lnet/selftest/rpc.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/selftest/rpc.c
  37  *
  38  * Author: Isaac Huang <isaac@clusterfs.com>
  39  *
  40  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  41  * - percpt data for service to improve smp performance
  42  * - code cleanup
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_LNET
  46
  47 #include "selftest.h"
  48
  49 typedef enum {
  50         SRPC_STATE_NONE,
  51         SRPC_STATE_NI_INIT,
  52         SRPC_STATE_EQ_INIT,
  53         SRPC_STATE_RUNNING,
  54         SRPC_STATE_STOPPING,
  55 } srpc_state_t;
  56
  57 static struct smoketest_rpc {
  58         spinlock_t       rpc_glock;     /* global lock */
  59         srpc_service_t  *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  60         lnet_handle_eq_t rpc_lnet_eq;   /* _the_ LNet event queue */
  61         srpc_state_t     rpc_state;
  62         srpc_counters_t  rpc_counters;
  63         __u64            rpc_matchbits; /* matchbits counter */
  64 } srpc_data;
  65
  66 static inline int
  67 srpc_serv_portal(int svc_id)
  68 {
  69         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  70                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  71 }
  72
  73 /* forward ref's */
  74 int srpc_handle_rpc (swi_workitem_t *wi);
  75
  76 void srpc_get_counters (srpc_counters_t *cnt)
  77 {
  78         spin_lock(&srpc_data.rpc_glock);
  79         *cnt = srpc_data.rpc_counters;
  80         spin_unlock(&srpc_data.rpc_glock);
  81 }
  82
  83 void srpc_set_counters (const srpc_counters_t *cnt)
  84 {
  85         spin_lock(&srpc_data.rpc_glock);
  86         srpc_data.rpc_counters = *cnt;
  87         spin_unlock(&srpc_data.rpc_glock);
  88 }
  89
  90 static int
  91 srpc_add_bulk_page(srpc_bulk_t *bk, struct page *pg, int i, int off, int nob)
  92 {
  93         LASSERT(off < PAGE_SIZE);
  94         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  95
  96         bk->bk_iovs[i].kiov_offset = off;
  97         bk->bk_iovs[i].kiov_page   = pg;
  98         bk->bk_iovs[i].kiov_len    = nob;
  99         return nob;
 100 }
 101
 102 void
 103 srpc_free_bulk (srpc_bulk_t *bk)
 104 {
 105         int         i;
 106         struct page *pg;
 107
 108         LASSERT (bk != NULL);
 109
 110         for (i = 0; i < bk->bk_niov; i++) {
 111                 pg = bk->bk_iovs[i].kiov_page;
 112                 if (pg == NULL) break;
 113
 114                 __free_page(pg);
 115         }
 116
 117         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 118         return;
 119 }
 120
 121 srpc_bulk_t *
 122 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 123                 unsigned bulk_len, int sink)
 124 {
 125         srpc_bulk_t  *bk;
 126         int           i;
 127
 128         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 129
 130         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 131                          offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 132         if (bk == NULL) {
 133                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 134                 return NULL;
 135         }
 136
 137         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
 138         bk->bk_sink   = sink;
 139         bk->bk_len    = bulk_len;
 140         bk->bk_niov   = bulk_npg;
 141
 142         for (i = 0; i < bulk_npg; i++) {
 143                 struct page *pg;
 144                 int         nob;
 145
 146                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 147                 if (pg == NULL) {
 148                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 149                         srpc_free_bulk(bk);
 150                         return NULL;
 151                 }
 152
 153                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 154                       bulk_off;
 155
 156                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 157                 bulk_len -= nob;
 158                 bulk_off = 0;
 159         }
 160
 161         return bk;
 162 }
 163
 164 static inline __u64
 165 srpc_next_id (void)
 166 {
 167         __u64 id;
 168
 169         spin_lock(&srpc_data.rpc_glock);
 170         id = srpc_data.rpc_matchbits++;
 171         spin_unlock(&srpc_data.rpc_glock);
 172         return id;
 173 }
 174
 175 static void
 176 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 177                      struct srpc_service_cd *scd,
 178                      struct srpc_buffer *buffer)
 179 {
 180         memset(rpc, 0, sizeof(*rpc));
 181         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
 182                           srpc_serv_is_framework(scd->scd_svc) ?
 183                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 184
 185         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 186
 187         rpc->srpc_scd      = scd;
 188         rpc->srpc_reqstbuf = buffer;
 189         rpc->srpc_peer     = buffer->buf_peer;
 190         rpc->srpc_self     = buffer->buf_self;
 191         LNetInvalidateHandle(&rpc->srpc_replymdh);
 192 }
 193
 194 static void
 195 srpc_service_fini(struct srpc_service *svc)
 196 {
 197         struct srpc_service_cd  *scd;
 198         struct srpc_server_rpc  *rpc;
 199         struct srpc_buffer      *buf;
 200         struct list_head                *q;
 201         int                     i;
 202
 203         if (svc->sv_cpt_data == NULL)
 204                 return;
 205
 206         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 207                 while (1) {
 208                         if (!list_empty(&scd->scd_buf_posted))
 209                                 q = &scd->scd_buf_posted;
 210                         else if (!list_empty(&scd->scd_buf_blocked))
 211                                 q = &scd->scd_buf_blocked;
 212                         else
 213                                 break;
 214
 215                         while (!list_empty(q)) {
 216                                 buf = list_entry(q->next,
 217                                                      struct srpc_buffer,
 218                                                      buf_list);
 219                                 list_del(&buf->buf_list);
 220                                 LIBCFS_FREE(buf, sizeof(*buf));
 221                         }
 222                 }
 223
 224                 LASSERT(list_empty(&scd->scd_rpc_active));
 225
 226                 while (!list_empty(&scd->scd_rpc_free)) {
 227                         rpc = list_entry(scd->scd_rpc_free.next,
 228                                              struct srpc_server_rpc,
 229                                              srpc_list);
 230                         list_del(&rpc->srpc_list);
 231                         LIBCFS_FREE(rpc, sizeof(*rpc));
 232                 }
 233         }
 234
 235         cfs_percpt_free(svc->sv_cpt_data);
 236         svc->sv_cpt_data = NULL;
 237 }
 238
 239 static int
 240 srpc_service_nrpcs(struct srpc_service *svc)
 241 {
 242         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 243
 244         return srpc_serv_is_framework(svc) ?
