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, 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 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 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 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 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 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 int
 505 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 506 {
 507         struct srpc_service     *sv = scd->scd_svc;
 508         struct srpc_msg         *msg = &buf->buf_msg;
 509         int                     rc;
 510
 511         LNetInvalidateHandle(&buf->buf_mdh);
 512         list_add(&buf->buf_list, &scd->scd_buf_posted);
 513         scd->scd_buf_nposted++;
 514         spin_unlock(&scd->scd_lock);
 515
 516         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 517                                       !srpc_serv_is_framework(sv),
 518                                       msg, sizeof(*msg), &buf->buf_mdh,
 519                                       &scd->scd_ev);
 520
 521         /* At this point, a RPC (new or delayed) may have arrived in
 522          * msg and its event handler has been called. So we must add
 523          * buf to scd_buf_posted _before_ dropping scd_lock */
 524
 525         spin_lock(&scd->scd_lock);
 526
 527         if (rc == 0) {
 528                 if (!sv->sv_shuttingdown)
 529                         return 0;
 530
 531                 spin_unlock(&scd->scd_lock);
 532                 /* srpc_shutdown_service might have tried to unlink me
 533                  * when my buf_mdh was still invalid */
 534                 LNetMDUnlink(buf->buf_mdh);
 535                 spin_lock(&scd->scd_lock);
 536                 return 0;
 537         }
 538
 539         scd->scd_buf_nposted--;
 540         if (sv->sv_shuttingdown)
 541                 return rc; /* don't allow to change scd_buf_posted */
 542
 543         list_del(&buf->buf_list);
 544         spin_unlock(&scd->scd_lock);
 545
 546         LIBCFS_FREE(buf, sizeof(*buf));
 547
 548         spin_lock(&scd->scd_lock);
 549         return rc;
 550 }
 551
 552 int
 553 srpc_add_buffer(struct swi_workitem *wi)
 554 {
 555         struct srpc_service_cd  *scd = wi->swi_workitem.wi_data;
 556         struct srpc_buffer      *buf;
 557         int                     rc = 0;
 558
 559         /* it's called by workitem scheduler threads, these threads
 560          * should have been set CPT affinity, so buffers will be posted
 561          * on CPT local list of Portal */
 562         spin_lock(&scd->scd_lock);
 563
 564         while (scd->scd_buf_adjust > 0 &&
 565                !scd->scd_svc->sv_shuttingdown) {
 566                 scd->scd_buf_adjust--; /* consume it */
 567                 scd->scd_buf_posting++;
 568
 569                 spin_unlock(&scd->scd_lock);
 570
 571                 LIBCFS_ALLOC(buf, sizeof(*buf));
 572                 if (buf == NULL) {
 573                         CERROR("Failed to add new buf to service: %s\n",
 574                                scd->scd_svc->sv_name);
 575                         spin_lock(&scd->scd_lock);
 576                         rc = -ENOMEM;
 577                         break;
 578                 }
 579
 580                 spin_lock(&scd->scd_lock);
 581                 if (scd->scd_svc->sv_shuttingdown) {
 582                         spin_unlock(&scd->scd_lock);
 583                         LIBCFS_FREE(buf, sizeof(*buf));
 584
 585                         spin_lock(&scd->scd_lock);
 586                         rc = -ESHUTDOWN;
 587                         break;
 588                 }
 589
 590                 rc = srpc_service_post_buffer(scd, buf);
 591                 if (rc != 0)
 592                         break; /* buf has been freed inside */
 593
 594                 LASSERT(scd->scd_buf_posting > 0);
 595                 scd->scd_buf_posting--;
 596                 scd->scd_buf_total++;
 597                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 598         }
 599
 600         if (rc != 0) {
 601                 scd->scd_buf_err_stamp = cfs_time_current_sec();
 602                 scd->scd_buf_err = rc;
 603
 604                 LASSERT(scd->scd_buf_posting > 0);
 605                 scd->scd_buf_posting--;
 606         }
 607
 608         spin_unlock(&scd->scd_lock);
 609         return 0;
 610 }
 611
 612 int
 613 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 614 {
 615         struct srpc_service_cd  *scd;
 616         int                     rc = 0;
 617         int                     i;
 618
 619         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 620
 621         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 622                 spin_lock(&scd->scd_lock);
 623
 624                 scd->scd_buf_err = 0;
 625                 scd->scd_buf_err_stamp = 0;
 626                 scd->scd_buf_posting = 0;
 627                 scd->scd_buf_adjust = nbuffer;
 628                 /* start to post buffers */
 629                 swi_schedule_workitem(&scd->scd_buf_wi);
 630                 spin_unlock(&scd->scd_lock);
 631
 632                 /* framework service only post buffer for one partition  */
 633                 if (srpc_serv_is_framework(sv))
 634                         break;
 635         }
 636
 637         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 638                 spin_lock(&scd->scd_lock);
 639                 /*
 640                  * NB: srpc_service_add_buffers() can be called inside
 641                  * thread context of lst_sched_serial, and we don't normally
 642                  * allow to sleep inside thread context of WI scheduler
 643                  * because it will block current scheduler thread from doing
 644                  * anything else, even worse, it could deadlock if it's
 645                  * waiting on result from another WI of the same scheduler.
