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  *
  31  * lnet/selftest/rpc.c
  32  *
  33  * Author: Isaac Huang <isaac@clusterfs.com>
  34  *
  35  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  36  * - percpt data for service to improve smp performance
  37  * - code cleanup
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LNET
  41
  42 #include "selftest.h"
  43
  44 enum srpc_state {
  45         SRPC_STATE_NONE,
  46         SRPC_STATE_NI_INIT,
  47         SRPC_STATE_EQ_INIT,
  48         SRPC_STATE_RUNNING,
  49         SRPC_STATE_STOPPING,
  50 };
  51
  52 static struct smoketest_rpc {
  53         spinlock_t       rpc_glock;     /* global lock */
  54         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  55         lnet_handler_t           rpc_lnet_handler;/* _the_ LNet event handler */
  56         enum srpc_state          rpc_state;
  57         struct srpc_counters     rpc_counters;
  58         __u64                    rpc_matchbits; /* matchbits counter */
  59 } srpc_data;
  60
  61 static inline int
  62 srpc_serv_portal(int svc_id)
  63 {
  64         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  65                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  66 }
  67
  68 /* forward ref's */
  69 static int srpc_handle_rpc(struct swi_workitem *wi);
  70
  71 void srpc_get_counters(struct srpc_counters *cnt)
  72 {
  73         spin_lock(&srpc_data.rpc_glock);
  74         *cnt = srpc_data.rpc_counters;
  75         spin_unlock(&srpc_data.rpc_glock);
  76 }
  77
  78 void srpc_set_counters(const struct srpc_counters *cnt)
  79 {
  80         spin_lock(&srpc_data.rpc_glock);
  81         srpc_data.rpc_counters = *cnt;
  82         spin_unlock(&srpc_data.rpc_glock);
  83 }
  84
  85 static int
  86 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
  87                    int nob)
  88 {
  89         LASSERT(off < PAGE_SIZE);
  90         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  91
  92         bk->bk_iovs[i].bv_offset = off;
  93         bk->bk_iovs[i].bv_page   = pg;
  94         bk->bk_iovs[i].bv_len    = nob;
  95         return nob;
  96 }
  97
  98 void
  99 srpc_free_bulk(struct srpc_bulk *bk)
 100 {
 101         int i;
 102         struct page *pg;
 103
 104         LASSERT(bk != NULL);
 105
 106         for (i = 0; i < bk->bk_niov; i++) {
 107                 pg = bk->bk_iovs[i].bv_page;
 108                 if (pg == NULL)
 109                         break;
 110
 111                 __free_page(pg);
 112         }
 113
 114         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
 115 }
 116
 117 struct srpc_bulk *
 118 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 119                 unsigned bulk_len, int sink)
 120 {
 121         struct srpc_bulk *bk;
 122         int i;
 123
 124         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 125
 126         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 127                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 128         if (bk == NULL) {
 129                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 130                 return NULL;
 131         }
 132
 133         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 134         bk->bk_sink   = sink;
 135         bk->bk_len    = bulk_len;
 136         bk->bk_niov   = bulk_npg;
 137
 138         for (i = 0; i < bulk_npg; i++) {
 139                 struct page *pg;
 140                 int nob;
 141
 142                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 143                 if (pg == NULL) {
 144                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 145                         srpc_free_bulk(bk);
 146                         return NULL;
 147                 }
 148
 149                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 150                       bulk_off;
 151
 152                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 153                 bulk_len -= nob;
 154                 bulk_off = 0;
 155         }
 156
 157         return bk;
 158 }
 159
 160 static inline __u64
 161 srpc_next_id (void)
 162 {
 163         __u64 id;
 164
 165         spin_lock(&srpc_data.rpc_glock);
 166         id = srpc_data.rpc_matchbits++;
 167         spin_unlock(&srpc_data.rpc_glock);
 168         return id;
 169 }
 170
 171 static void
 172 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 173                      struct srpc_service_cd *scd,
 174                      struct srpc_buffer *buffer)
 175 {
 176         memset(rpc, 0, sizeof(*rpc));
 177         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 178                           srpc_serv_is_framework(scd->scd_svc) ?