 245                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 246 }
 247
 248 int srpc_add_buffer(struct swi_workitem *wi);
 249
 250 static int
 251 srpc_service_init(struct srpc_service *svc)
 252 {
 253         struct srpc_service_cd  *scd;
 254         struct srpc_server_rpc  *rpc;
 255         int                     nrpcs;
 256         int                     i;
 257         int                     j;
 258
 259         svc->sv_shuttingdown = 0;
 260
 261         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 262                                             sizeof(struct srpc_service_cd));
 263         if (svc->sv_cpt_data == NULL)
 264                 return -ENOMEM;
 265
 266         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 267                         1 : cfs_cpt_number(lnet_cpt_table());
 268         nrpcs = srpc_service_nrpcs(svc);
 269
 270         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 271                 scd->scd_cpt = i;
 272                 scd->scd_svc = svc;
 273                 spin_lock_init(&scd->scd_lock);
 274                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 275                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 276                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 277                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 278
 279                 scd->scd_ev.ev_data = scd;
 280                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 281
 282                 /* NB: don't use lst_sched_serial for adding buffer,
 283                  * see details in srpc_service_add_buffers() */
 284                 swi_init_workitem(&scd->scd_buf_wi, scd,
 285                                   srpc_add_buffer, lst_sched_test[i]);
 286
 287                 if (i != 0 && srpc_serv_is_framework(svc)) {
 288                         /* NB: framework service only needs srpc_service_cd for
 289                          * one partition, but we allocate for all to make
 290                          * it easier to implement, it will waste a little
 291                          * memory but nobody should care about this */
 292                         continue;
 293                 }
 294
 295                 for (j = 0; j < nrpcs; j++) {
 296                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 297                                          i, sizeof(*rpc));
 298                         if (rpc == NULL) {
 299                                 srpc_service_fini(svc);
 300                                 return -ENOMEM;
 301                         }
 302                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 303                 }
 304         }
 305
 306         return 0;
 307 }
 308
 309 int
 310 srpc_add_service(struct srpc_service *sv)
 311 {
 312         int id = sv->sv_id;
 313
 314         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 315
 316         if (srpc_service_init(sv) != 0)
 317                 return -ENOMEM;
 318
 319         spin_lock(&srpc_data.rpc_glock);
 320
 321         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 322
 323         if (srpc_data.rpc_services[id] != NULL) {
 324                 spin_unlock(&srpc_data.rpc_glock);
 325                 goto failed;
 326         }
 327
 328         srpc_data.rpc_services[id] = sv;
 329         spin_unlock(&srpc_data.rpc_glock);
 330
 331         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 332         return 0;
 333
 334  failed:
 335         srpc_service_fini(sv);
 336         return -EBUSY;
 337 }
 338
 339 int
 340 srpc_remove_service (srpc_service_t *sv)
 341 {
 342         int id = sv->sv_id;
 343
 344         spin_lock(&srpc_data.rpc_glock);
 345
 346         if (srpc_data.rpc_services[id] != sv) {
 347                 spin_unlock(&srpc_data.rpc_glock);
 348                 return -ENOENT;
 349         }
 350
 351         srpc_data.rpc_services[id] = NULL;
 352         spin_unlock(&srpc_data.rpc_glock);
 353         return 0;
 354 }
 355
 356 static int
 357 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 358                        int len, int options, lnet_process_id_t peer,
 359                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 360 {
 361         int              rc;
 362         lnet_md_t        md;
 363         lnet_handle_me_t meh;
 364
 365         rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 366                           local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
 367         if (rc != 0) {
 368                 CERROR ("LNetMEAttach failed: %d\n", rc);
 369                 LASSERT (rc == -ENOMEM);
 370                 return -ENOMEM;
 371         }
 372
 373         md.threshold = 1;
 374         md.user_ptr  = ev;
 375         md.start     = buf;
 376         md.length    = len;
 377         md.options   = options;
 378         md.eq_handle = srpc_data.rpc_lnet_eq;
 379
 380         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 381         if (rc != 0) {
 382                 CERROR ("LNetMDAttach failed: %d\n", rc);
 383                 LASSERT (rc == -ENOMEM);
 384
 385                 rc = LNetMEUnlink(meh);
 386                 LASSERT (rc == 0);
 387                 return -ENOMEM;
 388         }
 389
 390         CDEBUG (D_NET,
 391                 "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 392                 libcfs_id2str(peer), portal, matchbits);
 393         return 0;
 394 }
 395
 396 static int
 397 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 398                       int options, lnet_process_id_t peer, lnet_nid_t self,
 399                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 400 {
 401         int       rc;
 402         lnet_md_t md;
 403
 404         md.user_ptr  = ev;
 405         md.start     = buf;
 406         md.length    = len;
 407         md.eq_handle = srpc_data.rpc_lnet_eq;
 408         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 409         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 410
 411         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 412         if (rc != 0) {
 413                 CERROR ("LNetMDBind failed: %d\n", rc);
 414                 LASSERT (rc == -ENOMEM);
 415                 return -ENOMEM;
 416         }
 417
 418         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 419          * they're only meaningful for MDs attached to an ME (i.e. passive
 420          * buffers... */
 421         if ((options & LNET_MD_OP_PUT) != 0) {
 422                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 423                              portal, matchbits, 0, 0);
 424         } else {
 425                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 426
 427                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 428         }
 429
 430         if (rc != 0) {
 431                 CERROR ("LNet%s(%s, %d, %lld) failed: %d\n",
 432                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 433                         libcfs_id2str(peer), portal, matchbits, rc);
 434
 435                 /* The forthcoming unlink event will complete this operation
 436                  * with failure, so fall through and return success here.
