         if (rpc == NULL) {
 332                                 srpc_service_fini(svc);
 333                                 return -ENOMEM;
 334                         }
 335                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 336                 }
 337         }
 338
 339         return 0;
 340 }
 341
 342 int
 343 srpc_add_service(struct srpc_service *sv)
 344 {
 345         int id = sv->sv_id;
 346
 347         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 348
 349         if (srpc_service_init(sv) != 0)
 350                 return -ENOMEM;
 351
 352         spin_lock(&srpc_data.rpc_glock);
 353
 354         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 355
 356         if (srpc_data.rpc_services[id] != NULL) {
 357                 spin_unlock(&srpc_data.rpc_glock);
 358                 goto failed;
 359         }
 360
 361         srpc_data.rpc_services[id] = sv;
 362         spin_unlock(&srpc_data.rpc_glock);
 363
 364         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 365         return 0;
 366
 367  failed:
 368         srpc_service_fini(sv);
 369         return -EBUSY;
 370 }
 371
 372 int
 373 srpc_remove_service (srpc_service_t *sv)
 374 {
 375         int id = sv->sv_id;
 376
 377         spin_lock(&srpc_data.rpc_glock);
 378
 379         if (srpc_data.rpc_services[id] != sv) {
 380                 spin_unlock(&srpc_data.rpc_glock);
 381                 return -ENOENT;
 382         }
 383
 384         srpc_data.rpc_services[id] = NULL;
 385         spin_unlock(&srpc_data.rpc_glock);
 386         return 0;
 387 }
 388
 389 static int
 390 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 391                        int len, int options, lnet_process_id_t peer,
 392                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 393 {
 394         int              rc;
 395         lnet_md_t        md;
 396         lnet_handle_me_t meh;
 397
 398         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 399                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 400         if (rc != 0) {
 401                 CERROR ("LNetMEAttach failed: %d\n", rc);
 402                 LASSERT (rc == -ENOMEM);
 403                 return -ENOMEM;
 404         }
 405
 406         md.threshold = 1;
 407         md.user_ptr  = ev;
 408         md.start     = buf;
 409         md.length    = len;
 410         md.options   = options;
 411         md.eq_handle = srpc_data.rpc_lnet_eq;
 412
 413         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 414         if (rc != 0) {
 415                 CERROR ("LNetMDAttach failed: %d\n", rc);
 416                 LASSERT (rc == -ENOMEM);
 417
 418                 rc = LNetMEUnlink(meh);
 419                 LASSERT (rc == 0);
 420                 return -ENOMEM;
 421         }
 422
 423         CDEBUG (D_NET,
 424                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 425                 libcfs_id2str(peer), portal, matchbits);
 426         return 0;
 427 }
 428
 429 static int
 430 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 431                       int options, lnet_process_id_t peer, lnet_nid_t self,
 432                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 433 {
 434         int       rc;
 435         lnet_md_t md;
 436
 437         md.user_ptr  = ev;
 438         md.start     = buf;
 439         md.length    = len;
 440         md.eq_handle = srpc_data.rpc_lnet_eq;
 441         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 442         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 443
 444         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 445         if (rc != 0) {
 446                 CERROR ("LNetMDBind failed: %d\n", rc);
 447                 LASSERT (rc == -ENOMEM);
 448                 return -ENOMEM;
 449         }
 450
 451         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 452          * they're only meaningful for MDs attached to an ME (i.e. passive
 453          * buffers... */
 454         if ((options & LNET_MD_OP_PUT) != 0) {
 455                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 456                              portal, matchbits, 0, 0);
 457         } else {
 458                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 459
 460                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 461         }
 462
 463         if (rc != 0) {
 464                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 465                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 466                         libcfs_id2str(peer), portal, matchbits, rc);
 467
 468                 /* The forthcoming unlink event will complete this operation
 469                  * with failure, so fall through and return success here.
 470                  */
 471                 rc = LNetMDUnlink(*mdh);
 472                 LASSERT (rc == 0);
 473         } else {
 474                 CDEBUG (D_NET,
 475                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 476                         libcfs_id2str(peer), portal, matchbits);
 477         }
 478         return 0;
 479 }
 480
 481 static int
 482 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 483                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 484 {
 485         return srpc_post_active_rdma(srpc_serv_portal(service), service,
 486                                      buf, len, LNET_MD_OP_PUT, peer,
 487                                      LNET_NID_ANY, mdh, ev);
 488 }
 489
 490 static int
 491 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 492                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 493 {
 494         lnet_process_id_t any = {0};
 495
 496         any.nid = LNET_NID_ANY;
 497         any.pid = LNET_PID_ANY;
 498
 499         return srpc_post_passive_rdma(srpc_serv_portal(service),
 500                                       local, service, buf, len,
 501                                       LNET_MD_OP_PUT, any, mdh, ev);