 646                  * However, it's safe at here because scd_buf_wi is scheduled
 647                  * by thread in a different WI scheduler (lst_sched_test),
 648                  * so we don't have any risk of deadlock, though this could
 649                  * block all WIs pending on lst_sched_serial for a moment
 650                  * which is not good but not fatal.
 651                  */
 652                 lst_wait_until(scd->scd_buf_err != 0 ||
 653                                (scd->scd_buf_adjust == 0 &&
 654                                 scd->scd_buf_posting == 0),
 655                                scd->scd_lock, "waiting for adding buffer\n");
 656
 657                 if (scd->scd_buf_err != 0 && rc == 0)
 658                         rc = scd->scd_buf_err;
 659
 660                 spin_unlock(&scd->scd_lock);
 661         }
 662
 663         return rc;
 664 }
 665
 666 void
 667 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 668 {
 669         struct srpc_service_cd  *scd;
 670         int                     num;
 671         int                     i;
 672
 673         LASSERT(!sv->sv_shuttingdown);
 674
 675         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 676                 spin_lock(&scd->scd_lock);
 677
 678                 num = scd->scd_buf_total + scd->scd_buf_posting;
 679                 scd->scd_buf_adjust -= min(nbuffer, num);
 680
 681                 spin_unlock(&scd->scd_lock);
 682         }
 683 }
 684
 685 /* returns 1 if sv has finished, otherwise 0 */
 686 int
 687 srpc_finish_service(struct srpc_service *sv)
 688 {
 689         struct srpc_service_cd  *scd;
 690         struct srpc_server_rpc  *rpc;
 691         int                     i;
 692
 693         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 694
 695         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 696                 spin_lock(&scd->scd_lock);
 697                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 698                         spin_unlock(&scd->scd_lock);
 699                         return 0;
 700                 }
 701
 702                 if (scd->scd_buf_nposted > 0) {
 703                         CDEBUG(D_NET, "waiting for %d posted buffers to "
 704                                "unlink\n", scd->scd_buf_nposted);
 705                         spin_unlock(&scd->scd_lock);
 706                         return 0;
 707                 }
 708
 709                 if (list_empty(&scd->scd_rpc_active)) {
 710                         spin_unlock(&scd->scd_lock);
 711                         continue;
 712                 }
 713
 714                 rpc = list_entry(scd->scd_rpc_active.next,
 715                                      struct srpc_server_rpc, srpc_list);
 716                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
 717                         "wi %s scheduled %d running %d, "
 718                         "ev fired %d type %d status %d lnet %d\n",
 719                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 720                         swi_state2str(rpc->srpc_wi.swi_state),
 721                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 722                         rpc->srpc_wi.swi_workitem.wi_running,
 723                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 724                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 725                 spin_unlock(&scd->scd_lock);
 726                 return 0;
 727         }
 728
 729         /* no lock needed from now on */
 730         srpc_service_fini(sv);
 731         return 1;
 732 }
 733
 734 /* called with sv->sv_lock held */
 735 void
 736 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
 737 {
 738         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 739                 if (srpc_service_post_buffer(scd, buf) != 0) {
 740                         CWARN("Failed to post %s buffer\n",
 741                               scd->scd_svc->sv_name);
 742                 }
 743                 return;
 744         }
 745
 746         /* service is shutting down, or we want to recycle some buffers */
 747         scd->scd_buf_total--;
 748
 749         if (scd->scd_buf_adjust < 0) {
 750                 scd->scd_buf_adjust++;
 751                 if (scd->scd_buf_adjust < 0 &&
 752                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 753                         CDEBUG(D_INFO,
 754                                "Try to recyle %d buffers but nothing left\n",
 755                                scd->scd_buf_adjust);
 756                         scd->scd_buf_adjust = 0;
 757                 }
 758         }
 759
 760         spin_unlock(&scd->scd_lock);
 761         LIBCFS_FREE(buf, sizeof(*buf));
 762         spin_lock(&scd->scd_lock);
 763 }
 764
 765 void
 766 srpc_abort_service(struct srpc_service *sv)
 767 {
 768         struct srpc_service_cd  *scd;
 769         struct srpc_server_rpc  *rpc;
 770         int                     i;
 771
 772         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 773                sv->sv_id, sv->sv_name);
 774
 775         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 776                 spin_lock(&scd->scd_lock);
 777
 778                 /* schedule in-flight RPCs to notice the abort, NB:
 779                  * racing with incoming RPCs; complete fix should make test
 780                  * RPCs carry session ID in its headers */
 781                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 782                         rpc->srpc_aborted = 1;
 783                         swi_schedule_workitem(&rpc->srpc_wi);
 784                 }
 785
 786                 spin_unlock(&scd->scd_lock);
 787         }
 788 }
 789
 790 void
 791 srpc_shutdown_service(srpc_service_t *sv)
 792 {
 793         struct srpc_service_cd  *scd;
 794         struct srpc_server_rpc  *rpc;
 795         srpc_buffer_t           *buf;
 796         int                     i;
 797
 798         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 799                sv->sv_id, sv->sv_name);
 800
 801         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 802                 spin_lock(&scd->scd_lock);
 803