 437                  */
 438                 rc = LNetMDUnlink(*mdh);
 439                 LASSERT (rc == 0);
 440         } else {
 441                 CDEBUG (D_NET,
 442                         "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 443                         libcfs_id2str(peer), portal, matchbits);
 444         }
 445         return 0;
 446 }
 447
 448 static int
 449 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 450                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 451 {
 452         lnet_process_id_t any = {0};
 453
 454         any.nid = LNET_NID_ANY;
 455         any.pid = LNET_PID_ANY;
 456
 457         return srpc_post_passive_rdma(srpc_serv_portal(service),
 458                                       local, service, buf, len,
 459                                       LNET_MD_OP_PUT, any, mdh, ev);
 460 }
 461
 462 static int
 463 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 464 __must_hold(&scd->scd_lock)
 465 {
 466         struct srpc_service     *sv = scd->scd_svc;
 467         struct srpc_msg         *msg = &buf->buf_msg;
 468         int                     rc;
 469
 470         LNetInvalidateHandle(&buf->buf_mdh);
 471         list_add(&buf->buf_list, &scd->scd_buf_posted);
 472         scd->scd_buf_nposted++;
 473         spin_unlock(&scd->scd_lock);
 474
 475         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 476                                       !srpc_serv_is_framework(sv),
 477                                       msg, sizeof(*msg), &buf->buf_mdh,
 478                                       &scd->scd_ev);
 479
 480         /* At this point, a RPC (new or delayed) may have arrived in
 481          * msg and its event handler has been called. So we must add
 482          * buf to scd_buf_posted _before_ dropping scd_lock */
 483
 484         spin_lock(&scd->scd_lock);
 485
 486         if (rc == 0) {
 487                 if (!sv->sv_shuttingdown)
 488                         return 0;
 489
 490                 spin_unlock(&scd->scd_lock);
 491                 /* srpc_shutdown_service might have tried to unlink me
 492                  * when my buf_mdh was still invalid */
 493                 LNetMDUnlink(buf->buf_mdh);
 494                 spin_lock(&scd->scd_lock);
 495                 return 0;
 496         }
 497
 498         scd->scd_buf_nposted--;
 499         if (sv->sv_shuttingdown)
 500                 return rc; /* don't allow to change scd_buf_posted */
 501
 502         list_del(&buf->buf_list);
 503         spin_unlock(&scd->scd_lock);
 504
 505         LIBCFS_FREE(buf, sizeof(*buf));
 506
 507         spin_lock(&scd->scd_lock);
 508         return rc;
 509 }
 510
 511 int
 512 srpc_add_buffer(struct swi_workitem *wi)
 513 {
 514         struct srpc_service_cd  *scd = wi->swi_workitem.wi_data;
 515         struct srpc_buffer      *buf;
 516         int                     rc = 0;
 517
 518         /* it's called by workitem scheduler threads, these threads
 519          * should have been set CPT affinity, so buffers will be posted
 520          * on CPT local list of Portal */
 521         spin_lock(&scd->scd_lock);
 522
 523         while (scd->scd_buf_adjust > 0 &&
 524                !scd->scd_svc->sv_shuttingdown) {
 525                 scd->scd_buf_adjust--; /* consume it */
 526                 scd->scd_buf_posting++;
 527
 528                 spin_unlock(&scd->scd_lock);
 529
 530                 LIBCFS_ALLOC(buf, sizeof(*buf));
 531                 if (buf == NULL) {
 532                         CERROR("Failed to add new buf to service: %s\n",
 533                                scd->scd_svc->sv_name);
 534                         spin_lock(&scd->scd_lock);
 535                         rc = -ENOMEM;
 536                         break;
 537                 }
 538
 539                 spin_lock(&scd->scd_lock);
 540                 if (scd->scd_svc->sv_shuttingdown) {
 541                         spin_unlock(&scd->scd_lock);
 542                         LIBCFS_FREE(buf, sizeof(*buf));
 543
 544                         spin_lock(&scd->scd_lock);
 545                         rc = -ESHUTDOWN;
 546                         break;
 547                 }
 548
 549                 rc = srpc_service_post_buffer(scd, buf);
 550                 if (rc != 0)
 551                         break; /* buf has been freed inside */
 552
 553                 LASSERT(scd->scd_buf_posting > 0);
 554                 scd->scd_buf_posting--;
 555                 scd->scd_buf_total++;
 556                 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
 557         }
 558
 559         if (rc != 0) {
 560                 scd->scd_buf_err_stamp = cfs_time_current_sec();
 561                 scd->scd_buf_err = rc;
 562
 563                 LASSERT(scd->scd_buf_posting > 0);
 564                 scd->scd_buf_posting--;
 565         }
 566
 567         spin_unlock(&scd->scd_lock);
 568         return 0;
 569 }
 570
 571 int
 572 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 573 {
 574         struct srpc_service_cd  *scd;
 575         int                     rc = 0;
 576         int                     i;
 577
 578         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 579
 580         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 581                 spin_lock(&scd->scd_lock);
 582
 583                 scd->scd_buf_err = 0;
 584                 scd->scd_buf_err_stamp = 0;
 585                 scd->scd_buf_posting = 0;
 586                 scd->scd_buf_adjust = nbuffer;
 587                 /* start to post buffers */
 588                 swi_schedule_workitem(&scd->scd_buf_wi);
 589                 spin_unlock(&scd->scd_lock);
 590
 591                 /* framework service only post buffer for one partition  */
 592                 if (srpc_serv_is_framework(sv))
 593                         break;
 594         }
 595
 596         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 597                 spin_lock(&scd->scd_lock);
 598                 /*
 599                  * NB: srpc_service_add_buffers() can be called inside
 600                  * thread context of lst_sched_serial, and we don't normally
 601                  * allow to sleep inside thread context of WI scheduler
 602                  * because it will block current scheduler thread from doing
 603                  * anything else, even worse, it could deadlock if it's
 604                  * waiting on result from another WI of the same scheduler.
 605                  * However, it's safe at here because scd_buf_wi is scheduled
 606                  * by thread in a different WI scheduler (lst_sched_test),
 607                  * so we don't have any risk of deadlock, though this could
 608                  * block all WIs pending on lst_sched_serial for a moment
 609                  * which is not good but not fatal.