 179                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 180
 181         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 182
 183         rpc->srpc_scd      = scd;
 184         rpc->srpc_reqstbuf = buffer;
 185         rpc->srpc_peer     = buffer->buf_peer;
 186         rpc->srpc_self     = buffer->buf_self;
 187         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 188 }
 189
 190 static void
 191 srpc_service_fini(struct srpc_service *svc)
 192 {
 193         struct srpc_service_cd *scd;
 194         struct srpc_server_rpc *rpc;
 195         struct srpc_buffer *buf;
 196         struct list_head *q;
 197         int i;
 198
 199         if (svc->sv_cpt_data == NULL)
 200                 return;
 201
 202         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 203                 while (1) {
 204                         if (!list_empty(&scd->scd_buf_posted))
 205                                 q = &scd->scd_buf_posted;
 206                         else if (!list_empty(&scd->scd_buf_blocked))
 207                                 q = &scd->scd_buf_blocked;
 208                         else
 209                                 break;
 210
 211                         while (!list_empty(q)) {
 212                                 buf = list_entry(q->next,
 213                                                  struct srpc_buffer,
 214                                                  buf_list);
 215                                 list_del(&buf->buf_list);
 216                                 LIBCFS_FREE(buf, sizeof(*buf));
 217                         }
 218                 }
 219
 220                 LASSERT(list_empty(&scd->scd_rpc_active));
 221
 222                 while (!list_empty(&scd->scd_rpc_free)) {
 223                         rpc = list_entry(scd->scd_rpc_free.next,
 224                                          struct srpc_server_rpc,
 225                                          srpc_list);
 226                         list_del(&rpc->srpc_list);
 227                         LIBCFS_FREE(rpc, sizeof(*rpc));
 228                 }
 229         }
 230
 231         cfs_percpt_free(svc->sv_cpt_data);
 232         svc->sv_cpt_data = NULL;
 233 }
 234
 235 static int
 236 srpc_service_nrpcs(struct srpc_service *svc)
 237 {
 238         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 239
 240         return srpc_serv_is_framework(svc) ?
 241                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 242 }
 243
 244 int srpc_add_buffer(struct swi_workitem *wi);
 245
 246 static int
 247 srpc_service_init(struct srpc_service *svc)
 248 {
 249         struct srpc_service_cd *scd;
 250         struct srpc_server_rpc *rpc;
 251         int nrpcs;
 252         int i;
 253         int j;
 254
 255         svc->sv_shuttingdown = 0;
 256
 257         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 258                                             sizeof(struct srpc_service_cd));
 259         if (svc->sv_cpt_data == NULL)
 260                 return -ENOMEM;
 261
 262         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 263                         1 : cfs_cpt_number(lnet_cpt_table());
 264         nrpcs = srpc_service_nrpcs(svc);
 265
 266         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 267                 scd->scd_cpt = i;
 268                 scd->scd_svc = svc;
 269                 spin_lock_init(&scd->scd_lock);
 270                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 271                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 272                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 273                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 274
 275                 scd->scd_ev.ev_data = scd;
 276                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 277
 278                 /* NB: don't use lst_sched_serial for adding buffer,
 279                  * see details in srpc_service_add_buffers() */
 280                 swi_init_workitem(&scd->scd_buf_wi,
 281                                   srpc_add_buffer, lst_sched_test[i]);
 282
 283                 if (i != 0 && srpc_serv_is_framework(svc)) {
 284                         /* NB: framework service only needs srpc_service_cd for
 285                          * one partition, but we allocate for all to make
 286                          * it easier to implement, it will waste a little
 287                          * memory but nobody should care about this */
 288                         continue;
 289                 }
 290
 291                 for (j = 0; j < nrpcs; j++) {
 292                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 293                                          i, sizeof(*rpc));
 294                         if (rpc == NULL) {
 295                                 srpc_service_fini(svc);
 296                                 return -ENOMEM;
 297                         }