 502 }
 503
 504 static int
 505 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 506 __must_hold(&scd->scd_lock)
 507 {
 508         struct srpc_service     *sv = scd->scd_svc;
 509         struct srpc_msg         *msg = &buf->buf_msg;
 510         int                     rc;
 511
 512         LNetInvalidateHandle(&buf->buf_mdh);
 513         list_add(&buf->buf_list, &scd->scd_buf_posted);
 514         scd->scd_buf_nposted++;
 515         spin_unlock(&scd->scd_lock);
 516
 517         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 518                                       !srpc_serv_is_framework(sv),
 519                                       msg, sizeof(*msg), &buf->buf_mdh,
 520                                       &scd->scd_ev);
 521
 522         /* At this point, a RPC (new or delayed) may have arrived in
 523          * msg and its event handler has been called. So we must add
 524          * buf to scd_buf_posted _before_ dropping scd_lock */
 525
 526         spin_lock(&scd->scd_lock);
 527
 528         if (rc == 0) {
 529                 if (!sv->sv_shuttingdown)
 530                         return 0;
 531
 532                 spin_unlock(&scd->scd_lock);
 533                 /* srpc_shutdown_service might have tried to unlink me
 534                  * when my buf_mdh was still invalid */
 535                 LNetMDUnlink(buf->buf_mdh);
 536                 spin_lock(&scd->scd_lock);
 537                 return 0;
 538         }
 539
 540         scd->scd_buf_nposted--;
 541         if (sv->sv_shuttingdown)
 542                 return rc; /* don't allow to change scd_buf_posted */
 543
 544         list_del(&buf->buf_list);
 545         spin_unlock(&scd->scd_lock);
 546
 547         LIBCFS_FREE(buf, sizeof(*buf));
 548
 549         spin_lock(&scd->scd_lock);
 550         return rc;
 551 }
 552
 553 int
 554 srpc_add_buffer(struct swi_workitem *wi)
 555 {
 556         struct srpc_service_cd  *scd = wi->swi_workitem.wi_data;
 557         struct srpc_buffer      *buf;
 558         int                     rc = 0;
 559
 560         /* it's called by workitem scheduler threads, these threads
 561          * should have been set CPT affinity, so buffers will be posted
 562          * on CPT local list of Portal */
 563         spin_lock(&scd->scd_lock);
 564
 565         while (scd->scd_buf_adjust > 0 &&
 566                !scd->scd_svc->sv_shuttingdown) {
 567                 scd->scd_buf_adjust--; /* consume it */
 568                 scd->scd_buf_posting++;
 569
 570                 spin_unlock(&scd->scd_lock);
 571
 572                 LIBCFS_ALLOC(buf, sizeof(*buf));
 573                 if (buf == NULL) {
 574                         CERROR("Failed to add new buf to service: %s\n",
 575                                scd->scd_svc->sv_name);
 576                         spin_lock(&scd->scd_lock);
 577                         rc = -ENOMEM;
 578                         break;
 579                 }
 580
 581                 spin_lock(&scd->scd_lock);
 582                 if (scd->scd_svc->sv_shuttingdown) {
 583                         spin_unlock(&scd->scd_lock);
 584                         LIBCFS_FREE(buf, sizeof(*buf));
 585
 586                         spin_lock(&scd->scd_lock);
 587                         rc = -ESHUTDOWN;
 588                         break;
 589                 }
 590
 591                 rc = srpc_service_post_buffer(scd, buf);
 592                 if (rc != 0)
 593                         break; /* buf has been freed inside */
 594
 595                 LASSERT(scd->scd_buf_posting > 0);
 596                 scd->scd_buf_posting--;
 597                 scd->scd_buf_total++;
 598                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 599         }
 600
 601         if (rc != 0) {
 602                 scd->scd_buf_err_stamp = cfs_time_current_sec();
 603                 scd->scd_buf_err = rc;
 604
 605                 LASSERT(scd->scd_buf_posting > 0);
 606                 scd->scd_buf_posting--;
 607         }
 608
 609         spin_unlock(&scd->scd_lock);
 610         return 0;
 611 }
 612
 613 int
 614 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 615 {
 616         struct srpc_service_cd  *scd;
 617         int                     rc = 0;
 618         int                     i;
 619
 620         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 621
 622         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 623                 spin_lock(&scd->scd_lock);
 624
 625                 scd->scd_buf_err = 0;
 626                 scd->scd_buf_err_stamp = 0;
 627                 scd->scd_buf_posting = 0;
 628                 scd->scd_buf_adjust = nbuffer;
 629                 /* start to post buffers */
 630                 swi_schedule_workitem(&scd->scd_buf_wi);
 631                 spin_unlock(&scd->scd_lock);
 632
 633                 /* framework service only post buffer for one partition  */
 634                 if (srpc_serv_is_framework(sv))
 635                         break;
 636         }
 637
 638         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 639                 spin_lock(&scd->scd_lock);
 640                 /*
 641                  * NB: srpc_service_add_buffers() can be called inside
 642                  * thread context of lst_sched_serial, and we don't normally
 643                  * allow to sleep inside thread context of WI scheduler
 644                  * because it will block current scheduler thread from doing
 645                  * anything else, even worse, it could deadlock if it's
 646                  * waiting on result from another WI of the same scheduler.
 647                  * However, it's safe at here because scd_buf_wi is scheduled
 648                  * by thread in a different WI scheduler (lst_sched_test),
 649                  * so we don't have any risk of deadlock, though this could
 650                  * block all WIs pending on lst_sched_serial for a moment
 651                  * which is not good but not fatal.