 804         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 805
 806         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 807                 spin_unlock(&scd->scd_lock);
 808
 809         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 810                 spin_lock(&scd->scd_lock);
 811
 812                 /* schedule in-flight RPCs to notice the shutdown */
 813                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 814                         swi_schedule_workitem(&rpc->srpc_wi);
 815
 816                 spin_unlock(&scd->scd_lock);
 817
 818                 /* OK to traverse scd_buf_posted without lock, since no one
 819                  * touches scd_buf_posted now */
 820                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 821                         LNetMDUnlink(buf->buf_mdh);
 822         }
 823 }
 824
 825 int
 826 srpc_send_request (srpc_client_rpc_t *rpc)
 827 {
 828         srpc_event_t *ev = &rpc->crpc_reqstev;
 829         int           rc;
 830
 831         ev->ev_fired = 0;
 832         ev->ev_data  = rpc;
 833         ev->ev_type  = SRPC_REQUEST_SENT;
 834
 835         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 836                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 837                                      &rpc->crpc_reqstmdh, ev);
 838         if (rc != 0) {
 839                 LASSERT (rc == -ENOMEM);
 840                 ev->ev_fired = 1;  /* no more event expected */
 841         }
 842         return rc;
 843 }
 844
 845 int
 846 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 847 {
 848         srpc_event_t *ev = &rpc->crpc_replyev;
 849         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 850         int           rc;
 851
 852         ev->ev_fired = 0;
 853         ev->ev_data  = rpc;
 854         ev->ev_type  = SRPC_REPLY_RCVD;
 855
 856         *id = srpc_next_id();
 857
 858         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 859                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 860                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 861                                     &rpc->crpc_replymdh, ev);
 862         if (rc != 0) {
 863                 LASSERT (rc == -ENOMEM);
 864                 ev->ev_fired = 1;  /* no more event expected */
 865         }
 866         return rc;
 867 }
 868
 869 int
 870 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 871 {
 872         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 873         srpc_event_t *ev = &rpc->crpc_bulkev;
 874         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 875         int           rc;
 876         int           opt;
 877
 878         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 879
 880         if (bk->bk_niov == 0) return 0; /* nothing to do */
 881
 882         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 883 #ifdef __KERNEL__
 884         opt |= LNET_MD_KIOV;
 885 #else
 886         opt |= LNET_MD_IOVEC;
 887 #endif
 888
 889         ev->ev_fired = 0;
 890         ev->ev_data  = rpc;
 891         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 892
 893         *id = srpc_next_id();
 894
 895         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 896                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 897                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 898         if (rc != 0) {
 899                 LASSERT (rc == -ENOMEM);
 900                 ev->ev_fired = 1;  /* no more event expected */
 901         }
 902         return rc;
 903 }
 904
 905 int
 906 srpc_do_bulk (srpc_server_rpc_t *rpc)
 907 {
 908         srpc_event_t  *ev = &rpc->srpc_ev;
 909         srpc_bulk_t   *bk = rpc->srpc_bulk;
 910         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 911         int            rc;
 912         int            opt;
 913
 914         LASSERT (bk != NULL);
 915
 916         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 917 #ifdef __KERNEL__
 918         opt |= LNET_MD_KIOV;
 919 #else
 920         opt |= LNET_MD_IOVEC;
 921 #endif
 922
 923         ev->ev_fired = 0;
 924         ev->ev_data  = rpc;
 925         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 926
 927         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 928                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 929                                    rpc->srpc_peer, rpc->srpc_self,
 930                                    &bk->bk_mdh, ev);
 931         if (rc != 0)
 932                 ev->ev_fired = 1;  /* no more event expected */
 933         return rc;
 934 }
 935
 936 /* only called from srpc_handle_rpc */
 937 void
 938 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
 939 {
 940         struct srpc_service_cd  *scd = rpc->srpc_scd;
 941         struct srpc_service     *sv  = scd->scd_svc;
 942         srpc_buffer_t           *buffer;
 943
 944         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 945
 946         rpc->srpc_status = status;
 947
 948         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 949                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 950                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 951                 swi_state2str(rpc->srpc_wi.swi_state), status);
 952
 953         if (status != 0) {
 954                 spin_lock(&srpc_data.rpc_glock);
 955                 srpc_data.rpc_counters.rpcs_dropped++;
 956                 spin_unlock(&srpc_data.rpc_glock);
 957         }
 958
 959         if (rpc->srpc_done != NULL)
 960                 (*rpc->srpc_done) (rpc);
 961         LASSERT(rpc->srpc_bulk == NULL);
 962
 963         spin_lock(&scd->scd_lock);
 964
 965         if (rpc->srpc_reqstbuf != NULL) {
 966                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 967                  * sv won't go away for scd_rpc_active must not be empty */
 968                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 969                 rpc->srpc_reqstbuf = NULL;
 970         }
 971
 972         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 973
 974         /*
 975          * No one can schedule me now since:
 976          * - I'm not on scd_rpc_active.
