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, 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 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 int
  91 srpc_add_bulk_page(srpc_bulk_t *bk, cfs_page_t *pg, int i, int nob)
  92 {
  93         nob = min(nob, (int)CFS_PAGE_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 = cfs_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         cfs_page_t *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                 cfs_free_page(pg);
 132         }
 133
 134 #ifndef __KERNEL__
 135         LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * 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         cfs_page_t  **pages;
 146         int           i;
 147
 148         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 149
 150         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 151                          offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 152         if (bk == NULL) {
 153                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 154                 return NULL;
 155         }
 156
 157         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 158         bk->bk_sink   = sink;
 159         bk->bk_len    = bulk_len;
 160         bk->bk_niov   = bulk_npg;
 161 #ifndef __KERNEL__
 162         LIBCFS_CPT_ALLOC(pages, lnet_cpt_table(), cpt,
 163                          sizeof(cfs_page_t *) * bulk_npg);
 164         if (pages == NULL) {
 165                 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 166                 CERROR("Can't allocate page array for %d pages\n", bulk_npg);
 167                 return NULL;
 168         }
 169
 170         memset(pages, 0, sizeof(cfs_page_t *) * bulk_npg);
 171         bk->bk_pages = pages;
 172 #else
 173         UNUSED(pages);
 174 #endif
 175
 176         for (i = 0; i < bulk_npg; i++) {
 177                 cfs_page_t *pg;
 178                 int         nob;
 179
 180                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, CFS_ALLOC_STD);
 181                 if (pg == NULL) {
 182                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 183                         srpc_free_bulk(bk);
 184                         return NULL;
 185                 }
 186
 187                 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
 188                 bulk_len -= nob;
 189         }
 190
 191         return bk;
 192 }
 193
 194 static inline __u64
 195 srpc_next_id (void)
 196 {
 197         __u64 id;
 198
 199         spin_lock(&srpc_data.rpc_glock);
 200         id = srpc_data.rpc_matchbits++;
 201         spin_unlock(&srpc_data.rpc_glock);
 202         return id;
 203 }
 204
 205 void
 206 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 207                      struct srpc_service_cd *scd,
 208                      struct srpc_buffer *buffer)
 209 {
 210         memset(rpc, 0, sizeof(*rpc));
 211         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
 212                           srpc_serv_is_framework(scd->scd_svc) ?
 213                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 214
 215         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 216
 217         rpc->srpc_scd      = scd;
 218         rpc->srpc_reqstbuf = buffer;
 219         rpc->srpc_peer     = buffer->buf_peer;
 220         rpc->srpc_self     = buffer->buf_self;
 221         LNetInvalidateHandle(&rpc->srpc_replymdh);
 222 }
 223
 224 static void
 225 srpc_service_fini(struct srpc_service *svc)
 226 {
 227         struct srpc_service_cd  *scd;
 228         struct srpc_server_rpc  *rpc;
 229         struct srpc_buffer      *buf;
 230         cfs_list_t              *q;
 231         int                     i;
 232
 233         if (svc->sv_cpt_data == NULL)
 234                 return;
 235
 236         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 237                 while (1) {
 238                         if (!cfs_list_empty(&scd->scd_buf_posted))
 239                                 q = &scd->scd_buf_posted;
 240                         else if (!cfs_list_empty(&scd->scd_buf_blocked))
 241                                 q = &scd->scd_buf_blocked;
 242                         else
 243                                 break;
 244
 245                         while (!cfs_list_empty(q)) {
 246                                 buf = cfs_list_entry(q->next,
 247                                                      struct srpc_buffer,
 248                                                      buf_list);
 249                                 cfs_list_del(&buf->buf_list);
 250                                 LIBCFS_FREE(buf, sizeof(*buf));
 251                         }
 252                 }
 253
 254                 LASSERT(cfs_list_empty(&scd->scd_rpc_active));
 255
 256                 while (!cfs_list_empty(&scd->scd_rpc_free)) {
 257                         rpc = cfs_list_entry(scd->scd_rpc_free.next,
 258                                              struct srpc_server_rpc,
 259                                              srpc_list);
 260                         cfs_list_del(&rpc->srpc_list);
 261                         LIBCFS_FREE(rpc, sizeof(*rpc));
 262                 }
 263         }
 264
 265         cfs_percpt_free(svc->sv_cpt_data);
 266         svc->sv_cpt_data = NULL;
 267 }
 268
 269 static int
 270 srpc_service_nrpcs(struct srpc_service *svc)
 271 {
 272         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 273
 274         return srpc_serv_is_framework(svc) ?
