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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/selftest/rpc.c
  33  *
  34  * Author: Isaac Huang <isaac@clusterfs.com>
  35  *
  36  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  37  * - percpt data for service to improve smp performance
  38  * - code cleanup
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include "selftest.h"
  44
  45 enum srpc_state {
  46         SRPC_STATE_NONE,
  47         SRPC_STATE_NI_INIT,
  48         SRPC_STATE_EQ_INIT,
  49         SRPC_STATE_RUNNING,
  50         SRPC_STATE_STOPPING,
  51 };
  52
  53 static struct smoketest_rpc {
  54         spinlock_t       rpc_glock;     /* global lock */
  55         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  56         lnet_handler_t           rpc_lnet_handler;/* _the_ LNet event handler */
  57         enum srpc_state          rpc_state;
  58         struct srpc_counters     rpc_counters;
  59         __u64                    rpc_matchbits; /* matchbits counter */
  60 } srpc_data;
  61
  62 static inline int
  63 srpc_serv_portal(int svc_id)
  64 {
  65         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  66                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  67 }
  68
  69 /* forward ref's */
  70 static int srpc_handle_rpc(struct swi_workitem *wi);
  71
  72 void srpc_get_counters(struct srpc_counters *cnt)
  73 {
  74         spin_lock(&srpc_data.rpc_glock);
  75         *cnt = srpc_data.rpc_counters;
  76         spin_unlock(&srpc_data.rpc_glock);
  77 }
  78
  79 void srpc_set_counters(const struct srpc_counters *cnt)
  80 {
  81         spin_lock(&srpc_data.rpc_glock);
  82         srpc_data.rpc_counters = *cnt;
  83         spin_unlock(&srpc_data.rpc_glock);
  84 }
  85
  86 static int
  87 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
  88                    int nob)
  89 {
  90         LASSERT(off < PAGE_SIZE);
  91         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  92
  93         bk->bk_iovs[i].bv_offset = off;
  94         bk->bk_iovs[i].bv_page   = pg;
  95         bk->bk_iovs[i].bv_len    = nob;
  96         return nob;
  97 }
  98
  99 void
 100 srpc_free_bulk(struct srpc_bulk *bk)
 101 {
 102         int i;
 103         struct page *pg;
 104
 105         LASSERT(bk != NULL);
 106
 107         for (i = 0; i < bk->bk_niov; i++) {
 108                 pg = bk->bk_iovs[i].bv_page;
 109                 if (pg == NULL)
 110                         break;
 111
 112                 __free_page(pg);
 113         }
 114
 115         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
 116 }
 117
 118 struct srpc_bulk *
 119 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 120                 unsigned bulk_len, int sink)
 121 {
 122         struct srpc_bulk *bk;
 123         int i;
 124
 125         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 126
 127         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 128                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 129         if (bk == NULL) {
 130                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 131                 return NULL;
 132         }
 133
 134         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 135         bk->bk_sink   = sink;
 136         bk->bk_len    = bulk_len;
 137         bk->bk_niov   = bulk_npg;
 138
 139         for (i = 0; i < bulk_npg; i++) {
 140                 struct page *pg;
 141                 int nob;
 142
 143                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 144                 if (pg == NULL) {
 145                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 146                         srpc_free_bulk(bk);
 147                         return NULL;
 148                 }
 149
 150                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 151                       bulk_off;
 152
 153                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 154                 bulk_len -= nob;
 155                 bulk_off = 0;
 156         }
 157
 158         return bk;
 159 }
 160
 161 static inline __u64
 162 srpc_next_id (void)
 163 {
 164         __u64 id;
 165
 166         spin_lock(&srpc_data.rpc_glock);
 167         id = srpc_data.rpc_matchbits++;
 168         spin_unlock(&srpc_data.rpc_glock);
 169         return id;
 170 }
 171
 172 static void
 173 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 174                      struct srpc_service_cd *scd,
 175                      struct srpc_buffer *buffer)
 176 {
 177         memset(rpc, 0, sizeof(*rpc));
 178         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 179                           srpc_serv_is_framework(scd->scd_svc) ?
 180                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 181
 182         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 183
 184         rpc->srpc_scd      = scd;
 185         rpc->srpc_reqstbuf = buffer;
 186         rpc->srpc_peer     = buffer->buf_peer;
 187         rpc->srpc_self     = buffer->buf_self;
 188         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 189 }
 190
 191 static void
 192 srpc_service_fini(struct srpc_service *svc)
 193 {
 194         struct srpc_service_cd *scd;
 195         struct srpc_server_rpc *rpc;
 196         struct srpc_buffer *buf;
 197         struct list_head *q;
 198         int i;
 199
 200         if (svc->sv_cpt_data == NULL)
 201                 return;
 202
 203         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 204                 while (1) {
 205                         if (!list_empty(&scd->scd_buf_posted))
 206                                 q = &scd->scd_buf_posted;
 207                         else if (!list_empty(&scd->scd_buf_blocked))
 208                                 q = &scd->scd_buf_blocked;
 209                         else
 210                                 break;
 211
 212                         while (!list_empty(q)) {
 213                                 buf = list_entry(q->next,
 214                                                  struct srpc_buffer,
 215                                                  buf_list);
 216                                 list_del(&buf->buf_list);
 217                                 LIBCFS_FREE(buf, sizeof(*buf));
 218                         }
 219                 }
 220
 221                 LASSERT(list_empty(&scd->scd_rpc_active));
 222
 223                 while (!list_empty(&scd->scd_rpc_free)) {
 224                         rpc = list_entry(scd->scd_rpc_free.next,
 225                                          struct srpc_server_rpc,
 226                                          srpc_list);
 227                         list_del(&rpc->srpc_list);
 228                         LIBCFS_FREE(rpc, sizeof(*rpc));
 229                 }
 230         }
 231
 232         cfs_percpt_free(svc->sv_cpt_data);
 233         svc->sv_cpt_data = NULL;
 234 }
 235
 236 static int
 237 srpc_service_nrpcs(struct srpc_service *svc)
 238 {
 239         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 240
 241         return srpc_serv_is_framework(svc) ?