 652                  */
 653                 lst_wait_until(scd->scd_buf_err != 0 ||
 654                                (scd->scd_buf_adjust == 0 &&
 655                                 scd->scd_buf_posting == 0),
 656                                scd->scd_lock, "waiting for adding buffer\n");
 657
 658                 if (scd->scd_buf_err != 0 && rc == 0)
 659                         rc = scd->scd_buf_err;
 660
 661                 spin_unlock(&scd->scd_lock);
 662         }
 663
 664         return rc;
 665 }
 666
 667 void
 668 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 669 {
 670         struct srpc_service_cd  *scd;
 671         int                     num;
 672         int                     i;
 673
 674         LASSERT(!sv->sv_shuttingdown);
 675
 676         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 677                 spin_lock(&scd->scd_lock);
 678
 679                 num = scd->scd_buf_total + scd->scd_buf_posting;
 680                 scd->scd_buf_adjust -= min(nbuffer, num);
 681
 682                 spin_unlock(&scd->scd_lock);
 683         }
 684 }
 685
 686 /* returns 1 if sv has finished, otherwise 0 */
 687 int
 688 srpc_finish_service(struct srpc_service *sv)
 689 {
 690         struct srpc_service_cd  *scd;
 691         struct srpc_server_rpc  *rpc;
 692         int                     i;
 693
 694         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 695
 696         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 697                 spin_lock(&scd->scd_lock);
 698                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 699                         spin_unlock(&scd->scd_lock);
 700                         return 0;
 701                 }
 702
 703                 if (scd->scd_buf_nposted > 0) {
 704                         CDEBUG(D_NET, "waiting for %d posted buffers to "
 705                                "unlink\n", scd->scd_buf_nposted);
 706                         spin_unlock(&scd->scd_lock);
 707                         return 0;
 708                 }
 709
 710                 if (list_empty(&scd->scd_rpc_active)) {
 711                         spin_unlock(&scd->scd_lock);
 712                         continue;
 713                 }
 714
 715                 rpc = list_entry(scd->scd_rpc_active.next,
 716                                      struct srpc_server_rpc, srpc_list);
 717                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
 718                         "wi %s scheduled %d running %d, "
 719                         "ev fired %d type %d status %d lnet %d\n",
 720                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 721                         swi_state2str(rpc->srpc_wi.swi_state),
 722                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 723                         rpc->srpc_wi.swi_workitem.wi_running,
 724                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 725                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 726                 spin_unlock(&scd->scd_lock);
 727                 return 0;
 728         }
 729
 730         /* no lock needed from now on */
 731         srpc_service_fini(sv);
 732         return 1;
 733 }
 734
 735 /* called with sv->sv_lock held */
 736 static void
 737 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
 738 __must_hold(&scd->scd_lock)
 739 {
 740         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 741                 if (srpc_service_post_buffer(scd, buf) != 0) {
 742                         CWARN("Failed to post %s buffer\n",
 743                               scd->scd_svc->sv_name);
 744                 }
 745                 return;
 746         }
 747
 748         /* service is shutting down, or we want to recycle some buffers */
 749         scd->scd_buf_total--;
 750
 751         if (scd->scd_buf_adjust < 0) {
 752                 scd->scd_buf_adjust++;
 753                 if (scd->scd_buf_adjust < 0 &&
 754                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 755                         CDEBUG(D_INFO,
 756                                "Try to recyle %d buffers but nothing left\n",
 757                                scd->scd_buf_adjust);
 758                         scd->scd_buf_adjust = 0;
 759                 }
 760         }
 761
 762         spin_unlock(&scd->scd_lock);
 763         LIBCFS_FREE(buf, sizeof(*buf));
 764         spin_lock(&scd->scd_lock);
 765 }
 766
 767 void
 768 srpc_abort_service(struct srpc_service *sv)
 769 {
 770         struct srpc_service_cd  *scd;
 771         struct srpc_server_rpc  *rpc;
 772         int                     i;
 773
 774         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 775                sv->sv_id, sv->sv_name);
 776
 777         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 778                 spin_lock(&scd->scd_lock);
 779
 780                 /* schedule in-flight RPCs to notice the abort, NB:
 781                  * racing with incoming RPCs; complete fix should make test
 782                  * RPCs carry session ID in its headers */
 783                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 784                         rpc->srpc_aborted = 1;
 785                         swi_schedule_workitem(&rpc->srpc_wi);
 786                 }
 787
 788                 spin_unlock(&scd->scd_lock);
 789         }
 790 }
 791
 792 void
 793 srpc_shutdown_service(srpc_service_t *sv)
 794 {
 795         struct srpc_service_cd  *scd;
 796         struct srpc_server_rpc  *rpc;
 797         srpc_buffer_t           *buf;
 798         int                     i;
 799
 800         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 801                sv->sv_id, sv->sv_name);
 802
 803         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 804                 spin_lock(&scd->scd_lock);
 805
 806         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 807
 808         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 809                 spin_unlock(&scd->scd_lock);
 810
 811         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 812                 spin_lock(&scd->scd_lock);
 813
 814                 /* schedule in-flight RPCs to notice the shutdown */
 815                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 816                         swi_schedule_workitem(&rpc->srpc_wi);
 817
 818                 spin_unlock(&scd->scd_lock);
 819
 820                 /* OK to traverse scd_buf_posted without lock, since no one
 821                  * touches scd_buf_posted now */
 822                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 823                         LNetMDUnlink(buf->buf_mdh);
 824         }
 825 }
 826
 827 static int
 828 srpc_send_request (srpc_client_rpc_t *rpc)
 829 {
 830         srpc_event_t *ev = &rpc->crpc_reqstev;