 275                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 276 }
 277
 278 int srpc_add_buffer(struct swi_workitem *wi);
 279
 280 static int
 281 srpc_service_init(struct srpc_service *svc)
 282 {
 283         struct srpc_service_cd  *scd;
 284         struct srpc_server_rpc  *rpc;
 285         int                     nrpcs;
 286         int                     i;
 287         int                     j;
 288
 289         svc->sv_shuttingdown = 0;
 290
 291         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 292                                             sizeof(struct srpc_service_cd));
 293         if (svc->sv_cpt_data == NULL)
 294                 return -ENOMEM;
 295
 296         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 297                         1 : cfs_cpt_number(lnet_cpt_table());
 298         nrpcs = srpc_service_nrpcs(svc);
 299
 300         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 301                 scd->scd_cpt = i;
 302                 scd->scd_svc = svc;
 303                 spin_lock_init(&scd->scd_lock);
 304                 CFS_INIT_LIST_HEAD(&scd->scd_rpc_free);
 305                 CFS_INIT_LIST_HEAD(&scd->scd_rpc_active);
 306                 CFS_INIT_LIST_HEAD(&scd->scd_buf_posted);
 307                 CFS_INIT_LIST_HEAD(&scd->scd_buf_blocked);
 308
 309                 scd->scd_ev.ev_data = scd;
 310                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 311
 312                 /* NB: don't use lst_sched_serial for adding buffer,
 313                  * see details in srpc_service_add_buffers() */
 314                 swi_init_workitem(&scd->scd_buf_wi, scd,
 315                                   srpc_add_buffer, lst_sched_test[i]);
 316
 317                 if (i != 0 && srpc_serv_is_framework(svc)) {
 318                         /* NB: framework service only needs srpc_service_cd for
 319                          * one partition, but we allocate for all to make
 320                          * it easier to implement, it will waste a little
 321                          * memory but nobody should care about this */
 322                         continue;
 323                 }
 324
 325                 for (j = 0; j < nrpcs; j++) {
 326                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 327                                          i, sizeof(*rpc));
 328                         if (rpc == NULL) {
 329                                 srpc_service_fini(svc);
 330                                 return -ENOMEM;
 331                         }
 332                         cfs_list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 333                 }
 334         }
 335
 336         return 0;
 337 }
 338
 339 int
 340 srpc_add_service(struct srpc_service *sv)
 341 {
 342         int id = sv->sv_id;
 343
 344         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 345
 346         if (srpc_service_init(sv) != 0)
 347                 return -ENOMEM;
 348
 349         spin_lock(&srpc_data.rpc_glock);
 350
 351         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 352
 353         if (srpc_data.rpc_services[id] != NULL) {
 354                 spin_unlock(&srpc_data.rpc_glock);
 355                 goto failed;
 356         }
 357
 358         srpc_data.rpc_services[id] = sv;
 359         spin_unlock(&srpc_data.rpc_glock);
 360
 361         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 362         return 0;
 363
 364  failed:
 365         srpc_service_fini(sv);
 366         return -EBUSY;
 367 }
 368
 369 int
 370 srpc_remove_service (srpc_service_t *sv)
 371 {
 372         int id = sv->sv_id;
 373
 374         spin_lock(&srpc_data.rpc_glock);
 375
 376         if (srpc_data.rpc_services[id] != sv) {
 377                 spin_unlock(&srpc_data.rpc_glock);
 378                 return -ENOENT;
 379         }
 380
 381         srpc_data.rpc_services[id] = NULL;
 382         spin_unlock(&srpc_data.rpc_glock);
 383         return 0;
 384 }
 385
 386 int
 387 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 388                        int len, int options, lnet_process_id_t peer,
 389                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 390 {
 391         int              rc;
 392         lnet_md_t        md;
 393         lnet_handle_me_t meh;
 394
 395         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 396                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 397         if (rc != 0) {
 398                 CERROR ("LNetMEAttach failed: %d\n", rc);
 399                 LASSERT (rc == -ENOMEM);
 400                 return -ENOMEM;
 401         }
 402
 403         md.threshold = 1;
 404         md.user_ptr  = ev;
 405         md.start     = buf;
 406         md.length    = len;
 407         md.options   = options;
 408         md.eq_handle = srpc_data.rpc_lnet_eq;
 409
 410         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 411         if (rc != 0) {
 412                 CERROR ("LNetMDAttach failed: %d\n", rc);
 413                 LASSERT (rc == -ENOMEM);
 414
 415                 rc = LNetMEUnlink(meh);
 416                 LASSERT (rc == 0);
 417                 return -ENOMEM;
 418         }
 419
 420         CDEBUG (D_NET,
 421                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 422                 libcfs_id2str(peer), portal, matchbits);
 423         return 0;
 424 }
 425
 426 int
 427 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 428                       int options, lnet_process_id_t peer, lnet_nid_t self,
 429                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 430 {
 431         int       rc;
 432         lnet_md_t md;
 433
 434         md.user_ptr  = ev;
 435         md.start     = buf;
 436         md.length    = len;
 437         md.eq_handle = srpc_data.rpc_lnet_eq;
 438         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 439         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 440
 441         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 442         if (rc != 0) {
 443                 CERROR ("LNetMDBind failed: %d\n", rc);
 444                 LASSERT (rc == -ENOMEM);
 445                 return -ENOMEM;
 446         }
 447
 448         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 449          * they're only meaningful for MDs attached to an ME (i.e. passive
 450          * buffers... */
 451         if ((options & LNET_MD_OP_PUT) != 0) {
 452                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 453                              portal, matchbits, 0, 0);
 454         } else {
 455                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 456
 457                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 458         }
 459
 460         if (rc != 0) {
 461                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 462                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 463                         libcfs_id2str(peer), portal, matchbits, rc);
 464
 465                 /* The forthcoming unlink event will complete this operation
 466                  * with failure, so fall through and return success here.
 467                  */
 468                 rc = LNetMDUnlink(*mdh);
 469                 LASSERT (rc == 0);
 470         } else {
 471                 CDEBUG (D_NET,
 472                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 473                         libcfs_id2str(peer), portal, matchbits);
 474         }
 475         return 0;
 476 }
 477
 478 int
 479 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 480                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 481 {
 482         return srpc_post_active_rdma(srpc_serv_portal(service), service,
 483                                      buf, len, LNET_MD_OP_PUT, peer,
 484                                      LNET_NID_ANY, mdh, ev);
 485 }
 486
 487 int
 488 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 489                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 490 {
 491         lnet_process_id_t any = {0};
 492
 493         any.nid = LNET_NID_ANY;
 494         any.pid = LNET_PID_ANY;
 495
 496         return srpc_post_passive_rdma(srpc_serv_portal(service),
 497                                       local, service, buf, len,
 498                                       LNET_MD_OP_PUT, any, mdh, ev);
 499 }
 500
 501 int
 502 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 503 {
 504         struct srpc_service     *sv = scd->scd_svc;
 505         struct srpc_msg         *msg = &buf->buf_msg;
 506         int                     rc;
 507
 508         LNetInvalidateHandle(&buf->buf_mdh);
 509         cfs_list_add(&buf->buf_list, &scd->scd_buf_posted);
 510         scd->scd_buf_nposted++;