 298                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 299                 }
 300         }
 301
 302         return 0;
 303 }
 304
 305 int
 306 srpc_add_service(struct srpc_service *sv)
 307 {
 308         int id = sv->sv_id;
 309
 310         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 311
 312         if (srpc_service_init(sv) != 0)
 313                 return -ENOMEM;
 314
 315         spin_lock(&srpc_data.rpc_glock);
 316
 317         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 318
 319         if (srpc_data.rpc_services[id] != NULL) {
 320                 spin_unlock(&srpc_data.rpc_glock);
 321                 goto failed;
 322         }
 323
 324         srpc_data.rpc_services[id] = sv;
 325         spin_unlock(&srpc_data.rpc_glock);
 326
 327         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 328         return 0;
 329
 330 failed:
 331         srpc_service_fini(sv);
 332         return -EBUSY;
 333 }
 334
 335 int
 336 srpc_remove_service(struct srpc_service *sv)
 337 {
 338         int id = sv->sv_id;
 339
 340         spin_lock(&srpc_data.rpc_glock);
 341
 342         if (srpc_data.rpc_services[id] != sv) {
 343                 spin_unlock(&srpc_data.rpc_glock);
 344                 return -ENOENT;
 345         }
 346
 347         srpc_data.rpc_services[id] = NULL;
 348         spin_unlock(&srpc_data.rpc_glock);
 349         return 0;
 350 }
 351
 352 static int
 353 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 354                        int len, int options, struct lnet_process_id peer4,
 355                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 356 {
 357         int rc;
 358         struct lnet_md md;
 359         struct lnet_me *me;
 360         struct lnet_processid peer;
 361
 362         peer.pid = peer4.pid;
 363         lnet_nid4_to_nid(peer4.nid, &peer.nid);
 364
 365         me = LNetMEAttach(portal, &peer, matchbits, 0, LNET_UNLINK,
 366                           local ? LNET_INS_LOCAL : LNET_INS_AFTER);
 367         if (IS_ERR(me)) {
 368                 rc = PTR_ERR(me);
 369                 CERROR("LNetMEAttach failed: %d\n", rc);
 370                 LASSERT(rc == -ENOMEM);
 371                 return -ENOMEM;
 372         }
 373
 374         md.threshold = 1;
 375         md.user_ptr  = ev;
 376         md.start     = buf;
 377         md.length    = len;
 378         md.options   = options;
 379         md.handler   = srpc_data.rpc_lnet_handler;
 380
 381         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdh);
 382         if (rc != 0) {
 383                 CERROR("LNetMDAttach failed: %d\n", rc);
 384                 LASSERT(rc == -ENOMEM);
 385
 386                 return -ENOMEM;
 387         }
 388
 389         CDEBUG(D_NET,
 390                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 391                libcfs_id2str(peer4), portal, matchbits);
 392         return 0;
 393 }
 394
 395 static int
 396 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 397                       int options, struct lnet_process_id peer4,
 398                       lnet_nid_t self4, struct lnet_handle_md *mdh,
 399                       struct srpc_event *ev)
 400 {
 401         int rc;
 402         struct lnet_md md;
 403         struct lnet_nid self;
 404         struct lnet_processid peer;
 405
 406         lnet_nid4_to_nid(self4, &self);
 407         lnet_pid4_to_pid(peer4, &peer);
 408
 409         md.user_ptr  = ev;
 410         md.start     = buf;
 411         md.length    = len;
 412         md.handler   = srpc_data.rpc_lnet_handler;
 413         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 414         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 415
 416         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
 417         if (rc != 0) {
 418                 CERROR("LNetMDBind failed: %d\n", rc);
 419                 LASSERT(rc == -ENOMEM);
 420                 return -ENOMEM;
 421         }
 422
 423         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 424          * they're only meaningful for MDs attached to an ME (i.e. passive
 425          * buffers...
 426          */
 427         if ((options & LNET_MD_OP_PUT) != 0) {
 428                 rc = LNetPut(&self, *mdh, LNET_NOACK_REQ, &peer,
 429                              portal, matchbits, 0, 0);
 430         } else {
 431                 LASSERT((options & LNET_MD_OP_GET) != 0);
 432
 433                 rc = LNetGet(&self, *mdh, &peer, portal, matchbits, 0, false);
 434         }
 435
 436         if (rc != 0) {
 437                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 438                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 439                        libcfs_id2str(peer4), portal, matchbits, rc);
 440
 441                 /* The forthcoming unlink event will complete this operation
 442                  * with failure, so fall through and return success here.