 831         int           rc;
 832
 833         ev->ev_fired = 0;
 834         ev->ev_data  = rpc;
 835         ev->ev_type  = SRPC_REQUEST_SENT;
 836
 837         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 838                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 839                                      &rpc->crpc_reqstmdh, ev);
 840         if (rc != 0) {
 841                 LASSERT (rc == -ENOMEM);
 842                 ev->ev_fired = 1;  /* no more event expected */
 843         }
 844         return rc;
 845 }
 846
 847 static int
 848 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 849 {
 850         srpc_event_t *ev = &rpc->crpc_replyev;
 851         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 852         int           rc;
 853
 854         ev->ev_fired = 0;
 855         ev->ev_data  = rpc;
 856         ev->ev_type  = SRPC_REPLY_RCVD;
 857
 858         *id = srpc_next_id();
 859
 860         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 861                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 862                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 863                                     &rpc->crpc_replymdh, ev);
 864         if (rc != 0) {
 865                 LASSERT (rc == -ENOMEM);
 866                 ev->ev_fired = 1;  /* no more event expected */
 867         }
 868         return rc;
 869 }
 870
 871 static int
 872 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 873 {
 874         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 875         srpc_event_t *ev = &rpc->crpc_bulkev;
 876         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 877         int           rc;
 878         int           opt;
 879
 880         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 881
 882         if (bk->bk_niov == 0) return 0; /* nothing to do */
 883
 884         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 885 #ifdef __KERNEL__
 886         opt |= LNET_MD_KIOV;
 887 #else
 888         opt |= LNET_MD_IOVEC;
 889 #endif
 890
 891         ev->ev_fired = 0;
 892         ev->ev_data  = rpc;
 893         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 894
 895         *id = srpc_next_id();
 896
 897         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 898                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 899                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 900         if (rc != 0) {
 901                 LASSERT (rc == -ENOMEM);
 902                 ev->ev_fired = 1;  /* no more event expected */
 903         }
 904         return rc;
 905 }
 906
 907 static int
 908 srpc_do_bulk (srpc_server_rpc_t *rpc)
 909 {
 910         srpc_event_t  *ev = &rpc->srpc_ev;
 911         srpc_bulk_t   *bk = rpc->srpc_bulk;
 912         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 913         int            rc;
 914         int            opt;
 915
 916         LASSERT (bk != NULL);
 917
 918         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 919 #ifdef __KERNEL__
 920         opt |= LNET_MD_KIOV;
 921 #else
 922         opt |= LNET_MD_IOVEC;
 923 #endif
 924
 925         ev->ev_fired = 0;
 926         ev->ev_data  = rpc;
 927         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 928
 929         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 930                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 931                                    rpc->srpc_peer, rpc->srpc_self,
 932                                    &bk->bk_mdh, ev);
 933         if (rc != 0)
 934                 ev->ev_fired = 1;  /* no more event expected */
 935         return rc;
 936 }
 937
 938 /* only called from srpc_handle_rpc */
 939 static void
 940 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
 941 {
 942         struct srpc_service_cd  *scd = rpc->srpc_scd;
 943         struct srpc_service     *sv  = scd->scd_svc;
 944         srpc_buffer_t           *buffer;
 945
 946         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 947
 948         rpc->srpc_status = status;
 949
 950         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 951                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 952                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 953                 swi_state2str(rpc->srpc_wi.swi_state), status);
 954
 955         if (status != 0) {
 956                 spin_lock(&srpc_data.rpc_glock);
 957                 srpc_data.rpc_counters.rpcs_dropped++;
 958                 spin_unlock(&srpc_data.rpc_glock);
 959         }
 960
 961         if (rpc->srpc_done != NULL)
 962                 (*rpc->srpc_done) (rpc);
 963         LASSERT(rpc->srpc_bulk == NULL);
 964
 965         spin_lock(&scd->scd_lock);
 966
 967         if (rpc->srpc_reqstbuf != NULL) {
 968                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 969                  * sv won't go away for scd_rpc_active must not be empty */
 970                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 971                 rpc->srpc_reqstbuf = NULL;
 972         }
 973
 974         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 975
 976         /*
 977          * No one can schedule me now since:
 978          * - I'm not on scd_rpc_active.
 979          * - all LNet events have been fired.
 980          * Cancel pending schedules and prevent future schedule attempts:
 981          */
 982         LASSERT(rpc->srpc_ev.ev_fired);
 983         swi_exit_workitem(&rpc->srpc_wi);
 984
 985         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 986                 buffer = list_entry(scd->scd_buf_blocked.next,
 987                                         srpc_buffer_t, buf_list);
 988                 list_del(&buffer->buf_list);
 989
 990                 srpc_init_server_rpc(rpc, scd, buffer);
 991                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 992                 swi_schedule_workitem(&rpc->srpc_wi);
 993         } else {
 994                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 995         }
 996
 997         spin_unlock(&scd->scd_lock);
 998         return;
 999 }
1000
1001 /* handles an incoming RPC */
1002 int
1003 srpc_handle_rpc(swi_workitem_t *wi)
1004 {
1005         struct srpc_server_rpc  *rpc = wi->swi_workitem.wi_data;
1006         struct srpc_service_cd  *scd = rpc->srpc_scd;
1007         struct srpc_service     *sv = scd->scd_svc;
1008         srpc_event_t            *ev = &rpc->srpc_ev;
1009         int                     rc = 0;
1010
1011         LASSERT(wi == &rpc->srpc_wi);
1012
1013         spin_lock(&scd->scd_lock);
1014
1015         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
1016                 spin_unlock(&scd->scd_lock);
1017
1018                 if (rpc->srpc_bulk != NULL)
1019                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