 242                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 243 }
 244
 245 int srpc_add_buffer(struct swi_workitem *wi);
 246
 247 static int
 248 srpc_service_init(struct srpc_service *svc)
 249 {
 250         struct srpc_service_cd *scd;
 251         struct srpc_server_rpc *rpc;
 252         int nrpcs;
 253         int i;
 254         int j;
 255
 256         svc->sv_shuttingdown = 0;
 257
 258         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 259                                             sizeof(struct srpc_service_cd));
 260         if (svc->sv_cpt_data == NULL)
 261                 return -ENOMEM;
 262
 263         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 264                         1 : cfs_cpt_number(lnet_cpt_table());
 265         nrpcs = srpc_service_nrpcs(svc);
 266
 267         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 268                 scd->scd_cpt = i;
 269                 scd->scd_svc = svc;
 270                 spin_lock_init(&scd->scd_lock);
 271                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 272                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 273                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 274                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 275
 276                 scd->scd_ev.ev_data = scd;
 277                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 278
 279                 /* NB: don't use lst_sched_serial for adding buffer,
 280                  * see details in srpc_service_add_buffers() */
 281                 swi_init_workitem(&scd->scd_buf_wi,
 282                                   srpc_add_buffer, lst_sched_test[i]);
 283
 284                 if (i != 0 && srpc_serv_is_framework(svc)) {
 285                         /* NB: framework service only needs srpc_service_cd for
 286                          * one partition, but we allocate for all to make
 287                          * it easier to implement, it will waste a little
 288                          * memory but nobody should care about this */
 289                         continue;
 290                 }
 291
 292                 for (j = 0; j < nrpcs; j++) {
 293                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 294                                          i, sizeof(*rpc));
 295                         if (rpc == NULL) {
 296                                 srpc_service_fini(svc);
 297                                 return -ENOMEM;
 298                         }
 299                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 300                 }
 301         }
 302
 303         return 0;
 304 }
 305
 306 int
 307 srpc_add_service(struct srpc_service *sv)
 308 {
 309         int id = sv->sv_id;
 310
 311         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 312
 313         if (srpc_service_init(sv) != 0)
 314                 return -ENOMEM;
 315
 316         spin_lock(&srpc_data.rpc_glock);
 317
 318         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 319
 320         if (srpc_data.rpc_services[id] != NULL) {
 321                 spin_unlock(&srpc_data.rpc_glock);
 322                 goto failed;
 323         }
 324
 325         srpc_data.rpc_services[id] = sv;
 326         spin_unlock(&srpc_data.rpc_glock);
 327
 328         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 329         return 0;
 330
 331 failed:
 332         srpc_service_fini(sv);
 333         return -EBUSY;
 334 }
 335
 336 int
 337 srpc_remove_service(struct srpc_service *sv)
 338 {
 339         int id = sv->sv_id;
 340
 341         spin_lock(&srpc_data.rpc_glock);
 342
 343         if (srpc_data.rpc_services[id] != sv) {
 344                 spin_unlock(&srpc_data.rpc_glock);
 345                 return -ENOENT;
 346         }
 347
 348         srpc_data.rpc_services[id] = NULL;
 349         spin_unlock(&srpc_data.rpc_glock);
 350         return 0;
 351 }
 352
 353 static int
 354 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 355                        int len, int options, struct lnet_process_id peer,
 356                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 357 {
 358         int rc;
 359         struct lnet_md md;
 360         struct lnet_me *me;
 361
 362         me = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
 363                           local ? LNET_INS_LOCAL : LNET_INS_AFTER);
 364         if (IS_ERR(me)) {
 365                 rc = PTR_ERR(me);
 366                 CERROR("LNetMEAttach failed: %d\n", rc);
 367                 LASSERT(rc == -ENOMEM);
 368                 return -ENOMEM;
 369         }
 370
 371         md.threshold = 1;
 372         md.user_ptr  = ev;
 373         md.start     = buf;
 374         md.length    = len;
 375         md.options   = options;
 376         md.handler   = srpc_data.rpc_lnet_handler;
 377
 378         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdh);
 379         if (rc != 0) {
 380                 CERROR("LNetMDAttach failed: %d\n", rc);
 381                 LASSERT(rc == -ENOMEM);
 382
 383                 return -ENOMEM;
 384         }
 385
 386         CDEBUG(D_NET,
 387                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 388                libcfs_id2str(peer), portal, matchbits);
 389         return 0;
 390 }
 391
 392 static int
 393 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 394                       int options, struct lnet_process_id peer,
 395                       lnet_nid_t self, struct lnet_handle_md *mdh,
 396                       struct srpc_event *ev)
 397 {
 398         int rc;
 399         struct lnet_md md;
 400
 401         md.user_ptr  = ev;
 402         md.start     = buf;
 403         md.length    = len;
 404         md.handler   = srpc_data.rpc_lnet_handler;
 405         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 406         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 407
 408         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
 409         if (rc != 0) {
 410                 CERROR("LNetMDBind failed: %d\n", rc);
 411                 LASSERT(rc == -ENOMEM);
 412                 return -ENOMEM;
 413         }
 414
 415         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 416          * they're only meaningful for MDs attached to an ME (i.e. passive
 417          * buffers...
 418          */
 419         if ((options & LNET_MD_OP_PUT) != 0) {
 420                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 421                              portal, matchbits, 0, 0);
 422         } else {
 423                 LASSERT((options & LNET_MD_OP_GET) != 0);
 424
 425                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0, false);
 426         }
 427
 428         if (rc != 0) {
 429                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 430                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 431                        libcfs_id2str(peer), portal, matchbits, rc);
 432
 433                 /* The forthcoming unlink event will complete this operation
 434                  * with failure, so fall through and return success here.