 610                  */
 611                 lst_wait_until(scd->scd_buf_err != 0 ||
 612                                (scd->scd_buf_adjust == 0 &&
 613                                 scd->scd_buf_posting == 0),
 614                                scd->scd_lock, "waiting for adding buffer\n");
 615
 616                 if (scd->scd_buf_err != 0 && rc == 0)
 617                         rc = scd->scd_buf_err;
 618
 619                 spin_unlock(&scd->scd_lock);
 620         }
 621
 622         return rc;
 623 }
 624
 625 void
 626 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 627 {
 628         struct srpc_service_cd  *scd;
 629         int                     num;
 630         int                     i;
 631
 632         LASSERT(!sv->sv_shuttingdown);
 633
 634         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 635                 spin_lock(&scd->scd_lock);
 636
 637                 num = scd->scd_buf_total + scd->scd_buf_posting;
 638                 scd->scd_buf_adjust -= min(nbuffer, num);
 639
 640                 spin_unlock(&scd->scd_lock);
 641         }
 642 }
 643
 644 /* returns 1 if sv has finished, otherwise 0 */
 645 int
 646 srpc_finish_service(struct srpc_service *sv)
 647 {
 648         struct srpc_service_cd  *scd;
 649         struct srpc_server_rpc  *rpc;
 650         int                     i;
 651
 652         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 653
 654         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 655                 spin_lock(&scd->scd_lock);
 656                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 657                         spin_unlock(&scd->scd_lock);
 658                         return 0;
 659                 }
 660
 661                 if (scd->scd_buf_nposted > 0) {
 662                         CDEBUG(D_NET, "waiting for %d posted buffers to "
 663                                "unlink\n", scd->scd_buf_nposted);
 664                         spin_unlock(&scd->scd_lock);
 665                         return 0;
 666                 }
 667
 668                 if (list_empty(&scd->scd_rpc_active)) {
 669                         spin_unlock(&scd->scd_lock);
 670                         continue;
 671                 }
 672
 673                 rpc = list_entry(scd->scd_rpc_active.next,
 674                                      struct srpc_server_rpc, srpc_list);
 675                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
 676                         "wi %s scheduled %d running %d, "
 677                         "ev fired %d type %d status %d lnet %d\n",
 678                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 679                         swi_state2str(rpc->srpc_wi.swi_state),
 680                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 681                         rpc->srpc_wi.swi_workitem.wi_running,
 682                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 683                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 684                 spin_unlock(&scd->scd_lock);
 685                 return 0;
 686         }
 687
 688         /* no lock needed from now on */
 689         srpc_service_fini(sv);
 690         return 1;
 691 }
 692
 693 /* called with sv->sv_lock held */
 694 static void
 695 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
 696 __must_hold(&scd->scd_lock)
 697 {
 698         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 699                 if (srpc_service_post_buffer(scd, buf) != 0) {
 700                         CWARN("Failed to post %s buffer\n",
 701                               scd->scd_svc->sv_name);
 702                 }
 703                 return;
 704         }
 705
 706         /* service is shutting down, or we want to recycle some buffers */
 707         scd->scd_buf_total--;
 708
 709         if (scd->scd_buf_adjust < 0) {
 710                 scd->scd_buf_adjust++;
 711                 if (scd->scd_buf_adjust < 0 &&
 712                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 713                         CDEBUG(D_INFO,
 714                                "Try to recyle %d buffers but nothing left\n",
 715                                scd->scd_buf_adjust);
 716                         scd->scd_buf_adjust = 0;
 717                 }
 718         }
 719
 720         spin_unlock(&scd->scd_lock);
 721         LIBCFS_FREE(buf, sizeof(*buf));
 722         spin_lock(&scd->scd_lock);
 723 }
 724
 725 void
 726 srpc_abort_service(struct srpc_service *sv)
 727 {
 728         struct srpc_service_cd  *scd;
 729         struct srpc_server_rpc  *rpc;
 730         int                     i;
 731
 732         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 733                sv->sv_id, sv->sv_name);
 734
 735         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 736                 spin_lock(&scd->scd_lock);
 737
 738                 /* schedule in-flight RPCs to notice the abort, NB:
 739                  * racing with incoming RPCs; complete fix should make test
 740                  * RPCs carry session ID in its headers */
 741                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 742                         rpc->srpc_aborted = 1;
 743                         swi_schedule_workitem(&rpc->srpc_wi);
 744                 }
 745
 746                 spin_unlock(&scd->scd_lock);
 747         }
 748 }
 749
 750 void
 751 srpc_shutdown_service(srpc_service_t *sv)
 752 {
 753         struct srpc_service_cd  *scd;
 754         struct srpc_server_rpc  *rpc;
 755         srpc_buffer_t           *buf;
 756         int                     i;
 757
 758         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 759                sv->sv_id, sv->sv_name);
 760
 761         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 762                 spin_lock(&scd->scd_lock);
 763
 764         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 765
 766         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 767                 spin_unlock(&scd->scd_lock);
 768
 769         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 770                 spin_lock(&scd->scd_lock);
 771
 772                 /* schedule in-flight RPCs to notice the shutdown */
 773                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 774                         swi_schedule_workitem(&rpc->srpc_wi);
 775
 776                 spin_unlock(&scd->scd_lock);
 777
 778                 /* OK to traverse scd_buf_posted without lock, since no one
 779                  * touches scd_buf_posted now */
 780                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 781                         LNetMDUnlink(buf->buf_mdh);
 782         }
 783 }
 784
 785 static int
 786 srpc_send_request (srpc_client_rpc_t *rpc)
 787 {
 788         srpc_event_t *ev = &rpc->crpc_reqstev;
 789         int           rc;
 790
 791         ev->ev_fired = 0;
 792         ev->ev_data  = rpc;
 793         ev->ev_type  = SRPC_REQUEST_SENT;
 794
 795         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 796                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 797                                    sizeof(srpc_msg_t), LNET_MD_OP_PUT,
 798                                    rpc->crpc_dest, LNET_NID_ANY,
 799                                    &rpc->crpc_reqstmdh, ev);
 800         if (rc != 0) {
 801                 LASSERT (rc == -ENOMEM);
 802                 ev->ev_fired = 1;  /* no more event expected */
 803         }
 804         return rc;
 805 }
 806
 807 static int
 808 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 809 {
 810         srpc_event_t *ev = &rpc->crpc_replyev;
 811         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 812         int           rc;
 813
 814         ev->ev_fired = 0;
 815         ev->ev_data  = rpc;
 816         ev->ev_type  = SRPC_REPLY_RCVD;
 817
 818         *id = srpc_next_id();
 819
 820         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 821                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 822                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 823                                     &rpc->crpc_replymdh, ev);
 824         if (rc != 0) {
 825                 LASSERT (rc == -ENOMEM);
 826                 ev->ev_fired = 1;  /* no more event expected */
 827         }
 828         return rc;
 829 }
 830
 831 static int
 832 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 833 {
 834         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 835         srpc_event_t *ev = &rpc->crpc_bulkev;
 836         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 837         int           rc;
 838         int           opt;
 839
 840         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 841
 842         if (bk->bk_niov == 0) return 0; /* nothing to do */
 843
 844         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 845         opt |= LNET_MD_KIOV;
 846
 847         ev->ev_fired = 0;
 848         ev->ev_data  = rpc;
 849         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 850
 851         *id = srpc_next_id();
 852
 853         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 854                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 855                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 856         if (rc != 0) {