1020                 LNetMDUnlink(rpc->srpc_replymdh);
1021
1022                 if (ev->ev_fired) { /* no more event, OK to finish */
1023                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
1024                         return 1;
1025                 }
1026                 return 0;
1027         }
1028
1029         spin_unlock(&scd->scd_lock);
1030
1031         switch (wi->swi_state) {
1032         default:
1033                 LBUG ();
1034         case SWI_STATE_NEWBORN: {
1035                 srpc_msg_t           *msg;
1036                 srpc_generic_reply_t *reply;
1037
1038                 msg = &rpc->srpc_reqstbuf->buf_msg;
1039                 reply = &rpc->srpc_replymsg.msg_body.reply;
1040
1041                 if (msg->msg_magic == 0) {
1042                         /* moaned already in srpc_lnet_ev_handler */
1043                         srpc_server_rpc_done(rpc, EBADMSG);
1044                         return 1;
1045                 }
1046
1047                 srpc_unpack_msg_hdr(msg);
1048                 if (msg->msg_version != SRPC_MSG_VERSION) {
1049                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1050                               msg->msg_version, SRPC_MSG_VERSION,
1051                               libcfs_id2str(rpc->srpc_peer));
1052                         reply->status = EPROTO;
1053                         /* drop through and send reply */
1054                 } else {
1055                         reply->status = 0;
1056                         rc = (*sv->sv_handler)(rpc);
1057                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1058                         if (rc != 0) {
1059                                 srpc_server_rpc_done(rpc, rc);
1060                                 return 1;
1061                         }
1062                 }
1063
1064                 wi->swi_state = SWI_STATE_BULK_STARTED;
1065
1066                 if (rpc->srpc_bulk != NULL) {
1067                         rc = srpc_do_bulk(rpc);
1068                         if (rc == 0)
1069                                 return 0; /* wait for bulk */
1070
1071                         LASSERT (ev->ev_fired);
1072                         ev->ev_status = rc;
1073                 }
1074         }
1075         case SWI_STATE_BULK_STARTED:
1076                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1077
1078                 if (rpc->srpc_bulk != NULL) {
1079                         rc = ev->ev_status;
1080
1081                         if (sv->sv_bulk_ready != NULL)
1082                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1083
1084                         if (rc != 0) {
1085                                 srpc_server_rpc_done(rpc, rc);
1086                                 return 1;
1087                         }
1088                 }
1089
1090                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1091                 rc = srpc_send_reply(rpc);
1092                 if (rc == 0)
1093                         return 0; /* wait for reply */
1094                 srpc_server_rpc_done(rpc, rc);
1095                 return 1;
1096
1097         case SWI_STATE_REPLY_SUBMITTED:
1098                 if (!ev->ev_fired) {
1099                         CERROR("RPC %p: bulk %p, service %d\n",
1100                                rpc, rpc->srpc_bulk, sv->sv_id);
1101                         CERROR("Event: status %d, type %d, lnet %d\n",
1102                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1103                         LASSERT (ev->ev_fired);
1104                 }
1105
1106                 wi->swi_state = SWI_STATE_DONE;
1107                 srpc_server_rpc_done(rpc, ev->ev_status);
1108                 return 1;
1109         }
1110
1111         return 0;
1112 }
1113
1114 static void
1115 srpc_client_rpc_expired (void *data)
1116 {
1117         srpc_client_rpc_t *rpc = data;
1118
1119         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1120                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1121                rpc->crpc_timeout);
1122
1123         spin_lock(&rpc->crpc_lock);
1124
1125         rpc->crpc_timeout = 0;
1126         srpc_abort_rpc(rpc, -ETIMEDOUT);
1127
1128         spin_unlock(&rpc->crpc_lock);
1129
1130         spin_lock(&srpc_data.rpc_glock);
1131         srpc_data.rpc_counters.rpcs_expired++;
1132         spin_unlock(&srpc_data.rpc_glock);
1133 }
1134
1135 static void
1136 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1137 {
1138         stt_timer_t *timer = &rpc->crpc_timer;
1139
1140         if (rpc->crpc_timeout == 0)
1141                 return;
1142
1143         INIT_LIST_HEAD(&timer->stt_list);
1144         timer->stt_data    = rpc;
1145         timer->stt_func    = srpc_client_rpc_expired;
1146         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1147                                           cfs_time_current_sec());
1148         stt_add_timer(timer);
1149         return;
1150 }
1151
1152 /*
1153  * Called with rpc->crpc_lock held.
1154  *
1155  * Upon exit the RPC expiry timer is not queued and the handler is not
1156  * running on any CPU. */
1157 static void
1158 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1159 {
1160         /* timer not planted or already exploded */
1161         if (rpc->crpc_timeout == 0)
1162                 return;
1163
1164         /* timer sucessfully defused */
1165         if (stt_del_timer(&rpc->crpc_timer))
1166                 return;
1167
1168 #ifdef __KERNEL__
1169         /* timer detonated, wait for it to explode */
1170         while (rpc->crpc_timeout != 0) {
1171                 spin_unlock(&rpc->crpc_lock);
1172
1173                 schedule();
1174
1175                 spin_lock(&rpc->crpc_lock);
1176         }
1177 #else
1178         LBUG(); /* impossible in single-threaded runtime */
1179 #endif
1180 }
1181
1182 static void
1183 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1184 {
1185         swi_workitem_t *wi = &rpc->crpc_wi;
1186
1187         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1188
1189         spin_lock(&rpc->crpc_lock);
1190
1191         rpc->crpc_closed = 1;
1192         if (rpc->crpc_status == 0)
1193                 rpc->crpc_status = status;
1194
1195         srpc_del_client_rpc_timer(rpc);
1196
1197         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1198                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1199                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1200                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1201
1202         /*
1203          * No one can schedule me now since:
1204          * - RPC timer has been defused.
1205          * - all LNet events have been fired.
1206          * - crpc_closed has been set, preventing srpc_abort_rpc from
1207          *   scheduling me.