 443                  */
 444                 rc = LNetMDUnlink(*mdh);
 445                 LASSERT(rc == 0);
 446         } else {
 447                 CDEBUG(D_NET,
 448                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 449                        libcfs_id2str(peer4), portal, matchbits);
 450         }
 451         return 0;
 452 }
 453
 454 static int
 455 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 456                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 457 {
 458         struct lnet_process_id any = {0};
 459
 460         any.nid = LNET_NID_ANY;
 461         any.pid = LNET_PID_ANY;
 462
 463         return srpc_post_passive_rdma(srpc_serv_portal(service),
 464                                       local, service, buf, len,
 465                                       LNET_MD_OP_PUT, any, mdh, ev);
 466 }
 467
 468 static int
 469 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 470 __must_hold(&scd->scd_lock)
 471 {
 472         struct srpc_service *sv = scd->scd_svc;
 473         struct srpc_msg *msg = &buf->buf_msg;
 474         int rc;
 475
 476         LNetInvalidateMDHandle(&buf->buf_mdh);
 477         list_add(&buf->buf_list, &scd->scd_buf_posted);
 478         scd->scd_buf_nposted++;
 479         spin_unlock(&scd->scd_lock);
 480
 481         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 482                                       !srpc_serv_is_framework(sv),
 483                                       msg, sizeof(*msg), &buf->buf_mdh,
 484                                       &scd->scd_ev);
 485
 486         /* At this point, a RPC (new or delayed) may have arrived in
 487          * msg and its event handler has been called. So we must add
 488          * buf to scd_buf_posted _before_ dropping scd_lock */
 489
 490         spin_lock(&scd->scd_lock);
 491
 492         if (rc == 0) {
 493                 if (!sv->sv_shuttingdown)
 494                         return 0;
 495
 496                 spin_unlock(&scd->scd_lock);
 497                 /* srpc_shutdown_service might have tried to unlink me
 498                  * when my buf_mdh was still invalid */
 499                 LNetMDUnlink(buf->buf_mdh);
 500                 spin_lock(&scd->scd_lock);
 501                 return 0;
 502         }
 503
 504         scd->scd_buf_nposted--;
 505         if (sv->sv_shuttingdown)
 506                 return rc; /* don't allow to change scd_buf_posted */
 507
 508         list_del(&buf->buf_list);
 509         spin_unlock(&scd->scd_lock);
 510
 511         LIBCFS_FREE(buf, sizeof(*buf));
 512
 513         spin_lock(&scd->scd_lock);
 514         return rc;
 515 }
 516
 517 int
 518 srpc_add_buffer(struct swi_workitem *wi)
 519 {
 520         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 521                                                    scd_buf_wi);
 522         struct srpc_buffer *buf;
 523         int rc = 0;
 524
 525         /* it's called by workitem scheduler threads, these threads
 526          * should have been set CPT affinity, so buffers will be posted
 527          * on CPT local list of Portal */
 528         spin_lock(&scd->scd_lock);
 529
 530         while (scd->scd_buf_adjust > 0 &&
 531                !scd->scd_svc->sv_shuttingdown) {
 532                 scd->scd_buf_adjust--; /* consume it */
 533                 scd->scd_buf_posting++;
 534
 535                 spin_unlock(&scd->scd_lock);
 536
 537                 LIBCFS_ALLOC(buf, sizeof(*buf));
 538                 if (buf == NULL) {
 539                         CERROR("Failed to add new buf to service: %s\n",
 540                                scd->scd_svc->sv_name);
 541                         spin_lock(&scd->scd_lock);
 542                         rc = -ENOMEM;
 543                         break;
 544                 }
 545
 546                 spin_lock(&scd->scd_lock);
 547                 if (scd->scd_svc->sv_shuttingdown) {
 548                         spin_unlock(&scd->scd_lock);
 549                         LIBCFS_FREE(buf, sizeof(*buf));
 550
 551                         spin_lock(&scd->scd_lock);
 552                         rc = -ESHUTDOWN;
 553                         break;
 554                 }
 555
 556                 rc = srpc_service_post_buffer(scd, buf);
 557                 if (rc != 0)
 558                         break; /* buf has been freed inside */
 559
 560                 LASSERT(scd->scd_buf_posting > 0);
 561                 scd->scd_buf_posting--;
 562                 scd->scd_buf_total++;
 563                 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
 564         }
 565
 566         if (rc != 0) {
 567                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 568                 scd->scd_buf_err = rc;
 569
 570                 LASSERT(scd->scd_buf_posting > 0);
 571                 scd->scd_buf_posting--;
 572         }
 573
 574         spin_unlock(&scd->scd_lock);
 575         return 0;
 576 }
 577
 578 int
 579 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 580 {
 581         struct srpc_service_cd *scd;
 582         int rc = 0;
 583         int i;
 584
 585         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 586
 587         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 588                 spin_lock(&scd->scd_lock);
 589
 590                 scd->scd_buf_err = 0;
 591                 scd->scd_buf_err_stamp = 0;
 592                 scd->scd_buf_posting = 0;
 593                 scd->scd_buf_adjust = nbuffer;
 594                 /* start to post buffers */
 595                 swi_schedule_workitem(&scd->scd_buf_wi);
 596                 spin_unlock(&scd->scd_lock);
 597
 598                 /* framework service only post buffer for one partition  */
 599                 if (srpc_serv_is_framework(sv))
 600                         break;
 601         }
 602
 603         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 604                 spin_lock(&scd->scd_lock);
 605                 /*
 606                  * NB: srpc_service_add_buffers() can be called inside
 607                  * thread context of lst_sched_serial, and we don't normally
 608                  * allow to sleep inside thread context of WI scheduler
 609                  * because it will block current scheduler thread from doing
 610                  * anything else, even worse, it could deadlock if it's
 611                  * waiting on result from another WI of the same scheduler.