 977          * - all LNet events have been fired.
 978          * Cancel pending schedules and prevent future schedule attempts:
 979          */
 980         LASSERT(rpc->srpc_ev.ev_fired);
 981         swi_exit_workitem(&rpc->srpc_wi);
 982
 983         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 984                 buffer = list_entry(scd->scd_buf_blocked.next,
 985                                         srpc_buffer_t, buf_list);
 986                 list_del(&buffer->buf_list);
 987
 988                 srpc_init_server_rpc(rpc, scd, buffer);
 989                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 990                 swi_schedule_workitem(&rpc->srpc_wi);
 991         } else {
 992                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 993         }
 994
 995         spin_unlock(&scd->scd_lock);
 996         return;
 997 }
 998
 999 /* handles an incoming RPC */
1000 int
1001 srpc_handle_rpc(swi_workitem_t *wi)
1002 {
1003         struct srpc_server_rpc  *rpc = wi->swi_workitem.wi_data;
1004         struct srpc_service_cd  *scd = rpc->srpc_scd;
1005         struct srpc_service     *sv = scd->scd_svc;
1006         srpc_event_t            *ev = &rpc->srpc_ev;
1007         int                     rc = 0;
1008
1009         LASSERT(wi == &rpc->srpc_wi);
1010
1011         spin_lock(&scd->scd_lock);
1012
1013         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
1014                 spin_unlock(&scd->scd_lock);
1015
1016                 if (rpc->srpc_bulk != NULL)
1017                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
1018                 LNetMDUnlink(rpc->srpc_replymdh);
1019
1020                 if (ev->ev_fired) { /* no more event, OK to finish */
1021                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
1022                         return 1;
1023                 }
1024                 return 0;
1025         }
1026
1027         spin_unlock(&scd->scd_lock);
1028
1029         switch (wi->swi_state) {
1030         default:
1031                 LBUG ();
1032         case SWI_STATE_NEWBORN: {
1033                 srpc_msg_t           *msg;
1034                 srpc_generic_reply_t *reply;
1035
1036                 msg = &rpc->srpc_reqstbuf->buf_msg;
1037                 reply = &rpc->srpc_replymsg.msg_body.reply;
1038
1039                 if (msg->msg_magic == 0) {
1040                         /* moaned already in srpc_lnet_ev_handler */
1041                         srpc_server_rpc_done(rpc, EBADMSG);
1042                         return 1;
1043                 }
1044
1045                 srpc_unpack_msg_hdr(msg);
1046                 if (msg->msg_version != SRPC_MSG_VERSION) {
1047                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1048                               msg->msg_version, SRPC_MSG_VERSION,
1049                               libcfs_id2str(rpc->srpc_peer));
1050                         reply->status = EPROTO;
1051                         /* drop through and send reply */
1052                 } else {
1053                         reply->status = 0;
1054                         rc = (*sv->sv_handler)(rpc);
1055                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1056                         if (rc != 0) {
1057                                 srpc_server_rpc_done(rpc, rc);
1058                                 return 1;
1059                         }
1060                 }
1061
1062                 wi->swi_state = SWI_STATE_BULK_STARTED;
1063
1064                 if (rpc->srpc_bulk != NULL) {
1065                         rc = srpc_do_bulk(rpc);
1066                         if (rc == 0)
1067                                 return 0; /* wait for bulk */
1068
1069                         LASSERT (ev->ev_fired);
1070                         ev->ev_status = rc;
1071                 }
1072         }
1073         case SWI_STATE_BULK_STARTED:
1074                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1075
1076                 if (rpc->srpc_bulk != NULL) {
1077                         rc = ev->ev_status;
1078
1079                         if (sv->sv_bulk_ready != NULL)
1080                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1081
1082                         if (rc != 0) {
1083                                 srpc_server_rpc_done(rpc, rc);
1084                                 return 1;
1085                         }
1086                 }
1087
1088                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1089                 rc = srpc_send_reply(rpc);
1090                 if (rc == 0)
1091                         return 0; /* wait for reply */
1092                 srpc_server_rpc_done(rpc, rc);
1093                 return 1;
1094
1095         case SWI_STATE_REPLY_SUBMITTED:
1096                 if (!ev->ev_fired) {
1097                         CERROR("RPC %p: bulk %p, service %d\n",
1098                                rpc, rpc->srpc_bulk, sv->sv_id);
1099                         CERROR("Event: status %d, type %d, lnet %d\n",
1100                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1101                         LASSERT (ev->ev_fired);
1102                 }
1103
1104                 wi->swi_state = SWI_STATE_DONE;
1105                 srpc_server_rpc_done(rpc, ev->ev_status);
1106                 return 1;
1107         }
1108
1109         return 0;
1110 }
1111
1112 void
1113 srpc_client_rpc_expired (void *data)
1114 {
1115         srpc_client_rpc_t *rpc = data;
1116
1117         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1118                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1119                rpc->crpc_timeout);
1120
1121         spin_lock(&rpc->crpc_lock);
1122
1123         rpc->crpc_timeout = 0;
1124         srpc_abort_rpc(rpc, -ETIMEDOUT);
1125
1126         spin_unlock(&rpc->crpc_lock);
1127
1128         spin_lock(&srpc_data.rpc_glock);
1129         srpc_data.rpc_counters.rpcs_expired++;
1130         spin_unlock(&srpc_data.rpc_glock);
1131 }
1132
1133 inline void
1134 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1135 {
1136         stt_timer_t *timer = &rpc->crpc_timer;
1137
1138         if (rpc->crpc_timeout == 0)
1139                 return;
1140
1141         INIT_LIST_HEAD(&timer->stt_list);
1142         timer->stt_data    = rpc;
1143         timer->stt_func    = srpc_client_rpc_expired;
1144         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1145                                           cfs_time_current_sec());
1146         stt_add_timer(timer);
1147         return;
1148 }
1149
1150 /*
1151  * Called with rpc->crpc_lock held.