1208          * Cancel pending schedules and prevent future schedule attempts:
1209          */
1210         LASSERT (!srpc_event_pending(rpc));
1211         swi_exit_workitem(wi);
1212
1213         spin_unlock(&rpc->crpc_lock);
1214
1215         (*rpc->crpc_done)(rpc);
1216         return;
1217 }
1218
1219 /* sends an outgoing RPC */
1220 int
1221 srpc_send_rpc (swi_workitem_t *wi)
1222 {
1223         int                rc = 0;
1224         srpc_client_rpc_t *rpc;
1225         srpc_msg_t        *reply;
1226         int                do_bulk;
1227
1228         LASSERT(wi != NULL);
1229
1230         rpc = wi->swi_workitem.wi_data;
1231
1232         LASSERT (rpc != NULL);
1233         LASSERT (wi == &rpc->crpc_wi);
1234
1235         reply = &rpc->crpc_replymsg;
1236         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1237
1238         spin_lock(&rpc->crpc_lock);
1239
1240         if (rpc->crpc_aborted) {
1241                 spin_unlock(&rpc->crpc_lock);
1242                 goto abort;
1243         }
1244
1245         spin_unlock(&rpc->crpc_lock);
1246
1247         switch (wi->swi_state) {
1248         default:
1249                 LBUG ();
1250         case SWI_STATE_NEWBORN:
1251                 LASSERT (!srpc_event_pending(rpc));
1252
1253                 rc = srpc_prepare_reply(rpc);
1254                 if (rc != 0) {
1255                         srpc_client_rpc_done(rpc, rc);
1256                         return 1;
1257                 }
1258
1259                 rc = srpc_prepare_bulk(rpc);
1260                 if (rc != 0) break;
1261
1262                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1263                 rc = srpc_send_request(rpc);
1264                 break;
1265
1266         case SWI_STATE_REQUEST_SUBMITTED:
1267                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1268                  * order; however, they're processed in a strict order:
1269                  * rqt, rpy, and bulk. */
1270                 if (!rpc->crpc_reqstev.ev_fired) break;
1271
1272                 rc = rpc->crpc_reqstev.ev_status;
1273                 if (rc != 0) break;
1274
1275                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1276                 /* perhaps more events, fall thru */
1277         case SWI_STATE_REQUEST_SENT: {
1278                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1279
1280                 if (!rpc->crpc_replyev.ev_fired) break;
1281
1282                 rc = rpc->crpc_replyev.ev_status;
1283                 if (rc != 0) break;
1284
1285                 srpc_unpack_msg_hdr(reply);
1286                 if (reply->msg_type != type ||
1287                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1288                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1289                         CWARN ("Bad message from %s: type %u (%d expected),"
1290                                " magic %u (%d expected).\n",
1291                                libcfs_id2str(rpc->crpc_dest),
1292                                reply->msg_type, type,
1293                                reply->msg_magic, SRPC_MSG_MAGIC);
1294                         rc = -EBADMSG;
1295                         break;
1296                 }
1297
1298                 if (do_bulk && reply->msg_body.reply.status != 0) {
1299                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1300                                "case peer didn't initiate bulk transfer\n",
1301                                reply->msg_body.reply.status,
1302                                libcfs_id2str(rpc->crpc_dest));
1303                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1304                 }
1305
1306                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1307         }
1308         case SWI_STATE_REPLY_RECEIVED:
1309                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1310
1311                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1312
1313                 /* Bulk buffer was unlinked due to remote error. Clear error
1314                  * since reply buffer still contains valid data.
1315                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1316                  * remote error. */
1317                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1318                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1319                         rc = 0;
1320
1321                 wi->swi_state = SWI_STATE_DONE;
1322                 srpc_client_rpc_done(rpc, rc);
1323                 return 1;
1324         }
1325
1326         if (rc != 0) {
1327                 spin_lock(&rpc->crpc_lock);
1328                 srpc_abort_rpc(rpc, rc);
1329                 spin_unlock(&rpc->crpc_lock);
1330         }
1331
1332 abort:
1333         if (rpc->crpc_aborted) {
1334                 LNetMDUnlink(rpc->crpc_reqstmdh);
1335                 LNetMDUnlink(rpc->crpc_replymdh);
1336                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1337
1338                 if (!srpc_event_pending(rpc)) {
1339                         srpc_client_rpc_done(rpc, -EINTR);
1340                         return 1;
1341                 }
1342         }
1343         return 0;
1344 }
1345
1346 srpc_client_rpc_t *
1347 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1348                         int nbulkiov, int bulklen,
1349                         void (*rpc_done)(srpc_client_rpc_t *),
1350                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1351 {
1352         srpc_client_rpc_t *rpc;
1353
1354         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1355                                    crpc_bulk.bk_iovs[nbulkiov]));
1356         if (rpc == NULL)
1357                 return NULL;
1358
1359         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1360                              bulklen, rpc_done, rpc_fini, priv);
1361         return rpc;
1362 }
1363
1364 /* called with rpc->crpc_lock held */
1365 void
1366 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1367 {
1368         LASSERT (why != 0);
1369
1370         if (rpc->crpc_aborted || /* already aborted */
1371             rpc->crpc_closed)    /* callback imminent */
1372                 return;
1373
1374         CDEBUG (D_NET,
1375                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1376                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1377                 swi_state2str(rpc->crpc_wi.swi_state), why);
1378
1379         rpc->crpc_aborted = 1;
1380         rpc->crpc_status  = why;
1381         swi_schedule_workitem(&rpc->crpc_wi);
1382         return;
1383 }
1384
1385 /* called with rpc->crpc_lock held */
1386 void
1387 srpc_post_rpc (srpc_client_rpc_t *rpc)
1388 {
1389         LASSERT (!rpc->crpc_aborted);
1390         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1391
1392         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1393                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1394                 rpc->crpc_timeout);
1395