 511         spin_unlock(&scd->scd_lock);
 512
 513         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 514                                       !srpc_serv_is_framework(sv),
 515                                       msg, sizeof(*msg), &buf->buf_mdh,
 516                                       &scd->scd_ev);
 517
 518         /* At this point, a RPC (new or delayed) may have arrived in
 519          * msg and its event handler has been called. So we must add
 520          * buf to scd_buf_posted _before_ dropping scd_lock */
 521
 522         spin_lock(&scd->scd_lock);
 523
 524         if (rc == 0) {
 525                 if (!sv->sv_shuttingdown)
 526                         return 0;
 527
 528                 spin_unlock(&scd->scd_lock);
 529                 /* srpc_shutdown_service might have tried to unlink me
 530                  * when my buf_mdh was still invalid */
 531                 LNetMDUnlink(buf->buf_mdh);
 532                 spin_lock(&scd->scd_lock);
 533                 return 0;
 534         }
 535
 536         scd->scd_buf_nposted--;
 537         if (sv->sv_shuttingdown)
 538                 return rc; /* don't allow to change scd_buf_posted */
 539
 540         cfs_list_del(&buf->buf_list);
 541         spin_unlock(&scd->scd_lock);
 542
 543         LIBCFS_FREE(buf, sizeof(*buf));
 544
 545         spin_lock(&scd->scd_lock);
 546         return rc;
 547 }
 548
 549 int
 550 srpc_add_buffer(struct swi_workitem *wi)
 551 {
 552         struct srpc_service_cd  *scd = wi->swi_workitem.wi_data;
 553         struct srpc_buffer      *buf;
 554         int                     rc = 0;
 555
 556         /* it's called by workitem scheduler threads, these threads
 557          * should have been set CPT affinity, so buffers will be posted
 558          * on CPT local list of Portal */
 559         spin_lock(&scd->scd_lock);
 560
 561         while (scd->scd_buf_adjust > 0 &&
 562                !scd->scd_svc->sv_shuttingdown) {
 563                 scd->scd_buf_adjust--; /* consume it */
 564                 scd->scd_buf_posting++;
 565
 566                 spin_unlock(&scd->scd_lock);
 567
 568                 LIBCFS_ALLOC(buf, sizeof(*buf));
 569                 if (buf == NULL) {
 570                         CERROR("Failed to add new buf to service: %s\n",
 571                                scd->scd_svc->sv_name);
 572                         spin_lock(&scd->scd_lock);
 573                         rc = -ENOMEM;
 574                         break;
 575                 }
 576
 577                 spin_lock(&scd->scd_lock);
 578                 if (scd->scd_svc->sv_shuttingdown) {
 579                         spin_unlock(&scd->scd_lock);
 580                         LIBCFS_FREE(buf, sizeof(*buf));
 581
 582                         spin_lock(&scd->scd_lock);
 583                         rc = -ESHUTDOWN;
 584                         break;
 585                 }
 586
 587                 rc = srpc_service_post_buffer(scd, buf);
 588                 if (rc != 0)
 589                         break; /* buf has been freed inside */
 590
 591                 LASSERT(scd->scd_buf_posting > 0);
 592                 scd->scd_buf_posting--;
 593                 scd->scd_buf_total++;
 594                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 595         }
 596
 597         if (rc != 0) {
 598                 scd->scd_buf_err_stamp = cfs_time_current_sec();
 599                 scd->scd_buf_err = rc;
 600
 601                 LASSERT(scd->scd_buf_posting > 0);
 602                 scd->scd_buf_posting--;
 603         }
 604
 605         spin_unlock(&scd->scd_lock);
 606         return 0;
 607 }
 608
 609 int
 610 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 611 {
 612         struct srpc_service_cd  *scd;
 613         int                     rc = 0;
 614         int                     i;
 615
 616         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 617
 618         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 619                 spin_lock(&scd->scd_lock);
 620
 621                 scd->scd_buf_err = 0;
 622                 scd->scd_buf_err_stamp = 0;
 623                 scd->scd_buf_posting = 0;
 624                 scd->scd_buf_adjust = nbuffer;
 625                 /* start to post buffers */
 626                 swi_schedule_workitem(&scd->scd_buf_wi);
 627                 spin_unlock(&scd->scd_lock);
 628
 629                 /* framework service only post buffer for one partition  */
 630                 if (srpc_serv_is_framework(sv))
 631                         break;
 632         }
 633
 634         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 635                 spin_lock(&scd->scd_lock);
 636                 /*
 637                  * NB: srpc_service_add_buffers() can be called inside
 638                  * thread context of lst_sched_serial, and we don't normally
 639                  * allow to sleep inside thread context of WI scheduler
 640                  * because it will block current scheduler thread from doing
 641                  * anything else, even worse, it could deadlock if it's
 642                  * waiting on result from another WI of the same scheduler.
 643                  * However, it's safe at here because scd_buf_wi is scheduled
 644                  * by thread in a different WI scheduler (lst_sched_test),
 645                  * so we don't have any risk of deadlock, though this could
 646                  * block all WIs pending on lst_sched_serial for a moment
 647                  * which is not good but not fatal.
 648                  */
 649                 lst_wait_until(scd->scd_buf_err != 0 ||
 650                                (scd->scd_buf_adjust == 0 &&
 651                                 scd->scd_buf_posting == 0),
 652                                scd->scd_lock, "waiting for adding buffer\n");
 653
 654                 if (scd->scd_buf_err != 0 && rc == 0)
 655                         rc = scd->scd_buf_err;
 656
 657                 spin_unlock(&scd->scd_lock);
 658         }
 659
 660         return rc;
 661 }
 662
 663 void
 664 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 665 {
 666         struct srpc_service_cd  *scd;
 667         int                     num;
 668         int                     i;
 669
 670         LASSERT(!sv->sv_shuttingdown);
 671
 672         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 673                 spin_lock(&scd->scd_lock);
 674
 675                 num = scd->scd_buf_total + scd->scd_buf_posting;
 676                 scd->scd_buf_adjust -= min(nbuffer, num);
 677
 678                 spin_unlock(&scd->scd_lock);
 679         }
 680 }
 681
 682 /* returns 1 if sv has finished, otherwise 0 */
 683 int
 684 srpc_finish_service(struct srpc_service *sv)
 685 {
 686         struct srpc_service_cd  *scd;
 687         struct srpc_server_rpc  *rpc;
 688         int                     i;
 689
 690         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 691
 692         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 693                 spin_lock(&scd->scd_lock);
 694                 if (!swi_deschedule_workitem(&scd->scd_buf_wi))
 695                         return 0;
 696
 697                 if (scd->scd_buf_nposted > 0) {
 698                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink",
 699                                scd->scd_buf_nposted);
 700                         spin_unlock(&scd->scd_lock);
 701                         return 0;
 702                 }
 703
 704                 if (cfs_list_empty(&scd->scd_rpc_active)) {
 705                         spin_unlock(&scd->scd_lock);
 706                         continue;
 707                 }
 708
 709                 rpc = cfs_list_entry(scd->scd_rpc_active.next,
 710                                      struct srpc_server_rpc, srpc_list);
 711                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
 712                         "wi %s scheduled %d running %d, "
 713                         "ev fired %d type %d status %d lnet %d\n",
 714                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 715                         swi_state2str(rpc->srpc_wi.swi_state),
 716                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 717                         rpc->srpc_wi.swi_workitem.wi_running,
 718                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 719                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 720                 spin_unlock(&scd->scd_lock);
 721                 return 0;
 722         }
 723
 724         /* no lock needed from now on */
 725         srpc_service_fini(sv);
 726         return 1;
 727 }
 728
 729 /* called with sv->sv_lock held */
 730 void