 435                  */
 436                 rc = LNetMDUnlink(*mdh);
 437                 LASSERT(rc == 0);
 438         } else {
 439                 CDEBUG(D_NET,
 440                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 441                        libcfs_id2str(peer), portal, matchbits);
 442         }
 443         return 0;
 444 }
 445
 446 static int
 447 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 448                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 449 {
 450         struct lnet_process_id any = {0};
 451
 452         any.nid = LNET_NID_ANY;
 453         any.pid = LNET_PID_ANY;
 454
 455         return srpc_post_passive_rdma(srpc_serv_portal(service),
 456                                       local, service, buf, len,
 457                                       LNET_MD_OP_PUT, any, mdh, ev);
 458 }
 459
 460 static int
 461 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 462 __must_hold(&scd->scd_lock)
 463 {
 464         struct srpc_service *sv = scd->scd_svc;
 465         struct srpc_msg *msg = &buf->buf_msg;
 466         int rc;
 467
 468         LNetInvalidateMDHandle(&buf->buf_mdh);
 469         list_add(&buf->buf_list, &scd->scd_buf_posted);
 470         scd->scd_buf_nposted++;
 471         spin_unlock(&scd->scd_lock);
 472
 473         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 474                                       !srpc_serv_is_framework(sv),
 475                                       msg, sizeof(*msg), &buf->buf_mdh,
 476                                       &scd->scd_ev);
 477
 478         /* At this point, a RPC (new or delayed) may have arrived in
 479          * msg and its event handler has been called. So we must add
 480          * buf to scd_buf_posted _before_ dropping scd_lock */
 481
 482         spin_lock(&scd->scd_lock);
 483
 484         if (rc == 0) {
 485                 if (!sv->sv_shuttingdown)
 486                         return 0;
 487
 488                 spin_unlock(&scd->scd_lock);
 489                 /* srpc_shutdown_service might have tried to unlink me
 490                  * when my buf_mdh was still invalid */
 491                 LNetMDUnlink(buf->buf_mdh);
 492                 spin_lock(&scd->scd_lock);
 493                 return 0;
 494         }
 495
 496         scd->scd_buf_nposted--;
 497         if (sv->sv_shuttingdown)
 498                 return rc; /* don't allow to change scd_buf_posted */
 499
 500         list_del(&buf->buf_list);
 501         spin_unlock(&scd->scd_lock);
 502
 503         LIBCFS_FREE(buf, sizeof(*buf));
 504
 505         spin_lock(&scd->scd_lock);
 506         return rc;
 507 }
 508
 509 int
 510 srpc_add_buffer(struct swi_workitem *wi)
 511 {
 512         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 513                                                    scd_buf_wi);
 514         struct srpc_buffer *buf;
 515         int rc = 0;
 516
 517         /* it's called by workitem scheduler threads, these threads
 518          * should have been set CPT affinity, so buffers will be posted
 519          * on CPT local list of Portal */
 520         spin_lock(&scd->scd_lock);
 521
 522         while (scd->scd_buf_adjust > 0 &&
 523                !scd->scd_svc->sv_shuttingdown) {
 524                 scd->scd_buf_adjust--; /* consume it */
 525                 scd->scd_buf_posting++;
 526
 527                 spin_unlock(&scd->scd_lock);
 528
 529                 LIBCFS_ALLOC(buf, sizeof(*buf));
 530                 if (buf == NULL) {
 531                         CERROR("Failed to add new buf to service: %s\n",
 532                                scd->scd_svc->sv_name);
 533                         spin_lock(&scd->scd_lock);
 534                         rc = -ENOMEM;
 535                         break;
 536                 }
 537
 538                 spin_lock(&scd->scd_lock);
 539                 if (scd->scd_svc->sv_shuttingdown) {
 540                         spin_unlock(&scd->scd_lock);
 541                         LIBCFS_FREE(buf, sizeof(*buf));
 542
 543                         spin_lock(&scd->scd_lock);
 544                         rc = -ESHUTDOWN;
 545                         break;
 546                 }
 547
 548                 rc = srpc_service_post_buffer(scd, buf);
 549                 if (rc != 0)
 550                         break; /* buf has been freed inside */
 551
 552                 LASSERT(scd->scd_buf_posting > 0);
 553                 scd->scd_buf_posting--;
 554                 scd->scd_buf_total++;
 555                 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
 556         }
 557
 558         if (rc != 0) {
 559                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 560                 scd->scd_buf_err = rc;
 561
 562                 LASSERT(scd->scd_buf_posting > 0);
 563                 scd->scd_buf_posting--;
 564         }
 565
 566         spin_unlock(&scd->scd_lock);
 567         return 0;
 568 }
 569
 570 int
 571 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 572 {
 573         struct srpc_service_cd *scd;
 574         int rc = 0;
 575         int i;
 576
 577         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 578
 579         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 580                 spin_lock(&scd->scd_lock);
 581
 582                 scd->scd_buf_err = 0;
 583                 scd->scd_buf_err_stamp = 0;
 584                 scd->scd_buf_posting = 0;
 585                 scd->scd_buf_adjust = nbuffer;
 586                 /* start to post buffers */
 587                 swi_schedule_workitem(&scd->scd_buf_wi);
 588                 spin_unlock(&scd->scd_lock);
 589
 590                 /* framework service only post buffer for one partition  */
 591                 if (srpc_serv_is_framework(sv))
 592                         break;
 593         }
 594
 595         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 596                 spin_lock(&scd->scd_lock);
 597                 /*
 598                  * NB: srpc_service_add_buffers() can be called inside
 599                  * thread context of lst_sched_serial, and we don't normally
 600                  * allow to sleep inside thread context of WI scheduler
 601                  * because it will block current scheduler thread from doing
 602                  * anything else, even worse, it could deadlock if it's
 603                  * waiting on result from another WI of the same scheduler.
 604                  * However, it's safe at here because scd_buf_wi is scheduled
 605                  * by thread in a different WI scheduler (lst_sched_test),
 606                  * so we don't have any risk of deadlock, though this could
 607                  * block all WIs pending on lst_sched_serial for a moment
 608                  * which is not good but not fatal.