1396         srpc_add_client_rpc_timer(rpc);
1397         swi_schedule_workitem(&rpc->crpc_wi);
1398         return;
1399 }
1400
1401
1402 int
1403 srpc_send_reply(struct srpc_server_rpc *rpc)
1404 {
1405         srpc_event_t            *ev = &rpc->srpc_ev;
1406         struct srpc_msg         *msg = &rpc->srpc_replymsg;
1407         struct srpc_buffer      *buffer = rpc->srpc_reqstbuf;
1408         struct srpc_service_cd  *scd = rpc->srpc_scd;
1409         struct srpc_service     *sv = scd->scd_svc;
1410         __u64                   rpyid;
1411         int                     rc;
1412
1413         LASSERT(buffer != NULL);
1414         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1415
1416         spin_lock(&scd->scd_lock);
1417
1418         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1419                 /* Repost buffer before replying since test client
1420                  * might send me another RPC once it gets the reply */
1421                 if (srpc_service_post_buffer(scd, buffer) != 0)
1422                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1423                 rpc->srpc_reqstbuf = NULL;
1424         }
1425
1426         spin_unlock(&scd->scd_lock);
1427
1428         ev->ev_fired = 0;
1429         ev->ev_data  = rpc;
1430         ev->ev_type  = SRPC_REPLY_SENT;
1431
1432         msg->msg_magic   = SRPC_MSG_MAGIC;
1433         msg->msg_version = SRPC_MSG_VERSION;
1434         msg->msg_type    = srpc_service2reply(sv->sv_id);
1435
1436         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1437                                    sizeof(*msg), LNET_MD_OP_PUT,
1438                                    rpc->srpc_peer, rpc->srpc_self,
1439                                    &rpc->srpc_replymdh, ev);
1440         if (rc != 0)
1441                 ev->ev_fired = 1;  /* no more event expected */
1442         return rc;
1443 }
1444
1445 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1446 static void
1447 srpc_lnet_ev_handler(lnet_event_t *ev)
1448 {
1449         struct srpc_service_cd  *scd;
1450         srpc_event_t      *rpcev = ev->md.user_ptr;
1451         srpc_client_rpc_t *crpc;
1452         srpc_server_rpc_t *srpc;
1453         srpc_buffer_t     *buffer;
1454         srpc_service_t    *sv;
1455         srpc_msg_t        *msg;
1456         srpc_msg_type_t    type;
1457
1458         LASSERT (!in_interrupt());
1459
1460         if (ev->status != 0) {
1461                 __u32 errors;
1462
1463                 spin_lock(&srpc_data.rpc_glock);
1464                 if (ev->status != -ECANCELED) /* cancellation is not error */
1465                         srpc_data.rpc_counters.errors++;
1466                 errors = srpc_data.rpc_counters.errors;
1467                 spin_unlock(&srpc_data.rpc_glock);
1468
1469                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1470                         ev->status, ev->type, errors);
1471         }
1472
1473         rpcev->ev_lnet = ev->type;
1474
1475         switch (rpcev->ev_type) {
1476         default:
1477                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1478                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1479                 LBUG ();
1480         case SRPC_REQUEST_SENT:
1481                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1482                         spin_lock(&srpc_data.rpc_glock);
1483                         srpc_data.rpc_counters.rpcs_sent++;
1484                         spin_unlock(&srpc_data.rpc_glock);
1485                 }
1486         case SRPC_REPLY_RCVD:
1487         case SRPC_BULK_REQ_RCVD:
1488                 crpc = rpcev->ev_data;
1489
1490                 if (rpcev != &crpc->crpc_reqstev &&
1491                     rpcev != &crpc->crpc_replyev &&
1492                     rpcev != &crpc->crpc_bulkev) {
1493                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1494                                rpcev, crpc, &crpc->crpc_reqstev,
1495                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1496                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1497                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1498                         LBUG ();
1499                 }
1500
1501                 spin_lock(&crpc->crpc_lock);
1502
1503                 LASSERT(rpcev->ev_fired == 0);
1504                 rpcev->ev_fired  = 1;
1505                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1506                                                 -EINTR : ev->status;
1507                 swi_schedule_workitem(&crpc->crpc_wi);
1508
1509                 spin_unlock(&crpc->crpc_lock);
1510                 break;
1511
1512         case SRPC_REQUEST_RCVD:
1513                 scd = rpcev->ev_data;
1514                 sv = scd->scd_svc;
1515
1516                 LASSERT(rpcev == &scd->scd_ev);
1517
1518                 spin_lock(&scd->scd_lock);
1519
1520                 LASSERT (ev->unlinked);
1521                 LASSERT (ev->type == LNET_EVENT_PUT ||
1522                          ev->type == LNET_EVENT_UNLINK);
1523                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1524                          sv->sv_shuttingdown);
1525
1526                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1527                 buffer->buf_peer = ev->initiator;
1528                 buffer->buf_self = ev->target.nid;
1529
1530                 LASSERT(scd->scd_buf_nposted > 0);
1531                 scd->scd_buf_nposted--;
1532
1533                 if (sv->sv_shuttingdown) {
1534                         /* Leave buffer on scd->scd_buf_nposted since
1535                          * srpc_finish_service needs to traverse it. */
1536                         spin_unlock(&scd->scd_lock);
1537                         break;
1538                 }
1539
1540                 if (scd->scd_buf_err_stamp != 0 &&
1541                     scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1542                         /* re-enable adding buffer */
1543                         scd->scd_buf_err_stamp = 0;
1544                         scd->scd_buf_err = 0;
1545                 }
1546
1547                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1548                     scd->scd_buf_adjust == 0 &&
1549                     scd->scd_buf_nposted < scd->scd_buf_low) {
1550                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1551                                                   SFW_TEST_WI_MIN);
1552                         swi_schedule_workitem(&scd->scd_buf_wi);
1553                 }
1554
1555                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1556                 msg = &buffer->buf_msg;
1557                 type = srpc_service2request(sv->sv_id);
1558
1559                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1560                     (msg->msg_type != type &&
1561                      msg->msg_type != __swab32(type)) ||