1152  *
1153  * Upon exit the RPC expiry timer is not queued and the handler is not
1154  * running on any CPU. */
1155 void
1156 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1157 {
1158         /* timer not planted or already exploded */
1159         if (rpc->crpc_timeout == 0)
1160                 return;
1161
1162         /* timer sucessfully defused */
1163         if (stt_del_timer(&rpc->crpc_timer))
1164                 return;
1165
1166 #ifdef __KERNEL__
1167         /* timer detonated, wait for it to explode */
1168         while (rpc->crpc_timeout != 0) {
1169                 spin_unlock(&rpc->crpc_lock);
1170
1171                 schedule();
1172
1173                 spin_lock(&rpc->crpc_lock);
1174         }
1175 #else
1176         LBUG(); /* impossible in single-threaded runtime */
1177 #endif
1178 }
1179
1180 void
1181 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1182 {
1183         swi_workitem_t *wi = &rpc->crpc_wi;
1184
1185         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1186
1187         spin_lock(&rpc->crpc_lock);
1188
1189         rpc->crpc_closed = 1;
1190         if (rpc->crpc_status == 0)
1191                 rpc->crpc_status = status;
1192
1193         srpc_del_client_rpc_timer(rpc);
1194
1195         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1196                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1197                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1198                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1199
1200         /*
1201          * No one can schedule me now since:
1202          * - RPC timer has been defused.
1203          * - all LNet events have been fired.
1204          * - crpc_closed has been set, preventing srpc_abort_rpc from
1205          *   scheduling me.
1206          * Cancel pending schedules and prevent future schedule attempts:
1207          */
1208         LASSERT (!srpc_event_pending(rpc));
1209         swi_exit_workitem(wi);
1210
1211         spin_unlock(&rpc->crpc_lock);
1212
1213         (*rpc->crpc_done)(rpc);
1214         return;
1215 }
1216
1217 /* sends an outgoing RPC */
1218 int
1219 srpc_send_rpc (swi_workitem_t *wi)
1220 {
1221         int                rc = 0;
1222         srpc_client_rpc_t *rpc;
1223         srpc_msg_t        *reply;
1224         int                do_bulk;
1225
1226         LASSERT(wi != NULL);
1227
1228         rpc = wi->swi_workitem.wi_data;
1229
1230         LASSERT (rpc != NULL);
1231         LASSERT (wi == &rpc->crpc_wi);
1232
1233         reply = &rpc->crpc_replymsg;
1234         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1235
1236         spin_lock(&rpc->crpc_lock);
1237
1238         if (rpc->crpc_aborted) {
1239                 spin_unlock(&rpc->crpc_lock);
1240                 goto abort;
1241         }
1242
1243         spin_unlock(&rpc->crpc_lock);
1244
1245         switch (wi->swi_state) {
1246         default:
1247                 LBUG ();
1248         case SWI_STATE_NEWBORN:
1249                 LASSERT (!srpc_event_pending(rpc));
1250
1251                 rc = srpc_prepare_reply(rpc);
1252                 if (rc != 0) {
1253                         srpc_client_rpc_done(rpc, rc);
1254                         return 1;
1255                 }
1256
1257                 rc = srpc_prepare_bulk(rpc);
1258                 if (rc != 0) break;
1259
1260                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1261                 rc = srpc_send_request(rpc);
1262                 break;
1263
1264         case SWI_STATE_REQUEST_SUBMITTED:
1265                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1266                  * order; however, they're processed in a strict order:
1267                  * rqt, rpy, and bulk. */
1268                 if (!rpc->crpc_reqstev.ev_fired) break;
1269
1270                 rc = rpc->crpc_reqstev.ev_status;
1271                 if (rc != 0) break;
1272
1273                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1274                 /* perhaps more events, fall thru */
1275         case SWI_STATE_REQUEST_SENT: {
1276                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1277
1278                 if (!rpc->crpc_replyev.ev_fired) break;
1279
1280                 rc = rpc->crpc_replyev.ev_status;
1281                 if (rc != 0) break;
1282
1283                 srpc_unpack_msg_hdr(reply);
1284                 if (reply->msg_type != type ||
1285                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1286                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1287                         CWARN ("Bad message from %s: type %u (%d expected),"
1288                                " magic %u (%d expected).\n",
1289                                libcfs_id2str(rpc->crpc_dest),
1290                                reply->msg_type, type,
1291                                reply->msg_magic, SRPC_MSG_MAGIC);
1292                         rc = -EBADMSG;
1293                         break;
1294                 }
1295
1296                 if (do_bulk && reply->msg_body.reply.status != 0) {
1297                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1298                                "case peer didn't initiate bulk transfer\n",
1299                                reply->msg_body.reply.status,
1300                                libcfs_id2str(rpc->crpc_dest));
1301                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1302                 }
1303
1304                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1305         }
1306         case SWI_STATE_REPLY_RECEIVED:
1307                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1308
1309                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1310
1311                 /* Bulk buffer was unlinked due to remote error. Clear error
1312                  * since reply buffer still contains valid data.