 857                 LASSERT (rc == -ENOMEM);
 858                 ev->ev_fired = 1;  /* no more event expected */
 859         }
 860         return rc;
 861 }
 862
 863 static int
 864 srpc_do_bulk (srpc_server_rpc_t *rpc)
 865 {
 866         srpc_event_t  *ev = &rpc->srpc_ev;
 867         srpc_bulk_t   *bk = rpc->srpc_bulk;
 868         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 869         int            rc;
 870         int            opt;
 871
 872         LASSERT (bk != NULL);
 873
 874         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 875         opt |= LNET_MD_KIOV;
 876
 877         ev->ev_fired = 0;
 878         ev->ev_data  = rpc;
 879         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 880
 881         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 882                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 883                                    rpc->srpc_peer, rpc->srpc_self,
 884                                    &bk->bk_mdh, ev);
 885         if (rc != 0)
 886                 ev->ev_fired = 1;  /* no more event expected */
 887         return rc;
 888 }
 889
 890 /* only called from srpc_handle_rpc */
 891 static void
 892 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
 893 {
 894         struct srpc_service_cd  *scd = rpc->srpc_scd;
 895         struct srpc_service     *sv  = scd->scd_svc;
 896         srpc_buffer_t           *buffer;
 897
 898         LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 899
 900         rpc->srpc_status = status;
 901
 902         CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
 903                 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 904                 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 905                 swi_state2str(rpc->srpc_wi.swi_state), status);
 906
 907         if (status != 0) {
 908                 spin_lock(&srpc_data.rpc_glock);
 909                 srpc_data.rpc_counters.rpcs_dropped++;
 910                 spin_unlock(&srpc_data.rpc_glock);
 911         }
 912
 913         if (rpc->srpc_done != NULL)
 914                 (*rpc->srpc_done) (rpc);
 915         LASSERT(rpc->srpc_bulk == NULL);
 916
 917         spin_lock(&scd->scd_lock);
 918
 919         if (rpc->srpc_reqstbuf != NULL) {
 920                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 921                  * sv won't go away for scd_rpc_active must not be empty */
 922                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 923                 rpc->srpc_reqstbuf = NULL;
 924         }
 925
 926         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 927
 928         /*
 929          * No one can schedule me now since:
 930          * - I'm not on scd_rpc_active.
 931          * - all LNet events have been fired.
 932          * Cancel pending schedules and prevent future schedule attempts:
 933          */
 934         LASSERT(rpc->srpc_ev.ev_fired);
 935         swi_exit_workitem(&rpc->srpc_wi);
 936
 937         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 938                 buffer = list_entry(scd->scd_buf_blocked.next,
 939                                         srpc_buffer_t, buf_list);
 940                 list_del(&buffer->buf_list);
 941
 942                 srpc_init_server_rpc(rpc, scd, buffer);
 943                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 944                 swi_schedule_workitem(&rpc->srpc_wi);
 945         } else {
 946                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 947         }
 948
 949         spin_unlock(&scd->scd_lock);
 950         return;
 951 }
 952
 953 /* handles an incoming RPC */
 954 int
 955 srpc_handle_rpc(swi_workitem_t *wi)
 956 {
 957         struct srpc_server_rpc  *rpc = wi->swi_workitem.wi_data;
 958         struct srpc_service_cd  *scd = rpc->srpc_scd;
 959         struct srpc_service     *sv = scd->scd_svc;
 960         srpc_event_t            *ev = &rpc->srpc_ev;
 961         int                     rc = 0;
 962
 963         LASSERT(wi == &rpc->srpc_wi);
 964
 965         spin_lock(&scd->scd_lock);
 966
 967         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 968                 spin_unlock(&scd->scd_lock);
 969
 970                 if (rpc->srpc_bulk != NULL)
 971                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 972                 LNetMDUnlink(rpc->srpc_replymdh);
 973
 974                 if (ev->ev_fired) { /* no more event, OK to finish */
 975                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 976                         return 1;
 977                 }
 978                 return 0;
 979         }
 980
 981         spin_unlock(&scd->scd_lock);
 982
 983         switch (wi->swi_state) {
 984         default:
 985                 LBUG ();
 986         case SWI_STATE_NEWBORN: {
 987                 srpc_msg_t           *msg;
 988                 srpc_generic_reply_t *reply;
 989
 990                 msg = &rpc->srpc_reqstbuf->buf_msg;
 991                 reply = &rpc->srpc_replymsg.msg_body.reply;
 992
 993                 if (msg->msg_magic == 0) {
 994                         /* moaned already in srpc_lnet_ev_handler */
 995                         srpc_server_rpc_done(rpc, EBADMSG);
 996                         return 1;
 997                 }
 998
 999                 srpc_unpack_msg_hdr(msg);
1000                 if (msg->msg_version != SRPC_MSG_VERSION) {
1001                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1002                               msg->msg_version, SRPC_MSG_VERSION,
1003                               libcfs_id2str(rpc->srpc_peer));
1004                         reply->status = EPROTO;
1005                         /* drop through and send reply */
1006                 } else {
1007                         reply->status = 0;
1008                         rc = (*sv->sv_handler)(rpc);
1009                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1010                         if (rc != 0) {
1011                                 srpc_server_rpc_done(rpc, rc);
1012                                 return 1;
1013                         }
1014                 }
1015
1016                 wi->swi_state = SWI_STATE_BULK_STARTED;
1017
1018                 if (rpc->srpc_bulk != NULL) {
1019                         rc = srpc_do_bulk(rpc);
1020                         if (rc == 0)
1021                                 return 0; /* wait for bulk */
1022
1023                         LASSERT (ev->ev_fired);
1024                         ev->ev_status = rc;
1025                 }
1026         }
1027         case SWI_STATE_BULK_STARTED:
1028                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1029
1030                 if (rpc->srpc_bulk != NULL) {
1031                         rc = ev->ev_status;
1032
1033                         if (sv->sv_bulk_ready != NULL)
1034                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1035
1036                         if (rc != 0) {
1037                                 srpc_server_rpc_done(rpc, rc);
1038                                 return 1;
1039                         }
1040                 }
1041
1042                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1043                 rc = srpc_send_reply(rpc);
1044                 if (rc == 0)
1045                         return 0; /* wait for reply */
1046                 srpc_server_rpc_done(rpc, rc);
1047                 return 1;
1048
1049         case SWI_STATE_REPLY_SUBMITTED:
1050                 if (!ev->ev_fired) {
1051                         CERROR("RPC %p: bulk %p, service %d\n",
1052                                rpc, rpc->srpc_bulk, sv->sv_id);
1053                         CERROR("Event: status %d, type %d, lnet %d\n",
1054                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1055                         LASSERT (ev->ev_fired);
1056                 }
1057
1058                 wi->swi_state = SWI_STATE_DONE;
1059                 srpc_server_rpc_done(rpc, ev->ev_status);
1060                 return 1;
1061         }
1062
1063         return 0;
1064 }
1065
1066 static void
1067 srpc_client_rpc_expired (void *data)
1068 {
1069         srpc_client_rpc_t *rpc = data;
1070
1071         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1072                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1073                rpc->crpc_timeout);
1074
1075         spin_lock(&rpc->crpc_lock);
1076
1077         rpc->crpc_timeout = 0;
1078         srpc_abort_rpc(rpc, -ETIMEDOUT);
1079
1080         spin_unlock(&rpc->crpc_lock);
1081
1082         spin_lock(&srpc_data.rpc_glock);
1083         srpc_data.rpc_counters.rpcs_expired++;
1084         spin_unlock(&srpc_data.rpc_glock);
1085 }
1086
1087 static void
1088 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1089 {
1090         stt_timer_t *timer = &rpc->crpc_timer;
1091
1092         if (rpc->crpc_timeout == 0)
1093                 return;
1094
1095         INIT_LIST_HEAD(&timer->stt_list);
1096         timer->stt_data    = rpc;
1097         timer->stt_func    = srpc_client_rpc_expired;
1098         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1099                                           cfs_time_current_sec());
1100         stt_add_timer(timer);
1101         return;
1102 }
1103
1104 /*
1105  * Called with rpc->crpc_lock held.