 609                  */
 610                 lst_wait_until(scd->scd_buf_err != 0 ||
 611                                (scd->scd_buf_adjust == 0 &&
 612                                 scd->scd_buf_posting == 0),
 613                                scd->scd_lock, "waiting for adding buffer\n");
 614
 615                 if (scd->scd_buf_err != 0 && rc == 0)
 616                         rc = scd->scd_buf_err;
 617
 618                 spin_unlock(&scd->scd_lock);
 619         }
 620
 621         return rc;
 622 }
 623
 624 void
 625 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 626 {
 627         struct srpc_service_cd *scd;
 628         int num;
 629         int i;
 630
 631         LASSERT(!sv->sv_shuttingdown);
 632
 633         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 634                 spin_lock(&scd->scd_lock);
 635
 636                 num = scd->scd_buf_total + scd->scd_buf_posting;
 637                 scd->scd_buf_adjust -= min(nbuffer, num);
 638
 639                 spin_unlock(&scd->scd_lock);
 640         }
 641 }
 642
 643 /* returns 1 if sv has finished, otherwise 0 */
 644 int
 645 srpc_finish_service(struct srpc_service *sv)
 646 {
 647         struct srpc_service_cd *scd;
 648         struct srpc_server_rpc *rpc;
 649         int i;
 650
 651         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 652
 653         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 654                 spin_lock(&scd->scd_lock);
 655                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 656                         spin_unlock(&scd->scd_lock);
 657                         return 0;
 658                 }
 659
 660                 if (scd->scd_buf_nposted > 0) {
 661                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 662                                scd->scd_buf_nposted);
 663                         spin_unlock(&scd->scd_lock);
 664                         return 0;
 665                 }
 666
 667                 if (list_empty(&scd->scd_rpc_active)) {
 668                         spin_unlock(&scd->scd_lock);
 669                         continue;
 670                 }
 671
 672                 rpc = list_entry(scd->scd_rpc_active.next,
 673                                  struct srpc_server_rpc, srpc_list);
 674                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, wi %s scheduled %d running %d, ev fired %d type %d status %d lnet %d\n",
 675                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 676                         swi_state2str(rpc->srpc_wi.swi_state),
 677                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 678                         rpc->srpc_wi.swi_workitem.wi_running,
 679                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 680                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 681                 spin_unlock(&scd->scd_lock);
 682                 return 0;
 683         }
 684
 685         /* no lock needed from now on */
 686         srpc_service_fini(sv);
 687         return 1;
 688 }
 689
 690 /* called with sv->sv_lock held */
 691 static void
 692 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 693                             struct srpc_buffer *buf)
 694 __must_hold(&scd->scd_lock)
 695 {
 696         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 697                 if (srpc_service_post_buffer(scd, buf) != 0) {
 698                         CWARN("Failed to post %s buffer\n",
 699                               scd->scd_svc->sv_name);
 700                 }
 701                 return;
 702         }
 703
 704         /* service is shutting down, or we want to recycle some buffers */
 705         scd->scd_buf_total--;
 706
 707         if (scd->scd_buf_adjust < 0) {
 708                 scd->scd_buf_adjust++;
 709                 if (scd->scd_buf_adjust < 0 &&
 710                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 711                         CDEBUG(D_INFO,
 712                                "Try to recyle %d buffers but nothing left\n",
 713                                scd->scd_buf_adjust);
 714                         scd->scd_buf_adjust = 0;
 715                 }
 716         }
 717
 718         spin_unlock(&scd->scd_lock);
 719         LIBCFS_FREE(buf, sizeof(*buf));
 720         spin_lock(&scd->scd_lock);
 721 }
 722
 723 void
 724 srpc_abort_service(struct srpc_service *sv)
 725 {
 726         struct srpc_service_cd *scd;
 727         struct srpc_server_rpc *rpc;
 728         int i;
 729
 730         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 731                sv->sv_id, sv->sv_name);
 732
 733         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 734                 spin_lock(&scd->scd_lock);
 735
 736                 /* schedule in-flight RPCs to notice the abort, NB:
 737                  * racing with incoming RPCs; complete fix should make test
 738                  * RPCs carry session ID in its headers
 739                  */
 740                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 741                         rpc->srpc_aborted = 1;
 742                         swi_schedule_workitem(&rpc->srpc_wi);
 743                 }
 744
 745                 spin_unlock(&scd->scd_lock);
 746         }
 747 }
 748
 749 void
 750 srpc_shutdown_service(struct srpc_service *sv)
 751 {
 752         struct srpc_service_cd *scd;
 753         struct srpc_server_rpc *rpc;
 754         struct srpc_buffer *buf;
 755         int i;
 756
 757         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 758                sv->sv_id, sv->sv_name);
 759
 760         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 761                 spin_lock(&scd->scd_lock);
 762
 763         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 764
 765         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 766                 spin_unlock(&scd->scd_lock);
 767
 768         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 769                 spin_lock(&scd->scd_lock);
 770
 771                 /* schedule in-flight RPCs to notice the shutdown */
 772                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 773                         swi_schedule_workitem(&rpc->srpc_wi);
 774
 775                 spin_unlock(&scd->scd_lock);
 776
 777                 /* OK to traverse scd_buf_posted without lock, since no one
 778                  * touches scd_buf_posted now
 779                  */
 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(struct srpc_client_rpc *rpc)
 787 {
 788         struct srpc_event *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(struct srpc_msg), 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(struct srpc_client_rpc *rpc)
 809 {
 810         struct srpc_event *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,
 822                                     sizeof(struct srpc_msg),
 823                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 824                                     &rpc->crpc_replymdh, ev);
 825         if (rc != 0) {
 826                 LASSERT(rc == -ENOMEM);
 827                 ev->ev_fired = 1;  /* no more event expected */
 828         }
 829         return rc;
 830 }
 831
 832 static int
 833 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 834 {
 835         struct srpc_bulk *bk = &rpc->crpc_bulk;
 836         struct srpc_event *ev = &rpc->crpc_bulkev;
 837         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 838         int rc;
 839         int opt;
 840
 841         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 842
 843         /* nothing to do */
 844         if (bk->bk_niov == 0)
 845                 return 0;
 846
 847         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 848         opt |= LNET_MD_KIOV;
 849
 850         ev->ev_fired = 0;
 851         ev->ev_data  = rpc;
 852         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 853
 854         *id = srpc_next_id();
 855
 856         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 857                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 858                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 859         if (rc != 0) {
 860                 LASSERT(rc == -ENOMEM);
 861                 ev->ev_fired = 1;  /* no more event expected */
 862         }
 863         return rc;
 864 }
 865
 866 static int
 867 srpc_do_bulk(struct srpc_server_rpc *rpc)
 868 {
 869         struct srpc_event *ev = &rpc->srpc_ev;
 870         struct srpc_bulk *bk = rpc->srpc_bulk;
 871         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 872         int rc;
 873         int opt;
 874
 875         LASSERT(bk != NULL);
 876
 877         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 878         opt |= LNET_MD_KIOV;
 879
 880         ev->ev_fired = 0;
 881         ev->ev_data  = rpc;
 882         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 883
 884         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 885                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 886                                    rpc->srpc_peer, rpc->srpc_self,
 887                                    &bk->bk_mdh, ev);
 888         if (rc != 0)
 889                 ev->ev_fired = 1;  /* no more event expected */
 890         return rc;
 891 }
 892
 893 /* only called from srpc_handle_rpc */
 894 static void
 895 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 896 {
 897         struct srpc_service_cd  *scd = rpc->srpc_scd;
 898         struct srpc_service     *sv  = scd->scd_svc;
 899         struct srpc_buffer *buffer;
 900
 901         LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 902
 903         rpc->srpc_status = status;
 904
 905         CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
 906                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 907                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 908                      swi_state2str(rpc->srpc_wi.swi_state), status);
 909
 910         if (status != 0) {
 911                 spin_lock(&srpc_data.rpc_glock);
 912                 srpc_data.rpc_counters.rpcs_dropped++;
 913                 spin_unlock(&srpc_data.rpc_glock);
 914         }
 915
 916         if (rpc->srpc_done != NULL)
 917                 (*rpc->srpc_done) (rpc);
 918         LASSERT(rpc->srpc_bulk == NULL);
 919
 920         spin_lock(&scd->scd_lock);
 921
 922         if (rpc->srpc_reqstbuf != NULL) {
 923                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 924                  * sv won't go away for scd_rpc_active must not be empty
 925                  */
 926                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 927                 rpc->srpc_reqstbuf = NULL;
 928         }
 929
 930         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 931
 932         /*
 933          * No one can schedule me now since:
 934          * - I'm not on scd_rpc_active.