1562                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1563                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1564                         CERROR ("Dropping RPC (%s) from %s: "
1565                                 "status %d mlength %d type %u magic %u.\n",
1566                                 sv->sv_name, libcfs_id2str(ev->initiator),
1567                                 ev->status, ev->mlength,
1568                                 msg->msg_type, msg->msg_magic);
1569
1570                         /* NB can't call srpc_service_recycle_buffer here since
1571                          * it may call LNetM[DE]Attach. The invalid magic tells
1572                          * srpc_handle_rpc to drop this RPC */
1573                         msg->msg_magic = 0;
1574                 }
1575
1576                 if (!list_empty(&scd->scd_rpc_free)) {
1577                         srpc = list_entry(scd->scd_rpc_free.next,
1578                                           struct srpc_server_rpc,
1579                                           srpc_list);
1580                         list_del(&srpc->srpc_list);
1581
1582                         srpc_init_server_rpc(srpc, scd, buffer);
1583                         list_add_tail(&srpc->srpc_list,
1584                                       &scd->scd_rpc_active);
1585                         swi_schedule_workitem(&srpc->srpc_wi);
1586                 } else {
1587                         list_add_tail(&buffer->buf_list,
1588                                       &scd->scd_buf_blocked);
1589                 }
1590
1591                 spin_unlock(&scd->scd_lock);
1592
1593                 spin_lock(&srpc_data.rpc_glock);
1594                 srpc_data.rpc_counters.rpcs_rcvd++;
1595                 spin_unlock(&srpc_data.rpc_glock);
1596                 break;
1597
1598         case SRPC_BULK_GET_RPLD:
1599                 LASSERT (ev->type == LNET_EVENT_SEND ||
1600                          ev->type == LNET_EVENT_REPLY ||
1601                          ev->type == LNET_EVENT_UNLINK);
1602
1603                 if (!ev->unlinked)
1604                         break; /* wait for final event */
1605
1606         case SRPC_BULK_PUT_SENT:
1607                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1608                         spin_lock(&srpc_data.rpc_glock);
1609
1610                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1611                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1612                         else
1613                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1614
1615                         spin_unlock(&srpc_data.rpc_glock);
1616                 }
1617         case SRPC_REPLY_SENT:
1618                 srpc = rpcev->ev_data;
1619                 scd  = srpc->srpc_scd;
1620
1621                 LASSERT(rpcev == &srpc->srpc_ev);
1622
1623                 spin_lock(&scd->scd_lock);
1624
1625                 rpcev->ev_fired  = 1;
1626                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1627                                    -EINTR : ev->status;
1628                 swi_schedule_workitem(&srpc->srpc_wi);
1629
1630                 spin_unlock(&scd->scd_lock);
1631                 break;
1632         }
1633 }
1634
1635 #ifndef __KERNEL__
1636
1637 int
1638 srpc_check_event (int timeout)
1639 {
1640         lnet_event_t ev;
1641         int          rc;
1642         int          i;
1643
1644         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1645                         timeout * 1000, &ev, &i);
1646         if (rc == 0) return 0;
1647
1648         LASSERT (rc == -EOVERFLOW || rc == 1);
1649
1650         /* We can't affort to miss any events... */
1651         if (rc == -EOVERFLOW) {
1652                 CERROR ("Dropped an event!!!\n");
1653                 abort();
1654         }
1655
1656         srpc_lnet_ev_handler(&ev);
1657         return 1;
1658 }
1659
1660 #endif
1661
1662 int
1663 srpc_startup (void)
1664 {
1665         int rc;
1666
1667         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1668         spin_lock_init(&srpc_data.rpc_glock);
1669
1670         /* 1 second pause to avoid timestamp reuse */
1671         cfs_pause(cfs_time_seconds(1));
1672         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1673
1674         srpc_data.rpc_state = SRPC_STATE_NONE;
1675
1676 #ifdef __KERNEL__
1677         rc = LNetNIInit(LNET_PID_LUSTRE);
1678 #else
1679         if (the_lnet.ln_server_mode_flag)
1680                 rc = LNetNIInit(LNET_PID_LUSTRE);
1681         else
1682                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1683 #endif
1684         if (rc < 0) {
1685                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1686                 return rc;
1687         }
1688
1689         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1690
1691         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1692 #ifdef __KERNEL__
1693         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1694 #else
1695         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1696 #endif
1697         if (rc != 0) {
1698                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1699                 goto bail;
1700         }
1701
1702         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1703         LASSERT(rc == 0);
1704         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1705         LASSERT(rc == 0);
1706
1707         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1708
1709         rc = stt_startup();
1710
1711 bail:
1712         if (rc != 0)
1713                 srpc_shutdown();
1714         else
1715                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1716
1717         return rc;
1718 }
1719
1720 void
1721 srpc_shutdown (void)
1722 {
1723         int i;
1724         int rc;
1725         int state;
1726
1727         state = srpc_data.rpc_state;
1728         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1729
1730         switch (state) {
1731         default:
1732                 LBUG ();
1733         case SRPC_STATE_RUNNING:
1734                 spin_lock(&srpc_data.rpc_glock);
1735
1736                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1737                         srpc_service_t *sv = srpc_data.rpc_services[i];
1738
1739                         LASSERTF (sv == NULL,
1740                                   "service not empty: id %d, name %s\n",
1741                                   i, sv->sv_name);
1742                 }
1743
1744                 spin_unlock(&srpc_data.rpc_glock);
1745
1746                 stt_shutdown();
1747
1748         case SRPC_STATE_EQ_INIT:
1749                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1750                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1751                 LASSERT (rc == 0);
1752                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1753                 LASSERT (rc == 0); /* the EQ should have no user by now */
1754
1755         case SRPC_STATE_NI_INIT:
1756                 LNetNIFini();
1757         }
1758
1759         return;
1760 }