 731 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
 732 {
 733         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 734                 if (srpc_service_post_buffer(scd, buf) != 0) {
 735                         CWARN("Failed to post %s buffer\n",
 736                               scd->scd_svc->sv_name);
 737                 }
 738                 return;
 739         }
 740
 741         /* service is shutting down, or we want to recycle some buffers */
 742         scd->scd_buf_total--;
 743
 744         if (scd->scd_buf_adjust < 0) {
 745                 scd->scd_buf_adjust++;
 746                 if (scd->scd_buf_adjust < 0 &&
 747                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 748                         CDEBUG(D_INFO,
 749                                "Try to recyle %d buffers but nothing left\n",
 750                                scd->scd_buf_adjust);
 751                         scd->scd_buf_adjust = 0;
 752                 }
 753         }
 754
 755         spin_unlock(&scd->scd_lock);
 756         LIBCFS_FREE(buf, sizeof(*buf));
 757         spin_lock(&scd->scd_lock);
 758 }
 759
 760 void
 761 srpc_abort_service(struct srpc_service *sv)
 762 {
 763         struct srpc_service_cd  *scd;
 764         struct srpc_server_rpc  *rpc;
 765         int                     i;
 766
 767         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 768                sv->sv_id, sv->sv_name);
 769
 770         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 771                 spin_lock(&scd->scd_lock);
 772
 773                 /* schedule in-flight RPCs to notice the abort, NB:
 774                  * racing with incoming RPCs; complete fix should make test
 775                  * RPCs carry session ID in its headers */
 776                 cfs_list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 777                         rpc->srpc_aborted = 1;
 778                         swi_schedule_workitem(&rpc->srpc_wi);
 779                 }
 780
 781                 spin_unlock(&scd->scd_lock);
 782         }
 783 }
 784
 785 void
 786 srpc_shutdown_service(srpc_service_t *sv)
 787 {
 788         struct srpc_service_cd  *scd;
 789         struct srpc_server_rpc  *rpc;
 790         srpc_buffer_t           *buf;
 791         int                     i;
 792
 793         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 794                sv->sv_id, sv->sv_name);
 795
 796         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 797                 spin_lock(&scd->scd_lock);
 798
 799         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 800
 801         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 802                 spin_unlock(&scd->scd_lock);
 803
 804         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 805                 spin_lock(&scd->scd_lock);
 806
 807                 /* schedule in-flight RPCs to notice the shutdown */
 808                 cfs_list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 809                         swi_schedule_workitem(&rpc->srpc_wi);
 810
 811                 spin_unlock(&scd->scd_lock);
 812
 813                 /* OK to traverse scd_buf_posted without lock, since no one
 814                  * touches scd_buf_posted now */
 815                 cfs_list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 816                         LNetMDUnlink(buf->buf_mdh);
 817         }
 818 }
 819
 820 int
 821 srpc_send_request (srpc_client_rpc_t *rpc)
 822 {
 823         srpc_event_t *ev = &rpc->crpc_reqstev;
 824         int           rc;
 825
 826         ev->ev_fired = 0;
 827         ev->ev_data  = rpc;
 828         ev->ev_type  = SRPC_REQUEST_SENT;
 829
 830         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 831                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 832                                      &rpc->crpc_reqstmdh, ev);
 833         if (rc != 0) {
 834                 LASSERT (rc == -ENOMEM);
 835                 ev->ev_fired = 1;  /* no more event expected */
 836         }
 837         return rc;
 838 }
 839
 840 int
 841 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 842 {
 843         srpc_event_t *ev = &rpc->crpc_replyev;
 844         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 845         int           rc;
 846
 847         ev->ev_fired = 0;
 848         ev->ev_data  = rpc;
 849         ev->ev_type  = SRPC_REPLY_RCVD;
 850
 851         *id = srpc_next_id();
 852
 853         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 854                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 855                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 856                                     &rpc->crpc_replymdh, ev);
 857         if (rc != 0) {
 858                 LASSERT (rc == -ENOMEM);
 859                 ev->ev_fired = 1;  /* no more event expected */
 860         }
 861         return rc;
 862 }
 863
 864 int
 865 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 866 {
 867         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 868         srpc_event_t *ev = &rpc->crpc_bulkev;
 869         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 870         int           rc;
 871         int           opt;
 872
 873         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 874
 875         if (bk->bk_niov == 0) return 0; /* nothing to do */
 876
 877         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 878 #ifdef __KERNEL__
 879         opt |= LNET_MD_KIOV;
 880 #else
 881         opt |= LNET_MD_IOVEC;
 882 #endif
 883
 884         ev->ev_fired = 0;
 885         ev->ev_data  = rpc;
 886         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 887
 888         *id = srpc_next_id();
 889
 890         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 891                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 892                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 893         if (rc != 0) {
 894                 LASSERT (rc == -ENOMEM);
 895                 ev->ev_fired = 1;  /* no more event expected */
 896         }
 897         return rc;
 898 }
 899
 900 int
 901 srpc_do_bulk (srpc_server_rpc_t *rpc)
 902 {
 903         srpc_event_t  *ev = &rpc->srpc_ev;
 904         srpc_bulk_t   *bk = rpc->srpc_bulk;
 905         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 906         int            rc;
 907         int            opt;
 908
 909         LASSERT (bk != NULL);
 910
 911         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 912 #ifdef __KERNEL__
 913         opt |= LNET_MD_KIOV;
 914 #else
 915         opt |= LNET_MD_IOVEC;
 916 #endif
 917
 918         ev->ev_fired = 0;
 919         ev->ev_data  = rpc;
 920         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 921
 922         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 923                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 924                                    rpc->srpc_peer, rpc->srpc_self,
 925                                    &bk->bk_mdh, ev);
 926         if (rc != 0)
 927                 ev->ev_fired = 1;  /* no more event expected */
 928         return rc;
 929 }
 930
 931 /* only called from srpc_handle_rpc */
 932 void
 933 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
 934 {
 935         struct srpc_service_cd  *scd = rpc->srpc_scd;
 936         struct srpc_service     *sv  = scd->scd_svc;
 937         srpc_buffer_t           *buffer;
 938
 939         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 940
 941         rpc->srpc_status = status;
 942
 943         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 944                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 945                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 946                 swi_state2str(rpc->srpc_wi.swi_state), status);
 947
 948         if (status != 0) {
 949                 spin_lock(&srpc_data.rpc_glock);
 950                 srpc_data.rpc_counters.rpcs_dropped++;
 951                 spin_unlock(&srpc_data.rpc_glock);
 952         }
 953
 954         if (rpc->srpc_done != NULL)
 955                 (*rpc->srpc_done) (rpc);
 956         LASSERT(rpc->srpc_bulk == NULL);
 957
 958         spin_lock(&scd->scd_lock);
 959
 960         if (rpc->srpc_reqstbuf != NULL) {
 961                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 962                  * sv won't go away for scd_rpc_active must not be empty */
 963                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 964                 rpc->srpc_reqstbuf = NULL;
 965         }
 966
 967         cfs_list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 968
 969         /*
 970          * No one can schedule me now since:
 971          * - I'm not on scd_rpc_active.