 935          * - all LNet events have been fired.
 936          * Cancel pending schedules and prevent future schedule attempts:
 937          */
 938         LASSERT(rpc->srpc_ev.ev_fired);
 939         swi_exit_workitem(&rpc->srpc_wi);
 940
 941         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 942                 buffer = list_entry(scd->scd_buf_blocked.next,
 943                                     struct srpc_buffer, buf_list);
 944                 list_del(&buffer->buf_list);
 945
 946                 srpc_init_server_rpc(rpc, scd, buffer);
 947                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 948                 swi_schedule_workitem(&rpc->srpc_wi);
 949         } else {
 950                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 951         }
 952
 953         spin_unlock(&scd->scd_lock);
 954 }
 955
 956 /* handles an incoming RPC */
 957 static int srpc_handle_rpc(struct swi_workitem *wi)
 958 {
 959         struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc,
 960                                                    srpc_wi);
 961         struct srpc_service_cd *scd = rpc->srpc_scd;
 962         struct srpc_service *sv = scd->scd_svc;
 963         struct srpc_event *ev = &rpc->srpc_ev;
 964         int rc = 0;
 965
 966         LASSERT(wi == &rpc->srpc_wi);
 967
 968         spin_lock(&scd->scd_lock);
 969
 970         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 971                 spin_unlock(&scd->scd_lock);
 972
 973                 if (rpc->srpc_bulk != NULL)
 974                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 975                 LNetMDUnlink(rpc->srpc_replymdh);
 976
 977                 if (ev->ev_fired) { /* no more event, OK to finish */
 978                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 979                         return 1;
 980                 }
 981                 return 0;
 982         }
 983
 984         spin_unlock(&scd->scd_lock);
 985
 986         switch (wi->swi_state) {
 987         default:
 988                 LBUG();
 989                 /* fallthrough */
 990         case SWI_STATE_NEWBORN: {
 991                 struct srpc_msg *msg;
 992                 struct srpc_generic_reply *reply;
 993
 994                 msg = &rpc->srpc_reqstbuf->buf_msg;
 995                 reply = &rpc->srpc_replymsg.msg_body.reply;
 996
 997                 if (msg->msg_magic == 0) {
 998                         /* moaned already in srpc_lnet_ev_handler */
 999                         srpc_server_rpc_done(rpc, EBADMSG);
1000                         return 1;
1001                 }
1002
1003                 srpc_unpack_msg_hdr(msg);
1004                 if (msg->msg_version != SRPC_MSG_VERSION) {
1005                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1006                               msg->msg_version, SRPC_MSG_VERSION,
1007                               libcfs_id2str(rpc->srpc_peer));
1008                         reply->status = EPROTO;
1009                         /* drop through and send reply */
1010                 } else {
1011                         reply->status = 0;
1012                         rc = (*sv->sv_handler)(rpc);
1013                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1014                         if (rc != 0) {
1015                                 srpc_server_rpc_done(rpc, rc);
1016                                 return 1;
1017                         }
1018                 }
1019
1020                 wi->swi_state = SWI_STATE_BULK_STARTED;
1021
1022                 if (rpc->srpc_bulk != NULL) {
1023                         rc = srpc_do_bulk(rpc);
1024                         if (rc == 0)
1025                                 return 0; /* wait for bulk */
1026
1027                         LASSERT(ev->ev_fired);
1028                         ev->ev_status = rc;
1029                 }
1030         }
1031         /* fallthrough */
1032         case SWI_STATE_BULK_STARTED:
1033                 LASSERT(rpc->srpc_bulk == NULL || ev->ev_fired);
1034
1035                 if (rpc->srpc_bulk != NULL) {
1036                         rc = ev->ev_status;
1037
1038                         if (sv->sv_bulk_ready != NULL)
1039                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1040
1041                         if (rc != 0) {
1042                                 srpc_server_rpc_done(rpc, rc);
1043                                 return 1;
1044                         }
1045                 }
1046
1047                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1048                 rc = srpc_send_reply(rpc);
1049                 if (rc == 0)
1050                         return 0; /* wait for reply */
1051                 srpc_server_rpc_done(rpc, rc);
1052                 return 1;
1053
1054         case SWI_STATE_REPLY_SUBMITTED:
1055                 if (!ev->ev_fired) {
1056                         CERROR("RPC %p: bulk %p, service %d\n",
1057                                rpc, rpc->srpc_bulk, sv->sv_id);
1058                         CERROR("Event: status %d, type %d, lnet %d\n",
1059                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1060                         LASSERT(ev->ev_fired);
1061                 }
1062
1063                 wi->swi_state = SWI_STATE_DONE;
1064                 srpc_server_rpc_done(rpc, ev->ev_status);
1065                 return 1;
1066         }
1067
1068         return 0;
1069 }
1070
1071 static void
1072 srpc_client_rpc_expired (void *data)
1073 {
1074         struct srpc_client_rpc *rpc = data;
1075
1076         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1077               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1078               rpc->crpc_timeout);
1079
1080         spin_lock(&rpc->crpc_lock);
1081
1082         rpc->crpc_timeout = 0;
1083         srpc_abort_rpc(rpc, -ETIMEDOUT);
1084
1085         spin_unlock(&rpc->crpc_lock);
1086
1087         spin_lock(&srpc_data.rpc_glock);
1088         srpc_data.rpc_counters.rpcs_expired++;
1089         spin_unlock(&srpc_data.rpc_glock);
1090 }
1091
1092 static void
1093 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1094 {
1095         struct stt_timer *timer = &rpc->crpc_timer;
1096
1097         if (rpc->crpc_timeout == 0)
1098                 return;
1099
1100         INIT_LIST_HEAD(&timer->stt_list);
1101         timer->stt_data    = rpc;
1102         timer->stt_func    = srpc_client_rpc_expired;
1103         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1104         stt_add_timer(timer);
1105 }
1106
1107 /*
1108  * Called with rpc->crpc_lock held.