1106  *
1107  * Upon exit the RPC expiry timer is not queued and the handler is not
1108  * running on any CPU. */
1109 static void
1110 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1111 {
1112         /* timer not planted or already exploded */
1113         if (rpc->crpc_timeout == 0)
1114                 return;
1115
1116         /* timer successfully defused */
1117         if (stt_del_timer(&rpc->crpc_timer))
1118                 return;
1119
1120         /* timer detonated, wait for it to explode */
1121         while (rpc->crpc_timeout != 0) {
1122                 spin_unlock(&rpc->crpc_lock);
1123
1124                 schedule();
1125
1126                 spin_lock(&rpc->crpc_lock);
1127         }
1128 }
1129
1130 static void
1131 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1132 {
1133         swi_workitem_t *wi = &rpc->crpc_wi;
1134
1135         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1136
1137         spin_lock(&rpc->crpc_lock);
1138
1139         rpc->crpc_closed = 1;
1140         if (rpc->crpc_status == 0)
1141                 rpc->crpc_status = status;
1142
1143         srpc_del_client_rpc_timer(rpc);
1144
1145         CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1146                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1147                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1148                 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1149
1150         /*
1151          * No one can schedule me now since:
1152          * - RPC timer has been defused.
1153          * - all LNet events have been fired.
1154          * - crpc_closed has been set, preventing srpc_abort_rpc from
1155          *   scheduling me.
1156          * Cancel pending schedules and prevent future schedule attempts:
1157          */
1158         LASSERT (!srpc_event_pending(rpc));
1159         swi_exit_workitem(wi);
1160
1161         spin_unlock(&rpc->crpc_lock);
1162
1163         (*rpc->crpc_done)(rpc);
1164         return;
1165 }
1166
1167 /* sends an outgoing RPC */
1168 int
1169 srpc_send_rpc (swi_workitem_t *wi)
1170 {
1171         int                rc = 0;
1172         srpc_client_rpc_t *rpc;
1173         srpc_msg_t        *reply;
1174         int                do_bulk;
1175
1176         LASSERT(wi != NULL);
1177
1178         rpc = wi->swi_workitem.wi_data;
1179
1180         LASSERT (rpc != NULL);
1181         LASSERT (wi == &rpc->crpc_wi);
1182
1183         reply = &rpc->crpc_replymsg;
1184         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1185
1186         spin_lock(&rpc->crpc_lock);
1187
1188         if (rpc->crpc_aborted) {
1189                 spin_unlock(&rpc->crpc_lock);
1190                 goto abort;
1191         }
1192
1193         spin_unlock(&rpc->crpc_lock);
1194
1195         switch (wi->swi_state) {
1196         default:
1197                 LBUG ();
1198         case SWI_STATE_NEWBORN:
1199                 LASSERT (!srpc_event_pending(rpc));
1200
1201                 rc = srpc_prepare_reply(rpc);
1202                 if (rc != 0) {
1203                         srpc_client_rpc_done(rpc, rc);
1204                         return 1;
1205                 }
1206
1207                 rc = srpc_prepare_bulk(rpc);
1208                 if (rc != 0) break;
1209
1210                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1211                 rc = srpc_send_request(rpc);
1212                 break;
1213
1214         case SWI_STATE_REQUEST_SUBMITTED:
1215                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1216                  * order; however, they're processed in a strict order:
1217                  * rqt, rpy, and bulk. */
1218                 if (!rpc->crpc_reqstev.ev_fired) break;
1219
1220                 rc = rpc->crpc_reqstev.ev_status;
1221                 if (rc != 0) break;
1222
1223                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1224                 /* perhaps more events, fall thru */
1225         case SWI_STATE_REQUEST_SENT: {
1226                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1227
1228                 if (!rpc->crpc_replyev.ev_fired) break;
1229
1230                 rc = rpc->crpc_replyev.ev_status;
1231                 if (rc != 0) break;
1232
1233                 srpc_unpack_msg_hdr(reply);
1234                 if (reply->msg_type != type ||
1235                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1236                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1237                         CWARN ("Bad message from %s: type %u (%d expected),"
1238                                " magic %u (%d expected).\n",
1239                                libcfs_id2str(rpc->crpc_dest),
1240                                reply->msg_type, type,
1241                                reply->msg_magic, SRPC_MSG_MAGIC);
1242                         rc = -EBADMSG;
1243                         break;
1244                 }
1245
1246                 if (do_bulk && reply->msg_body.reply.status != 0) {
1247                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1248                                "case peer didn't initiate bulk transfer\n",
1249                                reply->msg_body.reply.status,
1250                                libcfs_id2str(rpc->crpc_dest));
1251                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1252                 }
1253
1254                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1255         }
1256         case SWI_STATE_REPLY_RECEIVED:
1257                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1258
1259                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1260
1261                 /* Bulk buffer was unlinked due to remote error. Clear error
1262                  * since reply buffer still contains valid data.