 972          * - all LNet events have been fired.
 973          * Cancel pending schedules and prevent future schedule attempts:
 974          */
 975         LASSERT(rpc->srpc_ev.ev_fired);
 976         swi_exit_workitem(&rpc->srpc_wi);
 977
 978         if (!sv->sv_shuttingdown && !cfs_list_empty(&scd->scd_buf_blocked)) {
 979                 buffer = cfs_list_entry(scd->scd_buf_blocked.next,
 980                                         srpc_buffer_t, buf_list);
 981                 cfs_list_del(&buffer->buf_list);
 982
 983                 srpc_init_server_rpc(rpc, scd, buffer);
 984                 cfs_list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 985                 swi_schedule_workitem(&rpc->srpc_wi);
 986         } else {
 987                 cfs_list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 988         }
 989
 990         spin_unlock(&scd->scd_lock);
 991         return;
 992 }
 993
 994 /* handles an incoming RPC */
 995 int
 996 srpc_handle_rpc(swi_workitem_t *wi)
 997 {
 998         struct srpc_server_rpc  *rpc = wi->swi_workitem.wi_data;
 999         struct srpc_service_cd  *scd = rpc->srpc_scd;
1000         struct srpc_service     *sv = scd->scd_svc;
1001         srpc_event_t            *ev = &rpc->srpc_ev;
1002         int                     rc = 0;
1003
1004         LASSERT(wi == &rpc->srpc_wi);
1005
1006         spin_lock(&scd->scd_lock);
1007
1008         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
1009                 spin_unlock(&scd->scd_lock);
1010
1011                 if (rpc->srpc_bulk != NULL)
1012                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
1013                 LNetMDUnlink(rpc->srpc_replymdh);
1014
1015                 if (ev->ev_fired) { /* no more event, OK to finish */
1016                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
1017                         return 1;
1018                 }
1019                 return 0;
1020         }
1021
1022         spin_unlock(&scd->scd_lock);
1023
1024         switch (wi->swi_state) {
1025         default:
1026                 LBUG ();
1027         case SWI_STATE_NEWBORN: {
1028                 srpc_msg_t           *msg;
1029                 srpc_generic_reply_t *reply;
1030
1031                 msg = &rpc->srpc_reqstbuf->buf_msg;
1032                 reply = &rpc->srpc_replymsg.msg_body.reply;
1033
1034                 if (msg->msg_magic == 0) {
1035                         /* moaned already in srpc_lnet_ev_handler */
1036                         srpc_server_rpc_done(rpc, EBADMSG);
1037                         return 1;
1038                 }
1039
1040                 srpc_unpack_msg_hdr(msg);
1041                 if (msg->msg_version != SRPC_MSG_VERSION) {
1042                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1043                               msg->msg_version, SRPC_MSG_VERSION,
1044                               libcfs_id2str(rpc->srpc_peer));
1045                         reply->status = EPROTO;
1046                         /* drop through and send reply */
1047                 } else {
1048                         reply->status = 0;
1049                         rc = (*sv->sv_handler)(rpc);
1050                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1051                         if (rc != 0) {
1052                                 srpc_server_rpc_done(rpc, rc);
1053                                 return 1;
1054                         }
1055                 }
1056
1057                 wi->swi_state = SWI_STATE_BULK_STARTED;
1058
1059                 if (rpc->srpc_bulk != NULL) {
1060                         rc = srpc_do_bulk(rpc);
1061                         if (rc == 0)
1062                                 return 0; /* wait for bulk */
1063
1064                         LASSERT (ev->ev_fired);
1065                         ev->ev_status = rc;
1066                 }
1067         }
1068         case SWI_STATE_BULK_STARTED:
1069                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1070
1071                 if (rpc->srpc_bulk != NULL) {
1072                         rc = ev->ev_status;
1073
1074                         if (sv->sv_bulk_ready != NULL)
1075                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1076
1077                         if (rc != 0) {
1078                                 srpc_server_rpc_done(rpc, rc);
1079                                 return 1;
1080                         }
1081                 }
1082
1083                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1084                 rc = srpc_send_reply(rpc);
1085                 if (rc == 0)
1086                         return 0; /* wait for reply */
1087                 srpc_server_rpc_done(rpc, rc);
1088                 return 1;
1089
1090         case SWI_STATE_REPLY_SUBMITTED:
1091                 if (!ev->ev_fired) {
1092                         CERROR("RPC %p: bulk %p, service %d\n",
1093                                rpc, rpc->srpc_bulk, sv->sv_id);
1094                         CERROR("Event: status %d, type %d, lnet %d\n",
1095                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1096                         LASSERT (ev->ev_fired);
1097                 }
1098
1099                 wi->swi_state = SWI_STATE_DONE;
1100                 srpc_server_rpc_done(rpc, ev->ev_status);
1101                 return 1;
1102         }
1103
1104         return 0;
1105 }
1106
1107 void
1108 srpc_client_rpc_expired (void *data)
1109 {
1110         srpc_client_rpc_t *rpc = data;
1111
1112         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1113                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1114                rpc->crpc_timeout);
1115
1116         spin_lock(&rpc->crpc_lock);
1117
1118         rpc->crpc_timeout = 0;
1119         srpc_abort_rpc(rpc, -ETIMEDOUT);
1120
1121         spin_unlock(&rpc->crpc_lock);
1122
1123         spin_lock(&srpc_data.rpc_glock);
1124         srpc_data.rpc_counters.rpcs_expired++;
1125         spin_unlock(&srpc_data.rpc_glock);
1126 }
1127
1128 inline void
1129 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
1130 {
1131         stt_timer_t *timer = &rpc->crpc_timer;
1132
1133         if (rpc->crpc_timeout == 0) return;
1134
1135         CFS_INIT_LIST_HEAD(&timer->stt_list);
1136         timer->stt_data    = rpc;
1137         timer->stt_func    = srpc_client_rpc_expired;
1138         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1139                                           cfs_time_current_sec());
1140         stt_add_timer(timer);
1141         return;
1142 }
1143
1144 /*
1145  * Called with rpc->crpc_lock held.