1313                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1314                  * remote error. */
1315                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1316                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1317                         rc = 0;
1318
1319                 wi->swi_state = SWI_STATE_DONE;
1320                 srpc_client_rpc_done(rpc, rc);
1321                 return 1;
1322         }
1323
1324         if (rc != 0) {
1325                 spin_lock(&rpc->crpc_lock);
1326                 srpc_abort_rpc(rpc, rc);
1327                 spin_unlock(&rpc->crpc_lock);
1328         }
1329
1330 abort:
1331         if (rpc->crpc_aborted) {
1332                 LNetMDUnlink(rpc->crpc_reqstmdh);
1333                 LNetMDUnlink(rpc->crpc_replymdh);
1334                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1335
1336                 if (!srpc_event_pending(rpc)) {
1337                         srpc_client_rpc_done(rpc, -EINTR);
1338                         return 1;
1339                 }
1340         }
1341         return 0;
1342 }
1343
1344 srpc_client_rpc_t *
1345 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1346                         int nbulkiov, int bulklen,
1347                         void (*rpc_done)(srpc_client_rpc_t *),
1348                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1349 {
1350         srpc_client_rpc_t *rpc;
1351
1352         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1353                                    crpc_bulk.bk_iovs[nbulkiov]));
1354         if (rpc == NULL)
1355                 return NULL;
1356
1357         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1358                              bulklen, rpc_done, rpc_fini, priv);
1359         return rpc;
1360 }
1361
1362 /* called with rpc->crpc_lock held */
1363 void
1364 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1365 {
1366         LASSERT (why != 0);
1367
1368         if (rpc->crpc_aborted || /* already aborted */
1369             rpc->crpc_closed)    /* callback imminent */
1370                 return;
1371
1372         CDEBUG (D_NET,
1373                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1374                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1375                 swi_state2str(rpc->crpc_wi.swi_state), why);
1376
1377         rpc->crpc_aborted = 1;
1378         rpc->crpc_status  = why;
1379         swi_schedule_workitem(&rpc->crpc_wi);
1380         return;
1381 }
1382
1383 /* called with rpc->crpc_lock held */
1384 void
1385 srpc_post_rpc (srpc_client_rpc_t *rpc)
1386 {
1387         LASSERT (!rpc->crpc_aborted);
1388         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1389
1390         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1391                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1392                 rpc->crpc_timeout);
1393
1394         srpc_add_client_rpc_timer(rpc);
1395         swi_schedule_workitem(&rpc->crpc_wi);
1396         return;
1397 }
1398
1399
1400 int
1401 srpc_send_reply(struct srpc_server_rpc *rpc)
1402 {
1403         srpc_event_t            *ev = &rpc->srpc_ev;
1404         struct srpc_msg         *msg = &rpc->srpc_replymsg;
1405         struct srpc_buffer      *buffer = rpc->srpc_reqstbuf;
1406         struct srpc_service_cd  *scd = rpc->srpc_scd;
1407         struct srpc_service     *sv = scd->scd_svc;
1408         __u64                   rpyid;
1409         int                     rc;
1410
1411         LASSERT(buffer != NULL);
1412         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1413
1414         spin_lock(&scd->scd_lock);
1415
1416         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1417                 /* Repost buffer before replying since test client
1418                  * might send me another RPC once it gets the reply */
1419                 if (srpc_service_post_buffer(scd, buffer) != 0)
1420                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1421                 rpc->srpc_reqstbuf = NULL;
1422         }
1423
1424         spin_unlock(&scd->scd_lock);
1425
1426         ev->ev_fired = 0;
1427         ev->ev_data  = rpc;
1428         ev->ev_type  = SRPC_REPLY_SENT;
1429
1430         msg->msg_magic   = SRPC_MSG_MAGIC;
1431         msg->msg_version = SRPC_MSG_VERSION;
1432         msg->msg_type    = srpc_service2reply(sv->sv_id);
1433
1434         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1435                                    sizeof(*msg), LNET_MD_OP_PUT,
1436                                    rpc->srpc_peer, rpc->srpc_self,
1437                                    &rpc->srpc_replymdh, ev);
1438         if (rc != 0)
1439                 ev->ev_fired = 1;  /* no more event expected */
1440         return rc;
1441 }
1442
1443 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1444 void
1445 srpc_lnet_ev_handler(lnet_event_t *ev)
1446 {
1447         struct srpc_service_cd  *scd;
1448         srpc_event_t      *rpcev = ev->md.user_ptr;
1449         srpc_client_rpc_t *crpc;
1450         srpc_server_rpc_t *srpc;
1451         srpc_buffer_t     *buffer;
1452         srpc_service_t    *sv;
1453         srpc_msg_t        *msg;
1454         srpc_msg_type_t    type;
1455
1456         LASSERT (!in_interrupt());
1457
1458         if (ev->status != 0) {
1459                 spin_lock(&srpc_data.rpc_glock);
1460                 srpc_data.rpc_counters.errors++;
1461                 spin_unlock(&srpc_data.rpc_glock);
1462         }
1463
1464         rpcev->ev_lnet = ev->type;
1465