1263                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1264                  * remote error. */
1265                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1266                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1267                         rc = 0;
1268
1269                 wi->swi_state = SWI_STATE_DONE;
1270                 srpc_client_rpc_done(rpc, rc);
1271                 return 1;
1272         }
1273
1274         if (rc != 0) {
1275                 spin_lock(&rpc->crpc_lock);
1276                 srpc_abort_rpc(rpc, rc);
1277                 spin_unlock(&rpc->crpc_lock);
1278         }
1279
1280 abort:
1281         if (rpc->crpc_aborted) {
1282                 LNetMDUnlink(rpc->crpc_reqstmdh);
1283                 LNetMDUnlink(rpc->crpc_replymdh);
1284                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1285
1286                 if (!srpc_event_pending(rpc)) {
1287                         srpc_client_rpc_done(rpc, -EINTR);
1288                         return 1;
1289                 }
1290         }
1291         return 0;
1292 }
1293
1294 srpc_client_rpc_t *
1295 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1296                         int nbulkiov, int bulklen,
1297                         void (*rpc_done)(srpc_client_rpc_t *),
1298                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1299 {
1300         srpc_client_rpc_t *rpc;
1301
1302         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1303                                    crpc_bulk.bk_iovs[nbulkiov]));
1304         if (rpc == NULL)
1305                 return NULL;
1306
1307         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1308                              bulklen, rpc_done, rpc_fini, priv);
1309         return rpc;
1310 }
1311
1312 /* called with rpc->crpc_lock held */
1313 void
1314 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1315 {
1316         LASSERT (why != 0);
1317
1318         if (rpc->crpc_aborted || /* already aborted */
1319             rpc->crpc_closed)    /* callback imminent */
1320                 return;
1321
1322         CDEBUG (D_NET,
1323                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1324                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1325                 swi_state2str(rpc->crpc_wi.swi_state), why);
1326
1327         rpc->crpc_aborted = 1;
1328         rpc->crpc_status  = why;
1329         swi_schedule_workitem(&rpc->crpc_wi);
1330         return;
1331 }
1332
1333 /* called with rpc->crpc_lock held */
1334 void
1335 srpc_post_rpc (srpc_client_rpc_t *rpc)
1336 {
1337         LASSERT (!rpc->crpc_aborted);
1338         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1339
1340         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1341                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1342                 rpc->crpc_timeout);
1343
1344         srpc_add_client_rpc_timer(rpc);
1345         swi_schedule_workitem(&rpc->crpc_wi);
1346         return;
1347 }
1348
1349
1350 int
1351 srpc_send_reply(struct srpc_server_rpc *rpc)
1352 {
1353         srpc_event_t            *ev = &rpc->srpc_ev;
1354         struct srpc_msg         *msg = &rpc->srpc_replymsg;
1355         struct srpc_buffer      *buffer = rpc->srpc_reqstbuf;
1356         struct srpc_service_cd  *scd = rpc->srpc_scd;
1357         struct srpc_service     *sv = scd->scd_svc;
1358         __u64                   rpyid;
1359         int                     rc;
1360
1361         LASSERT(buffer != NULL);
1362         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1363
1364         spin_lock(&scd->scd_lock);
1365
1366         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1367                 /* Repost buffer before replying since test client
1368                  * might send me another RPC once it gets the reply */
1369                 if (srpc_service_post_buffer(scd, buffer) != 0)
1370                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1371                 rpc->srpc_reqstbuf = NULL;
1372         }
1373
1374         spin_unlock(&scd->scd_lock);
1375
1376         ev->ev_fired = 0;
1377         ev->ev_data  = rpc;
1378         ev->ev_type  = SRPC_REPLY_SENT;
1379
1380         msg->msg_magic   = SRPC_MSG_MAGIC;
1381         msg->msg_version = SRPC_MSG_VERSION;
1382         msg->msg_type    = srpc_service2reply(sv->sv_id);
1383
1384         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1385                                    sizeof(*msg), LNET_MD_OP_PUT,
1386                                    rpc->srpc_peer, rpc->srpc_self,
1387                                    &rpc->srpc_replymdh, ev);
1388         if (rc != 0)
1389                 ev->ev_fired = 1;  /* no more event expected */
1390         return rc;
1391 }
1392
1393 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1394 static void
1395 srpc_lnet_ev_handler(lnet_event_t *ev)
1396 {
1397         struct srpc_service_cd  *scd;
1398         srpc_event_t      *rpcev = ev->md.user_ptr;
1399         srpc_client_rpc_t *crpc;
1400         srpc_server_rpc_t *srpc;
1401         srpc_buffer_t     *buffer;
1402         srpc_service_t    *sv;
1403         srpc_msg_t        *msg;
1404         srpc_msg_type_t    type;
1405
1406         LASSERT (!in_interrupt());
1407
1408         if (ev->status != 0) {
1409                 __u32 errors;
1410
1411                 spin_lock(&srpc_data.rpc_glock);
1412                 if (ev->status != -ECANCELED) /* cancellation is not error */
1413                         srpc_data.rpc_counters.errors++;
1414                 errors = srpc_data.rpc_counters.errors;
1415                 spin_unlock(&srpc_data.rpc_glock);
1416
1417                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1418                         ev->status, ev->type, errors);
1419         }
1420
1421         rpcev->ev_lnet = ev->type;
1422
1423         switch (rpcev->ev_type) {
1424         default:
1425                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1426                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1427                 LBUG ();
1428         case SRPC_REQUEST_SENT:
1429                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1430                         spin_lock(&srpc_data.rpc_glock);
1431                         srpc_data.rpc_counters.rpcs_sent++;
1432                         spin_unlock(&srpc_data.rpc_glock);
1433                 }
1434         case SRPC_REPLY_RCVD:
1435         case SRPC_BULK_REQ_RCVD:
1436                 crpc = rpcev->ev_data;
1437
1438                 if (rpcev != &crpc->crpc_reqstev &&
1439                     rpcev != &crpc->crpc_replyev &&
1440                     rpcev != &crpc->crpc_bulkev) {
1441                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1442                                rpcev, crpc, &crpc->crpc_reqstev,
1443                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1444                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1445                                rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1446                         LBUG ();
1447                 }
1448
1449                 spin_lock(&crpc->crpc_lock);
1450
1451                 LASSERT(rpcev->ev_fired == 0);
1452                 rpcev->ev_fired  = 1;
1453                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1454                                                 -EINTR : ev->status;
1455                 swi_schedule_workitem(&crpc->crpc_wi);
1456
1457                 spin_unlock(&crpc->crpc_lock);
1458                 break;
1459
1460         case SRPC_REQUEST_RCVD:
1461                 scd = rpcev->ev_data;
1462                 sv = scd->scd_svc;
1463
1464                 LASSERT(rpcev == &scd->scd_ev);
1465
1466                 spin_lock(&scd->scd_lock);
1467
1468                 LASSERT (ev->unlinked);
1469                 LASSERT (ev->type == LNET_EVENT_PUT ||
1470                          ev->type == LNET_EVENT_UNLINK);