1109  *
1110  * Upon exit the RPC expiry timer is not queued and the handler is not
1111  * running on any CPU.
1112  */
1113 static void
1114 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1115 {
1116         /* timer not planted or already exploded */
1117         if (rpc->crpc_timeout == 0)
1118                 return;
1119
1120         /* timer successfully defused */
1121         if (stt_del_timer(&rpc->crpc_timer))
1122                 return;
1123
1124         /* timer detonated, wait for it to explode */
1125         while (rpc->crpc_timeout != 0) {
1126                 spin_unlock(&rpc->crpc_lock);
1127
1128                 schedule();
1129
1130                 spin_lock(&rpc->crpc_lock);
1131         }
1132 }
1133
1134 static void
1135 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1136 {
1137         struct swi_workitem *wi = &rpc->crpc_wi;
1138
1139         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1140
1141         spin_lock(&rpc->crpc_lock);
1142
1143         rpc->crpc_closed = 1;
1144         if (rpc->crpc_status == 0)
1145                 rpc->crpc_status = status;
1146
1147         srpc_del_client_rpc_timer(rpc);
1148
1149         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1150                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1151                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1152                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1153
1154         /*
1155          * No one can schedule me now since:
1156          * - RPC timer has been defused.
1157          * - all LNet events have been fired.
1158          * - crpc_closed has been set, preventing srpc_abort_rpc from
1159          *   scheduling me.
1160          * Cancel pending schedules and prevent future schedule attempts:
1161          */
1162         LASSERT(!srpc_event_pending(rpc));
1163         swi_exit_workitem(wi);
1164
1165         spin_unlock(&rpc->crpc_lock);
1166
1167         (*rpc->crpc_done)(rpc);
1168 }
1169
1170 /* sends an outgoing RPC */
1171 int
1172 srpc_send_rpc(struct swi_workitem *wi)
1173 {
1174         int rc = 0;
1175         struct srpc_client_rpc *rpc;
1176         struct srpc_msg *reply;
1177         int do_bulk;
1178
1179         LASSERT(wi != NULL);
1180
1181         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1182
1183         LASSERT(rpc != NULL);
1184         LASSERT(wi == &rpc->crpc_wi);
1185
1186         reply = &rpc->crpc_replymsg;
1187         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1188
1189         spin_lock(&rpc->crpc_lock);
1190
1191         if (rpc->crpc_aborted) {
1192                 spin_unlock(&rpc->crpc_lock);
1193                 goto abort;
1194         }
1195
1196         spin_unlock(&rpc->crpc_lock);
1197
1198         switch (wi->swi_state) {
1199         default:
1200                 LBUG();
1201         case SWI_STATE_NEWBORN:
1202                 LASSERT(!srpc_event_pending(rpc));
1203
1204                 rc = srpc_prepare_reply(rpc);
1205                 if (rc != 0) {
1206                         srpc_client_rpc_done(rpc, rc);
1207                         return 1;
1208                 }
1209
1210                 rc = srpc_prepare_bulk(rpc);
1211                 if (rc != 0)
1212                         break;
1213
1214                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1215                 rc = srpc_send_request(rpc);
1216                 break;
1217
1218         case SWI_STATE_REQUEST_SUBMITTED:
1219                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1220                  * order; however, they're processed in a strict order:
1221                  * rqt, rpy, and bulk.
1222                  */
1223                 if (!rpc->crpc_reqstev.ev_fired)
1224                         break;
1225
1226                 rc = rpc->crpc_reqstev.ev_status;
1227                 if (rc != 0)
1228                         break;
1229
1230                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1231                 /* fallthrough */
1232         case SWI_STATE_REQUEST_SENT: {
1233                 enum srpc_msg_type type;
1234
1235                 type = srpc_service2reply(rpc->crpc_service);
1236
1237                 if (!rpc->crpc_replyev.ev_fired)
1238                         break;
1239
1240                 rc = rpc->crpc_replyev.ev_status;
1241                 if (rc != 0)
1242                         break;
1243
1244                 srpc_unpack_msg_hdr(reply);
1245                 if (reply->msg_type != type ||
1246                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1247                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1248                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1249                               libcfs_id2str(rpc->crpc_dest),
1250                               reply->msg_type, type,
1251                               reply->msg_magic, SRPC_MSG_MAGIC);
1252                         rc = -EBADMSG;
1253                         break;
1254                 }
1255
1256                 if (do_bulk && reply->msg_body.reply.status != 0) {
1257                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1258                               reply->msg_body.reply.status,
1259                               libcfs_id2str(rpc->crpc_dest));
1260                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1261                 }
1262
1263                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1264         }
1265         /* fallthrough */
1266         case SWI_STATE_REPLY_RECEIVED:
1267                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1268                         break;
1269
1270                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1271
1272                 /* Bulk buffer was unlinked due to remote error. Clear error
1273                  * since reply buffer still contains valid data.
1274                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1275                  * remote error.