1146  *
1147  * Upon exit the RPC expiry timer is not queued and the handler is not
1148  * running on any CPU. */
1149 void
1150 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1151 {
1152         /* timer not planted or already exploded */
1153         if (rpc->crpc_timeout == 0)
1154                 return;
1155
1156         /* timer sucessfully defused */
1157         if (stt_del_timer(&rpc->crpc_timer))
1158                 return;
1159
1160 #ifdef __KERNEL__
1161         /* timer detonated, wait for it to explode */
1162         while (rpc->crpc_timeout != 0) {
1163                 spin_unlock(&rpc->crpc_lock);
1164
1165                 cfs_schedule();
1166
1167                 spin_lock(&rpc->crpc_lock);
1168         }
1169 #else
1170         LBUG(); /* impossible in single-threaded runtime */
1171 #endif
1172 }
1173
1174 void
1175 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1176 {
1177         swi_workitem_t *wi = &rpc->crpc_wi;
1178
1179         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1180
1181         spin_lock(&rpc->crpc_lock);
1182
1183         rpc->crpc_closed = 1;
1184         if (rpc->crpc_status == 0)
1185                 rpc->crpc_status = status;
1186
1187         srpc_del_client_rpc_timer(rpc);
1188
1189         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1190                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1191                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1192                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1193
1194         /*
1195          * No one can schedule me now since:
1196          * - RPC timer has been defused.
1197          * - all LNet events have been fired.
1198          * - crpc_closed has been set, preventing srpc_abort_rpc from
1199          *   scheduling me.
1200          * Cancel pending schedules and prevent future schedule attempts:
1201          */
1202         LASSERT (!srpc_event_pending(rpc));
1203         swi_exit_workitem(wi);
1204
1205         spin_unlock(&rpc->crpc_lock);
1206
1207         (*rpc->crpc_done)(rpc);
1208         return;
1209 }
1210
1211 /* sends an outgoing RPC */
1212 int
1213 srpc_send_rpc (swi_workitem_t *wi)
1214 {
1215         int                rc = 0;
1216         srpc_client_rpc_t *rpc;
1217         srpc_msg_t        *reply;
1218         int                do_bulk;
1219
1220         LASSERT(wi != NULL);
1221
1222         rpc = wi->swi_workitem.wi_data;
1223
1224         LASSERT (rpc != NULL);
1225         LASSERT (wi == &rpc->crpc_wi);
1226
1227         reply = &rpc->crpc_replymsg;
1228         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1229
1230         spin_lock(&rpc->crpc_lock);
1231
1232         if (rpc->crpc_aborted) {
1233                 spin_unlock(&rpc->crpc_lock);
1234                 goto abort;
1235         }
1236
1237         spin_unlock(&rpc->crpc_lock);
1238
1239         switch (wi->swi_state) {
1240         default:
1241                 LBUG ();
1242         case SWI_STATE_NEWBORN:
1243                 LASSERT (!srpc_event_pending(rpc));
1244
1245                 rc = srpc_prepare_reply(rpc);
1246                 if (rc != 0) {
1247                         srpc_client_rpc_done(rpc, rc);
1248                         return 1;
1249                 }
1250
1251                 rc = srpc_prepare_bulk(rpc);
1252                 if (rc != 0) break;
1253
1254                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1255                 rc = srpc_send_request(rpc);
1256                 break;
1257
1258         case SWI_STATE_REQUEST_SUBMITTED:
1259                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1260                  * order; however, they're processed in a strict order:
1261                  * rqt, rpy, and bulk. */
1262                 if (!rpc->crpc_reqstev.ev_fired) break;
1263
1264                 rc = rpc->crpc_reqstev.ev_status;
1265                 if (rc != 0) break;
1266
1267                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1268                 /* perhaps more events, fall thru */
1269         case SWI_STATE_REQUEST_SENT: {
1270                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1271
1272                 if (!rpc->crpc_replyev.ev_fired) break;
1273
1274                 rc = rpc->crpc_replyev.ev_status;
1275                 if (rc != 0) break;
1276
1277                 srpc_unpack_msg_hdr(reply);
1278                 if (reply->msg_type != type ||
1279                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1280                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1281                         CWARN ("Bad message from %s: type %u (%d expected),"
1282                                " magic %u (%d expected).\n",
1283                                libcfs_id2str(rpc->crpc_dest),
1284                                reply->msg_type, type,
1285                                reply->msg_magic, SRPC_MSG_MAGIC);
1286                         rc = -EBADMSG;
1287                         break;
1288                 }
1289
1290                 if (do_bulk && reply->msg_body.reply.status != 0) {
1291                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1292                                "case peer didn't initiate bulk transfer\n",
1293                                reply->msg_body.reply.status,
1294                                libcfs_id2str(rpc->crpc_dest));
1295                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1296                 }
1297
1298                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1299         }
1300         case SWI_STATE_REPLY_RECEIVED:
1301                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1302
1303                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1304
1305                 /* Bulk buffer was unlinked due to remote error. Clear error
1306                  * since reply buffer still contains valid data.