1466         switch (rpcev->ev_type) {
1467         default:
1468                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1469                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1470                 LBUG ();
1471         case SRPC_REQUEST_SENT:
1472                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1473                         spin_lock(&srpc_data.rpc_glock);
1474                         srpc_data.rpc_counters.rpcs_sent++;
1475                         spin_unlock(&srpc_data.rpc_glock);
1476                 }
1477         case SRPC_REPLY_RCVD:
1478         case SRPC_BULK_REQ_RCVD:
1479                 crpc = rpcev->ev_data;
1480
1481                 if (rpcev != &crpc->crpc_reqstev &&
1482                     rpcev != &crpc->crpc_replyev &&
1483                     rpcev != &crpc->crpc_bulkev) {
1484                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1485                                rpcev, crpc, &crpc->crpc_reqstev,
1486                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1487                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1488                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1489                         LBUG ();
1490                 }
1491
1492                 spin_lock(&crpc->crpc_lock);
1493
1494                 LASSERT(rpcev->ev_fired == 0);
1495                 rpcev->ev_fired  = 1;
1496                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1497                                                 -EINTR : ev->status;
1498                 swi_schedule_workitem(&crpc->crpc_wi);
1499
1500                 spin_unlock(&crpc->crpc_lock);
1501                 break;
1502
1503         case SRPC_REQUEST_RCVD:
1504                 scd = rpcev->ev_data;
1505                 sv = scd->scd_svc;
1506
1507                 LASSERT(rpcev == &scd->scd_ev);
1508
1509                 spin_lock(&scd->scd_lock);
1510
1511                 LASSERT (ev->unlinked);
1512                 LASSERT (ev->type == LNET_EVENT_PUT ||
1513                          ev->type == LNET_EVENT_UNLINK);
1514                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1515                          sv->sv_shuttingdown);
1516
1517                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1518                 buffer->buf_peer = ev->initiator;
1519                 buffer->buf_self = ev->target.nid;
1520
1521                 LASSERT(scd->scd_buf_nposted > 0);
1522                 scd->scd_buf_nposted--;
1523
1524                 if (sv->sv_shuttingdown) {
1525                         /* Leave buffer on scd->scd_buf_nposted since
1526                          * srpc_finish_service needs to traverse it. */
1527                         spin_unlock(&scd->scd_lock);
1528                         break;
1529                 }
1530
1531                 if (scd->scd_buf_err_stamp != 0 &&
1532                     scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1533                         /* re-enable adding buffer */
1534                         scd->scd_buf_err_stamp = 0;
1535                         scd->scd_buf_err = 0;
1536                 }
1537
1538                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1539                     scd->scd_buf_adjust == 0 &&
1540                     scd->scd_buf_nposted < scd->scd_buf_low) {
1541                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1542                                                   SFW_TEST_WI_MIN);
1543                         swi_schedule_workitem(&scd->scd_buf_wi);
1544                 }
1545
1546                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1547                 msg = &buffer->buf_msg;
1548                 type = srpc_service2request(sv->sv_id);
1549
1550                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1551                     (msg->msg_type != type &&
1552                      msg->msg_type != __swab32(type)) ||
1553                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1554                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1555                         CERROR ("Dropping RPC (%s) from %s: "
1556                                 "status %d mlength %d type %u magic %u.\n",
1557                                 sv->sv_name, libcfs_id2str(ev->initiator),
1558                                 ev->status, ev->mlength,
1559                                 msg->msg_type, msg->msg_magic);
1560
1561                         /* NB can't call srpc_service_recycle_buffer here since
1562                          * it may call LNetM[DE]Attach. The invalid magic tells
1563                          * srpc_handle_rpc to drop this RPC */
1564                         msg->msg_magic = 0;
1565                 }
1566
1567                 if (!list_empty(&scd->scd_rpc_free)) {
1568                         srpc = list_entry(scd->scd_rpc_free.next,
1569                                           struct srpc_server_rpc,
1570                                           srpc_list);
1571                         list_del(&srpc->srpc_list);
1572
1573                         srpc_init_server_rpc(srpc, scd, buffer);
1574                         list_add_tail(&srpc->srpc_list,
1575                                       &scd->scd_rpc_active);
1576                         swi_schedule_workitem(&srpc->srpc_wi);
1577                 } else {
1578                         list_add_tail(&buffer->buf_list,
1579                                       &scd->scd_buf_blocked);
1580                 }
1581