1276                  */
1277                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1278                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1279                         rc = 0;
1280
1281                 wi->swi_state = SWI_STATE_DONE;
1282                 srpc_client_rpc_done(rpc, rc);
1283                 return 1;
1284         }
1285
1286         if (rc != 0) {
1287                 spin_lock(&rpc->crpc_lock);
1288                 srpc_abort_rpc(rpc, rc);
1289                 spin_unlock(&rpc->crpc_lock);
1290         }
1291
1292 abort:
1293         if (rpc->crpc_aborted) {
1294                 LNetMDUnlink(rpc->crpc_reqstmdh);
1295                 LNetMDUnlink(rpc->crpc_replymdh);
1296                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1297
1298                 if (!srpc_event_pending(rpc)) {
1299                         srpc_client_rpc_done(rpc, -EINTR);
1300                         return 1;
1301                 }
1302         }
1303         return 0;
1304 }
1305
1306 struct srpc_client_rpc *
1307 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1308                        int nbulkiov, int bulklen,
1309                        void (*rpc_done)(struct srpc_client_rpc *),
1310                        void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1311 {
1312         struct srpc_client_rpc *rpc;
1313
1314         LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1315                                    crpc_bulk.bk_iovs[nbulkiov]));
1316         if (rpc == NULL)
1317                 return NULL;
1318
1319         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1320                              bulklen, rpc_done, rpc_fini, priv);
1321         return rpc;
1322 }
1323
1324 /* called with rpc->crpc_lock held */
1325 void
1326 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1327 {
1328         LASSERT(why != 0);
1329
1330         if (rpc->crpc_aborted || /* already aborted */
1331             rpc->crpc_closed)    /* callback imminent */
1332                 return;
1333
1334         CDEBUG(D_NET,
1335                "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1336                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1337                swi_state2str(rpc->crpc_wi.swi_state), why);
1338
1339         rpc->crpc_aborted = 1;
1340         rpc->crpc_status  = why;
1341         swi_schedule_workitem(&rpc->crpc_wi);
1342 }
1343
1344 /* called with rpc->crpc_lock held */
1345 void
1346 srpc_post_rpc(struct srpc_client_rpc *rpc)
1347 {
1348         LASSERT(!rpc->crpc_aborted);
1349         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1350
1351         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1352                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1353                rpc->crpc_timeout);
1354
1355         srpc_add_client_rpc_timer(rpc);
1356         swi_schedule_workitem(&rpc->crpc_wi);
1357 }
1358
1359
1360 int
1361 srpc_send_reply(struct srpc_server_rpc *rpc)
1362 {
1363         struct srpc_event *ev = &rpc->srpc_ev;
1364         struct srpc_msg *msg = &rpc->srpc_replymsg;
1365         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1366         struct srpc_service_cd *scd = rpc->srpc_scd;
1367         struct srpc_service *sv = scd->scd_svc;
1368         __u64 rpyid;
1369         int rc;
1370
1371         LASSERT(buffer != NULL);
1372         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1373
1374         spin_lock(&scd->scd_lock);
1375
1376         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1377                 /* Repost buffer before replying since test client
1378                  * might send me another RPC once it gets the reply
1379                  */
1380                 if (srpc_service_post_buffer(scd, buffer) != 0)
1381                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1382                 rpc->srpc_reqstbuf = NULL;
1383         }
1384
1385         spin_unlock(&scd->scd_lock);
1386
1387         ev->ev_fired = 0;
1388         ev->ev_data  = rpc;
1389         ev->ev_type  = SRPC_REPLY_SENT;
1390
1391         msg->msg_magic   = SRPC_MSG_MAGIC;
1392         msg->msg_version = SRPC_MSG_VERSION;
1393         msg->msg_type    = srpc_service2reply(sv->sv_id);
1394
1395         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1396                                    sizeof(*msg), LNET_MD_OP_PUT,
1397                                    rpc->srpc_peer, rpc->srpc_self,
1398                                    &rpc->srpc_replymdh, ev);
1399         if (rc != 0)
1400                 ev->ev_fired = 1;  /* no more event expected */
1401         return rc;
1402 }
1403
1404 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1405 static void
1406 srpc_lnet_ev_handler(struct lnet_event *ev)
1407 {
1408         struct srpc_service_cd *scd;
1409         struct srpc_event *rpcev = ev->md_user_ptr;
1410         struct srpc_client_rpc *crpc;
1411         struct srpc_server_rpc *srpc;
1412         struct srpc_buffer *buffer;
1413         struct srpc_service *sv;
1414         struct srpc_msg *msg;
1415         enum srpc_msg_type type;
1416
1417         LASSERT(!in_interrupt());
1418
1419         if (ev->status != 0) {
1420                 __u32 errors;
1421
1422                 spin_lock(&srpc_data.rpc_glock);
1423                 if (ev->status != -ECANCELED) /* cancellation is not error */
1424                         srpc_data.rpc_counters.errors++;
1425                 errors = srpc_data.rpc_counters.errors;
1426                 spin_unlock(&srpc_data.rpc_glock);
1427
1428                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1429                         ev->status, ev->type, errors);
1430         }
1431
1432         rpcev->ev_lnet = ev->type;
1433
1434         switch (rpcev->ev_type) {
1435         default:
1436                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1437                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1438                 LBUG();
1439                 /* fallthrough */
1440         case SRPC_REQUEST_SENT:
1441                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1442                         spin_lock(&srpc_data.rpc_glock);
1443                         srpc_data.rpc_counters.rpcs_sent++;
1444                         spin_unlock(&srpc_data.rpc_glock);
1445                 }
1446                 /* fallthrough */
1447         case SRPC_REPLY_RCVD:
1448         case SRPC_BULK_REQ_RCVD:
1449                 crpc = rpcev->ev_data;
1450
1451                 if (rpcev != &crpc->crpc_reqstev &&
1452                     rpcev != &crpc->crpc_replyev &&
1453                     rpcev != &crpc->crpc_bulkev) {
1454                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1455                                rpcev, crpc, &crpc->crpc_reqstev,
1456                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1457                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1458                                rpcev->ev_status, rpcev->ev_type,
1459                                rpcev->ev_lnet);
1460                         LBUG();
1461                 }
1462
1463                 spin_lock(&crpc->crpc_lock);
1464
1465                 LASSERT(rpcev->ev_fired == 0);
1466                 rpcev->ev_fired  = 1;
1467                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1468                                    -EINTR : ev->status;
1469                 swi_schedule_workitem(&crpc->crpc_wi);
1470
1471                 spin_unlock(&crpc->crpc_lock);
1472                 break;
1473
1474         case SRPC_REQUEST_RCVD:
1475                 scd = rpcev->ev_data;
1476                 sv = scd->scd_svc;
1477
1478                 LASSERT(rpcev == &scd->scd_ev);
1479
1480                 spin_lock(&scd->scd_lock);
1481
1482                 LASSERT(ev->unlinked);
1483                 LASSERT(ev->type == LNET_EVENT_PUT ||
1484                         ev->type == LNET_EVENT_UNLINK);
1485                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1486                         sv->sv_shuttingdown);
1487
1488                 buffer = container_of(ev->md_start, struct srpc_buffer,
1489                                       buf_msg);
1490                 buffer->buf_peer = ev->source;
1491                 buffer->buf_self = ev->target.nid;
1492
1493                 LASSERT(scd->scd_buf_nposted > 0);
1494                 scd->scd_buf_nposted--;
1495
1496                 if (sv->sv_shuttingdown) {
1497                         /* Leave buffer on scd->scd_buf_nposted since
1498                          * srpc_finish_service needs to traverse it.