1307                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1308                  * remote error. */
1309                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1310                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1311                         rc = 0;
1312
1313                 wi->swi_state = SWI_STATE_DONE;
1314                 srpc_client_rpc_done(rpc, rc);
1315                 return 1;
1316         }
1317
1318         if (rc != 0) {
1319                 spin_lock(&rpc->crpc_lock);
1320                 srpc_abort_rpc(rpc, rc);
1321                 spin_unlock(&rpc->crpc_lock);
1322         }
1323
1324 abort:
1325         if (rpc->crpc_aborted) {
1326                 LNetMDUnlink(rpc->crpc_reqstmdh);
1327                 LNetMDUnlink(rpc->crpc_replymdh);
1328                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1329
1330                 if (!srpc_event_pending(rpc)) {
1331                         srpc_client_rpc_done(rpc, -EINTR);
1332                         return 1;
1333                 }
1334         }
1335         return 0;
1336 }
1337
1338 srpc_client_rpc_t *
1339 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1340                         int nbulkiov, int bulklen,
1341                         void (*rpc_done)(srpc_client_rpc_t *),
1342                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1343 {
1344         srpc_client_rpc_t *rpc;
1345
1346         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1347                                    crpc_bulk.bk_iovs[nbulkiov]));
1348         if (rpc == NULL)
1349                 return NULL;
1350
1351         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1352                              bulklen, rpc_done, rpc_fini, priv);
1353         return rpc;
1354 }
1355
1356 /* called with rpc->crpc_lock held */
1357 void
1358 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1359 {
1360         LASSERT (why != 0);
1361
1362         if (rpc->crpc_aborted || /* already aborted */
1363             rpc->crpc_closed)    /* callback imminent */
1364                 return;
1365
1366         CDEBUG (D_NET,
1367                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1368                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1369                 swi_state2str(rpc->crpc_wi.swi_state), why);
1370
1371         rpc->crpc_aborted = 1;
1372         rpc->crpc_status  = why;
1373         swi_schedule_workitem(&rpc->crpc_wi);
1374         return;
1375 }
1376
1377 /* called with rpc->crpc_lock held */
1378 void
1379 srpc_post_rpc (srpc_client_rpc_t *rpc)
1380 {
1381         LASSERT (!rpc->crpc_aborted);
1382         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1383
1384         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1385                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1386                 rpc->crpc_timeout);
1387
1388         srpc_add_client_rpc_timer(rpc);
1389         swi_schedule_workitem(&rpc->crpc_wi);
1390         return;
1391 }
1392
1393
1394 int
1395 srpc_send_reply(struct srpc_server_rpc *rpc)
1396 {
1397         srpc_event_t            *ev = &rpc->srpc_ev;
1398         struct srpc_msg         *msg = &rpc->srpc_replymsg;
1399         struct srpc_buffer      *buffer = rpc->srpc_reqstbuf;
1400         struct srpc_service_cd  *scd = rpc->srpc_scd;
1401         struct srpc_service     *sv = scd->scd_svc;
1402         __u64                   rpyid;
1403         int                     rc;
1404
1405         LASSERT(buffer != NULL);
1406         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1407
1408         spin_lock(&scd->scd_lock);
1409
1410         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1411                 /* Repost buffer before replying since test client
1412                  * might send me another RPC once it gets the reply */
1413                 if (srpc_service_post_buffer(scd, buffer) != 0)
1414                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1415                 rpc->srpc_reqstbuf = NULL;
1416         }
1417
1418         spin_unlock(&scd->scd_lock);
1419
1420         ev->ev_fired = 0;
1421         ev->ev_data  = rpc;
1422         ev->ev_type  = SRPC_REPLY_SENT;
1423
1424         msg->msg_magic   = SRPC_MSG_MAGIC;
1425         msg->msg_version = SRPC_MSG_VERSION;
1426         msg->msg_type    = srpc_service2reply(sv->sv_id);
1427
1428         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1429                                    sizeof(*msg), LNET_MD_OP_PUT,
1430                                    rpc->srpc_peer, rpc->srpc_self,
1431                                    &rpc->srpc_replymdh, ev);
1432         if (rc != 0)
1433                 ev->ev_fired = 1;  /* no more event expected */
1434         return rc;
1435 }
1436
1437 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1438 void
1439 srpc_lnet_ev_handler(lnet_event_t *ev)
1440 {
1441         struct srpc_service_cd  *scd;
1442         srpc_event_t      *rpcev = ev->md.user_ptr;
1443         srpc_client_rpc_t *crpc;
1444         srpc_server_rpc_t *srpc;
1445         srpc_buffer_t     *buffer;
1446         srpc_service_t    *sv;
1447         srpc_msg_t        *msg;
1448         srpc_msg_type_t    type;
1449
1450         LASSERT (!cfs_in_interrupt());
1451
1452         if (ev->status != 0) {
1453                 spin_lock(&srpc_data.rpc_glock);
1454                 srpc_data.rpc_counters.errors++;
1455                 spin_unlock(&srpc_data.rpc_glock);
1456         }
1457
1458         rpcev->ev_lnet = ev->type;
1459
1460         switch (rpcev->ev_type) {
1461         default:
1462                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1463                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1464                 LBUG ();
1465         case SRPC_REQUEST_SENT:
1466                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1467                         spin_lock(&srpc_data.rpc_glock);
1468                         srpc_data.rpc_counters.rpcs_sent++;
1469                         spin_unlock(&srpc_data.rpc_glock);
1470                 }
1471         case SRPC_REPLY_RCVD:
1472         case SRPC_BULK_REQ_RCVD:
1473                 crpc = rpcev->ev_data;
1474
1475                 if (rpcev != &crpc->crpc_reqstev &&
1476                     rpcev != &crpc->crpc_replyev &&
1477                     rpcev != &crpc->crpc_bulkev) {
1478                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1479                                rpcev, crpc, &crpc->crpc_reqstev,
1480                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1481                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1482                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1483                         LBUG ();
1484                 }
1485
1486                 spin_lock(&crpc->crpc_lock);
1487
1488                 LASSERT(rpcev->ev_fired == 0);
1489                 rpcev->ev_fired  = 1;
1490                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1491                                                 -EINTR : ev->status;
1492                 swi_schedule_workitem(&crpc->crpc_wi);
1493
1494                 spin_unlock(&crpc->crpc_lock);
1495                 break;
1496
1497         case SRPC_REQUEST_RCVD:
1498                 scd = rpcev->ev_data;
1499                 sv = scd->scd_svc;
1500
1501                 LASSERT(rpcev == &scd->scd_ev);
1502
1503                 spin_lock(&scd->scd_lock);
1504
1505                 LASSERT (ev->unlinked);
1506                 LASSERT (ev->type == LNET_EVENT_PUT ||
1507                          ev->type == LNET_EVENT_UNLINK);
1508                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1509                          sv->sv_shuttingdown);
1510
1511                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1512                 buffer->buf_peer = ev->initiator;
1513                 buffer->buf_self = ev->target.nid;
1514
1515                 LASSERT(scd->scd_buf_nposted > 0);
1516                 scd->scd_buf_nposted--;
1517
1518                 if (sv->sv_shuttingdown) {
1519                         /* Leave buffer on scd->scd_buf_nposted since
1520                          * srpc_finish_service needs to traverse it. */
1521                         spin_unlock(&scd->scd_lock);
1522                         break;