1582                 spin_unlock(&scd->scd_lock);
1583
1584                 spin_lock(&srpc_data.rpc_glock);
1585                 srpc_data.rpc_counters.rpcs_rcvd++;
1586                 spin_unlock(&srpc_data.rpc_glock);
1587                 break;
1588
1589         case SRPC_BULK_GET_RPLD:
1590                 LASSERT (ev->type == LNET_EVENT_SEND ||
1591                          ev->type == LNET_EVENT_REPLY ||
1592                          ev->type == LNET_EVENT_UNLINK);
1593
1594                 if (!ev->unlinked)
1595                         break; /* wait for final event */
1596
1597         case SRPC_BULK_PUT_SENT:
1598                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1599                         spin_lock(&srpc_data.rpc_glock);
1600
1601                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1602                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1603                         else
1604                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1605
1606                         spin_unlock(&srpc_data.rpc_glock);
1607                 }
1608         case SRPC_REPLY_SENT:
1609                 srpc = rpcev->ev_data;
1610                 scd  = srpc->srpc_scd;
1611
1612                 LASSERT(rpcev == &srpc->srpc_ev);
1613
1614                 spin_lock(&scd->scd_lock);
1615
1616                 rpcev->ev_fired  = 1;
1617                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1618                                    -EINTR : ev->status;
1619                 swi_schedule_workitem(&srpc->srpc_wi);
1620
1621                 spin_unlock(&scd->scd_lock);
1622                 break;
1623         }
1624 }
1625
1626 #ifndef __KERNEL__
1627
1628 int
1629 srpc_check_event (int timeout)
1630 {
1631         lnet_event_t ev;
1632         int          rc;
1633         int          i;
1634
1635         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1636                         timeout * 1000, &ev, &i);
1637         if (rc == 0) return 0;
1638
1639         LASSERT (rc == -EOVERFLOW || rc == 1);
1640
1641         /* We can't affort to miss any events... */
1642         if (rc == -EOVERFLOW) {
1643                 CERROR ("Dropped an event!!!\n");
1644                 abort();
1645         }
1646
1647         srpc_lnet_ev_handler(&ev);
1648         return 1;
1649 }
1650
1651 #endif
1652
1653 int
1654 srpc_startup (void)
1655 {
1656         int rc;
1657
1658         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1659         spin_lock_init(&srpc_data.rpc_glock);
1660
1661         /* 1 second pause to avoid timestamp reuse */
1662         cfs_pause(cfs_time_seconds(1));
1663         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1664
1665         srpc_data.rpc_state = SRPC_STATE_NONE;
1666
1667 #ifdef __KERNEL__
1668         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1669 #else
1670         if (the_lnet.ln_server_mode_flag)
1671                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1672         else
1673                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1674 #endif
1675         if (rc < 0) {
1676                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1677                 return rc;
1678         }
1679
1680         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1681
1682         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1683 #ifdef __KERNEL__
1684         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1685 #else
1686         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1687 #endif
1688         if (rc != 0) {
1689                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1690                 goto bail;
1691         }
1692
1693         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1694         LASSERT(rc == 0);
1695         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1696         LASSERT(rc == 0);
1697
1698         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1699
1700         rc = stt_startup();
1701
1702 bail:
1703         if (rc != 0)
1704                 srpc_shutdown();
1705         else
1706                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1707
1708         return rc;
1709 }
1710
1711 void
1712 srpc_shutdown (void)
1713 {
1714         int i;
1715         int rc;
1716         int state;
1717
1718         state = srpc_data.rpc_state;
1719         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1720
1721         switch (state) {
1722         default:
1723                 LBUG ();
1724         case SRPC_STATE_RUNNING:
1725                 spin_lock(&srpc_data.rpc_glock);
1726
1727                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1728                         srpc_service_t *sv = srpc_data.rpc_services[i];
1729
1730                         LASSERTF (sv == NULL,
1731                                   "service not empty: id %d, name %s\n",
1732                                   i, sv->sv_name);
1733                 }
1734
1735                 spin_unlock(&srpc_data.rpc_glock);
1736
1737                 stt_shutdown();
1738
1739         case SRPC_STATE_EQ_INIT:
1740                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1741                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1742                 LASSERT (rc == 0);
1743                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1744                 LASSERT (rc == 0); /* the EQ should have no user by now */
1745
1746         case SRPC_STATE_NI_INIT:
1747                 LNetNIFini();
1748         }
1749
1750         return;
1751 }