1499                          */
1500                         spin_unlock(&scd->scd_lock);
1501                         break;
1502                 }
1503
1504                 if (scd->scd_buf_err_stamp != 0 &&
1505                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1506                         /* re-enable adding buffer */
1507                         scd->scd_buf_err_stamp = 0;
1508                         scd->scd_buf_err = 0;
1509                 }
1510
1511                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1512                     scd->scd_buf_adjust == 0 &&
1513                     scd->scd_buf_nposted < scd->scd_buf_low) {
1514                         scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1515                                                   SFW_TEST_WI_MIN);
1516                         swi_schedule_workitem(&scd->scd_buf_wi);
1517                 }
1518
1519                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1520                 msg = &buffer->buf_msg;
1521                 type = srpc_service2request(sv->sv_id);
1522
1523                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1524                     (msg->msg_type != type &&
1525                      msg->msg_type != __swab32(type)) ||
1526                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1527                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1528                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1529                                sv->sv_name, libcfs_id2str(ev->initiator),
1530                                ev->status, ev->mlength,
1531                                msg->msg_type, msg->msg_magic);
1532
1533                         /* NB can't call srpc_service_recycle_buffer here since
1534                          * it may call LNetM[DE]Attach. The invalid magic tells
1535                          * srpc_handle_rpc to drop this RPC
1536                          */
1537                         msg->msg_magic = 0;
1538                 }
1539
1540                 if (!list_empty(&scd->scd_rpc_free)) {
1541                         srpc = list_entry(scd->scd_rpc_free.next,
1542                                           struct srpc_server_rpc,
1543                                           srpc_list);
1544                         list_del(&srpc->srpc_list);
1545
1546                         srpc_init_server_rpc(srpc, scd, buffer);
1547                         list_add_tail(&srpc->srpc_list,
1548                                       &scd->scd_rpc_active);
1549                         swi_schedule_workitem(&srpc->srpc_wi);
1550                 } else {
1551                         list_add_tail(&buffer->buf_list,
1552                                       &scd->scd_buf_blocked);
1553                 }
1554
1555                 spin_unlock(&scd->scd_lock);
1556
1557                 spin_lock(&srpc_data.rpc_glock);
1558                 srpc_data.rpc_counters.rpcs_rcvd++;
1559                 spin_unlock(&srpc_data.rpc_glock);
1560                 break;
1561
1562         case SRPC_BULK_GET_RPLD:
1563                 LASSERT(ev->type == LNET_EVENT_SEND ||
1564                         ev->type == LNET_EVENT_REPLY ||
1565                         ev->type == LNET_EVENT_UNLINK);
1566
1567                 if (!ev->unlinked)
1568                         break; /* wait for final event */
1569                 /* fallthrough */
1570         case SRPC_BULK_PUT_SENT:
1571                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1572                         spin_lock(&srpc_data.rpc_glock);
1573
1574                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1575                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1576                         else
1577                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1578
1579                         spin_unlock(&srpc_data.rpc_glock);
1580                 }
1581                 /* fallthrough */
1582         case SRPC_REPLY_SENT:
1583                 srpc = rpcev->ev_data;
1584                 scd  = srpc->srpc_scd;
1585
1586                 LASSERT(rpcev == &srpc->srpc_ev);
1587
1588                 spin_lock(&scd->scd_lock);
1589
1590                 rpcev->ev_fired  = 1;
1591                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1592                                    -EINTR : ev->status;
1593                 swi_schedule_workitem(&srpc->srpc_wi);
1594
1595                 spin_unlock(&scd->scd_lock);
1596                 break;
1597         }
1598 }
1599
1600
1601 int
1602 srpc_startup (void)
1603 {
1604         int rc;
1605
1606         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1607         spin_lock_init(&srpc_data.rpc_glock);
1608
1609         /* 1 second pause to avoid timestamp reuse */
1610         schedule_timeout_uninterruptible(cfs_time_seconds(1));
1611         srpc_data.rpc_matchbits = ((__u64) ktime_get_real_seconds()) << 48;
1612
1613         srpc_data.rpc_state = SRPC_STATE_NONE;
1614
1615         rc = LNetNIInit(LNET_PID_LUSTRE);
1616         if (rc < 0) {
1617                 CERROR("LNetNIInit() has failed: %d\n", rc);
1618                 return rc;
1619         }
1620
1621         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1622
1623         srpc_data.rpc_lnet_handler = srpc_lnet_ev_handler;
1624
1625         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1626         LASSERT(rc == 0);
1627         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1628         LASSERT(rc == 0);
1629
1630         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1631
1632         rc = stt_startup();
1633
1634         if (rc != 0)
1635                 srpc_shutdown();
1636         else
1637                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1638
1639         return rc;
1640 }
1641
1642 void
1643 srpc_shutdown (void)
1644 {
1645         int i;
1646         int rc;
1647         int state;
1648
1649         state = srpc_data.rpc_state;
1650         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1651
1652         switch (state) {
1653         default:
1654                 LBUG();
1655                 /* fallthrough */
1656         case SRPC_STATE_RUNNING:
1657                 spin_lock(&srpc_data.rpc_glock);
1658
1659                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1660                         struct srpc_service *sv = srpc_data.rpc_services[i];
1661
1662                         LASSERTF(sv == NULL,
1663                                  "service not empty: id %d, name %s\n",
1664                                  i, sv->sv_name);
1665                 }
1666
1667                 spin_unlock(&srpc_data.rpc_glock);
1668
1669                 stt_shutdown();
1670                 /* fallthrough */
1671
1672         case SRPC_STATE_EQ_INIT:
1673                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1674                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1675                 LASSERT(rc == 0);
1676                 lnet_assert_handler_unused(srpc_data.rpc_lnet_handler);
1677                 /* fallthrough */
1678
1679         case SRPC_STATE_NI_INIT:
1680                 LNetNIFini();
1681         }
1682 }