1471                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1472                          sv->sv_shuttingdown);
1473
1474                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1475                 buffer->buf_peer = ev->initiator;
1476                 buffer->buf_self = ev->target.nid;
1477
1478                 LASSERT(scd->scd_buf_nposted > 0);
1479                 scd->scd_buf_nposted--;
1480
1481                 if (sv->sv_shuttingdown) {
1482                         /* Leave buffer on scd->scd_buf_nposted since
1483                          * srpc_finish_service needs to traverse it. */
1484                         spin_unlock(&scd->scd_lock);
1485                         break;
1486                 }
1487
1488                 if (scd->scd_buf_err_stamp != 0 &&
1489                     scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1490                         /* re-enable adding buffer */
1491                         scd->scd_buf_err_stamp = 0;
1492                         scd->scd_buf_err = 0;
1493                 }
1494
1495                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1496                     scd->scd_buf_adjust == 0 &&
1497                     scd->scd_buf_nposted < scd->scd_buf_low) {
1498                         scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1499                                                   SFW_TEST_WI_MIN);
1500                         swi_schedule_workitem(&scd->scd_buf_wi);
1501                 }
1502
1503                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1504                 msg = &buffer->buf_msg;
1505                 type = srpc_service2request(sv->sv_id);
1506
1507                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1508                     (msg->msg_type != type &&
1509                      msg->msg_type != __swab32(type)) ||
1510                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1511                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1512                         CERROR ("Dropping RPC (%s) from %s: "
1513                                 "status %d mlength %d type %u magic %u.\n",
1514                                 sv->sv_name, libcfs_id2str(ev->initiator),
1515                                 ev->status, ev->mlength,
1516                                 msg->msg_type, msg->msg_magic);
1517
1518                         /* NB can't call srpc_service_recycle_buffer here since
1519                          * it may call LNetM[DE]Attach. The invalid magic tells
1520                          * srpc_handle_rpc to drop this RPC */
1521                         msg->msg_magic = 0;
1522                 }
1523
1524                 if (!list_empty(&scd->scd_rpc_free)) {
1525                         srpc = list_entry(scd->scd_rpc_free.next,
1526                                           struct srpc_server_rpc,
1527                                           srpc_list);
1528                         list_del(&srpc->srpc_list);
1529
1530                         srpc_init_server_rpc(srpc, scd, buffer);
1531                         list_add_tail(&srpc->srpc_list,
1532                                       &scd->scd_rpc_active);
1533                         swi_schedule_workitem(&srpc->srpc_wi);
1534                 } else {
1535                         list_add_tail(&buffer->buf_list,
1536                                       &scd->scd_buf_blocked);
1537                 }
1538
1539                 spin_unlock(&scd->scd_lock);
1540
1541                 spin_lock(&srpc_data.rpc_glock);
1542                 srpc_data.rpc_counters.rpcs_rcvd++;
1543                 spin_unlock(&srpc_data.rpc_glock);
1544                 break;
1545
1546         case SRPC_BULK_GET_RPLD:
1547                 LASSERT (ev->type == LNET_EVENT_SEND ||
1548                          ev->type == LNET_EVENT_REPLY ||
1549                          ev->type == LNET_EVENT_UNLINK);
1550
1551                 if (!ev->unlinked)
1552                         break; /* wait for final event */
1553
1554         case SRPC_BULK_PUT_SENT:
1555                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1556                         spin_lock(&srpc_data.rpc_glock);
1557
1558                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1559                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1560                         else
1561                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1562
1563                         spin_unlock(&srpc_data.rpc_glock);
1564                 }
1565         case SRPC_REPLY_SENT:
1566                 srpc = rpcev->ev_data;
1567                 scd  = srpc->srpc_scd;
1568
1569                 LASSERT(rpcev == &srpc->srpc_ev);
1570
1571                 spin_lock(&scd->scd_lock);
1572
1573                 rpcev->ev_fired  = 1;
1574                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1575                                    -EINTR : ev->status;
1576                 swi_schedule_workitem(&srpc->srpc_wi);
1577
1578                 spin_unlock(&scd->scd_lock);
1579                 break;
1580         }
1581 }
1582
1583
1584 int
1585 srpc_startup (void)
1586 {
1587         int rc;
1588
1589         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1590         spin_lock_init(&srpc_data.rpc_glock);
1591
1592         /* 1 second pause to avoid timestamp reuse */
1593         set_current_state(TASK_UNINTERRUPTIBLE);
1594         schedule_timeout(cfs_time_seconds(1));
1595         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1596
1597         srpc_data.rpc_state = SRPC_STATE_NONE;
1598
1599         rc = LNetNIInit(LNET_PID_LUSTRE);
1600         if (rc < 0) {
1601                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1602                 return rc;
1603         }
1604
1605         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1606
1607         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1608         rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1609         if (rc != 0) {
1610                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1611                 goto bail;
1612         }
1613
1614         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1615         LASSERT(rc == 0);
1616         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1617         LASSERT(rc == 0);
1618
1619         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1620
1621         rc = stt_startup();
1622
1623 bail:
1624         if (rc != 0)
1625                 srpc_shutdown();
1626         else
1627                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1628
1629         return rc;
1630 }
1631
1632 void
1633 srpc_shutdown (void)
1634 {
1635         int i;
1636         int rc;
1637         int state;
1638
1639         state = srpc_data.rpc_state;
1640         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1641
1642         switch (state) {
1643         default:
1644                 LBUG ();
1645         case SRPC_STATE_RUNNING:
1646                 spin_lock(&srpc_data.rpc_glock);
1647
1648                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1649                         srpc_service_t *sv = srpc_data.rpc_services[i];
1650
1651                         LASSERTF (sv == NULL,
1652                                   "service not empty: id %d, name %s\n",
1653                                   i, sv->sv_name);
1654                 }
1655
1656                 spin_unlock(&srpc_data.rpc_glock);
1657
1658                 stt_shutdown();
1659
1660         case SRPC_STATE_EQ_INIT:
1661                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1662                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1663                 LASSERT (rc == 0);
1664                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1665                 LASSERT (rc == 0); /* the EQ should have no user by now */
1666
1667         case SRPC_STATE_NI_INIT:
1668                 LNetNIFini();
1669         }
1670
1671         return;
1672 }