1523                 }
1524
1525                 if (scd->scd_buf_err_stamp != 0 &&
1526                     scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1527                         /* re-enable adding buffer */
1528                         scd->scd_buf_err_stamp = 0;
1529                         scd->scd_buf_err = 0;
1530                 }
1531
1532                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1533                     scd->scd_buf_adjust == 0 &&
1534                     scd->scd_buf_nposted < scd->scd_buf_low) {
1535                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1536                                                   SFW_TEST_WI_MIN);
1537                         swi_schedule_workitem(&scd->scd_buf_wi);
1538                 }
1539
1540                 cfs_list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1541                 msg = &buffer->buf_msg;
1542                 type = srpc_service2request(sv->sv_id);
1543
1544                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1545                     (msg->msg_type != type &&
1546                      msg->msg_type != __swab32(type)) ||
1547                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1548                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1549                         CERROR ("Dropping RPC (%s) from %s: "
1550                                 "status %d mlength %d type %u magic %u.\n",
1551                                 sv->sv_name, libcfs_id2str(ev->initiator),
1552                                 ev->status, ev->mlength,
1553                                 msg->msg_type, msg->msg_magic);
1554
1555                         /* NB can't call srpc_service_recycle_buffer here since
1556                          * it may call LNetM[DE]Attach. The invalid magic tells
1557                          * srpc_handle_rpc to drop this RPC */
1558                         msg->msg_magic = 0;
1559                 }
1560
1561                 if (!cfs_list_empty(&scd->scd_rpc_free)) {
1562                         srpc = cfs_list_entry(scd->scd_rpc_free.next,
1563                                               struct srpc_server_rpc,
1564                                               srpc_list);
1565                         cfs_list_del(&srpc->srpc_list);
1566
1567                         srpc_init_server_rpc(srpc, scd, buffer);
1568                         cfs_list_add_tail(&srpc->srpc_list,
1569                                           &scd->scd_rpc_active);
1570                         swi_schedule_workitem(&srpc->srpc_wi);
1571                 } else {
1572                         cfs_list_add_tail(&buffer->buf_list,
1573                                           &scd->scd_buf_blocked);
1574                 }
1575
1576                 spin_unlock(&scd->scd_lock);
1577
1578                 spin_lock(&srpc_data.rpc_glock);
1579                 srpc_data.rpc_counters.rpcs_rcvd++;
1580                 spin_unlock(&srpc_data.rpc_glock);
1581                 break;
1582
1583         case SRPC_BULK_GET_RPLD:
1584                 LASSERT (ev->type == LNET_EVENT_SEND ||
1585                          ev->type == LNET_EVENT_REPLY ||
1586                          ev->type == LNET_EVENT_UNLINK);
1587
1588                 if (!ev->unlinked)
1589                         break; /* wait for final event */
1590
1591         case SRPC_BULK_PUT_SENT:
1592                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1593                         spin_lock(&srpc_data.rpc_glock);
1594
1595                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1596                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1597                         else
1598                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1599
1600                         spin_unlock(&srpc_data.rpc_glock);
1601                 }
1602         case SRPC_REPLY_SENT:
1603                 srpc = rpcev->ev_data;
1604                 scd  = srpc->srpc_scd;
1605
1606                 LASSERT(rpcev == &srpc->srpc_ev);
1607
1608                 spin_lock(&scd->scd_lock);
1609
1610                 rpcev->ev_fired  = 1;
1611                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1612                                    -EINTR : ev->status;
1613                 swi_schedule_workitem(&srpc->srpc_wi);
1614
1615                 spin_unlock(&scd->scd_lock);
1616                 break;
1617         }
1618 }
1619
1620 #ifndef __KERNEL__
1621
1622 int
1623 srpc_check_event (int timeout)
1624 {
1625         lnet_event_t ev;
1626         int          rc;
1627         int          i;
1628
1629         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1630                         timeout * 1000, &ev, &i);
1631         if (rc == 0) return 0;
1632
1633         LASSERT (rc == -EOVERFLOW || rc == 1);
1634
1635         /* We can't affort to miss any events... */
1636         if (rc == -EOVERFLOW) {
1637                 CERROR ("Dropped an event!!!\n");
1638                 abort();
1639         }
1640
1641         srpc_lnet_ev_handler(&ev);
1642         return 1;
1643 }
1644
1645 #endif
1646
1647 int
1648 srpc_startup (void)
1649 {
1650         int rc;
1651
1652         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1653         spin_lock_init(&srpc_data.rpc_glock);
1654
1655         /* 1 second pause to avoid timestamp reuse */
1656         cfs_pause(cfs_time_seconds(1));
1657         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1658
1659         srpc_data.rpc_state = SRPC_STATE_NONE;
1660
1661 #ifdef __KERNEL__
1662         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1663 #else
1664         if (the_lnet.ln_server_mode_flag)
1665                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1666         else
1667                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1668 #endif
1669         if (rc < 0) {
1670                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1671                 return rc;
1672         }
1673
1674         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1675
1676         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1677 #ifdef __KERNEL__
1678         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1679 #else
1680         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1681 #endif
1682         if (rc != 0) {
1683                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1684                 goto bail;
1685         }
1686
1687         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1688         LASSERT(rc == 0);
1689         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1690         LASSERT(rc == 0);
1691
1692         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1693
1694         rc = stt_startup();
1695
1696 bail:
1697         if (rc != 0)
1698                 srpc_shutdown();
1699         else
1700                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1701
1702         return rc;
1703 }
1704
1705 void
1706 srpc_shutdown (void)
1707 {
1708         int i;
1709         int rc;
1710         int state;
1711
1712         state = srpc_data.rpc_state;
1713         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1714
1715         switch (state) {
1716         default:
1717                 LBUG ();
1718         case SRPC_STATE_RUNNING:
1719                 spin_lock(&srpc_data.rpc_glock);
1720
1721                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1722                         srpc_service_t *sv = srpc_data.rpc_services[i];
1723
1724                         LASSERTF (sv == NULL,
1725                                   "service not empty: id %d, name %s\n",
1726                                   i, sv->sv_name);
1727                 }
1728
1729                 spin_unlock(&srpc_data.rpc_glock);
1730
1731                 stt_shutdown();
1732
1733         case SRPC_STATE_EQ_INIT:
1734                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1735                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1736                 LASSERT (rc == 0);
1737                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1738                 LASSERT (rc == 0); /* the EQ should have no user by now */
1739
1740         case SRPC_STATE_NI_INIT:
1741                 LNetNIFini();
1742         }
1743
1744         return;
1745 }