 612                  * However, it's safe at here because scd_buf_wi is scheduled
 613                  * by thread in a different WI scheduler (lst_sched_test),
 614                  * so we don't have any risk of deadlock, though this could
 615                  * block all WIs pending on lst_sched_serial for a moment
 616                  * which is not good but not fatal.
 617                  */
 618                 lst_wait_until(scd->scd_buf_err != 0 ||
 619                                (scd->scd_buf_adjust == 0 &&
 620                                 scd->scd_buf_posting == 0),
 621                                scd->scd_lock, "waiting for adding buffer\n");
 622
 623                 if (scd->scd_buf_err != 0 && rc == 0)
 624                         rc = scd->scd_buf_err;
 625
 626                 spin_unlock(&scd->scd_lock);
 627         }
 628
 629         return rc;
 630 }
 631
 632 void
 633 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 634 {
 635         struct srpc_service_cd *scd;
 636         int num;
 637         int i;
 638
 639         LASSERT(!sv->sv_shuttingdown);
 640
 641         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 642                 spin_lock(&scd->scd_lock);
 643
 644                 num = scd->scd_buf_total + scd->scd_buf_posting;
 645                 scd->scd_buf_adjust -= min(nbuffer, num);
 646
 647                 spin_unlock(&scd->scd_lock);
 648         }
 649 }
 650
 651 /* returns 1 if sv has finished, otherwise 0 */
 652 int
 653 srpc_finish_service(struct srpc_service *sv)
 654 {
 655         struct srpc_service_cd *scd;
 656         struct srpc_server_rpc *rpc;
 657         int i;
 658
 659         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 660
 661         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 662                 spin_lock(&scd->scd_lock);
 663                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 664                         spin_unlock(&scd->scd_lock);
 665                         return 0;
 666                 }
 667
 668                 if (scd->scd_buf_nposted > 0) {
 669                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 670                                scd->scd_buf_nposted);
 671                         spin_unlock(&scd->scd_lock);
 672                         return 0;
 673                 }
 674
 675                 if (list_empty(&scd->scd_rpc_active)) {
 676                         spin_unlock(&scd->scd_lock);
 677                         continue;
 678                 }
 679
 680                 rpc = list_entry(scd->scd_rpc_active.next,
 681                                  struct srpc_server_rpc, srpc_list);
 682                 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",
 683                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 684                         swi_state2str(rpc->srpc_wi.swi_state),
 685                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 686                         rpc->srpc_wi.swi_workitem.wi_running,
 687                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 688                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 689                 spin_unlock(&scd->scd_lock);
 690                 return 0;
 691         }
 692
 693         /* no lock needed from now on */
 694         srpc_service_fini(sv);
 695         return 1;
 696 }
 697
 698 /* called with sv->sv_lock held */
 699 static void
 700 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 701                             struct srpc_buffer *buf)
 702 __must_hold(&scd->scd_lock)
 703 {
 704         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 705                 if (srpc_service_post_buffer(scd, buf) != 0) {
 706                         CWARN("Failed to post %s buffer\n",
 707                               scd->scd_svc->sv_name);
 708                 }
 709                 return;
 710         }
 711
 712         /* service is shutting down, or we want to recycle some buffers */
 713         scd->scd_buf_total--;
 714
 715         if (scd->scd_buf_adjust < 0) {
 716                 scd->scd_buf_adjust++;
 717                 if (scd->scd_buf_adjust < 0 &&
 718                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 719                         CDEBUG(D_INFO,
 720                                "Try to recyle %d buffers but nothing left\n",
 721                                scd->scd_buf_adjust);
 722                         scd->scd_buf_adjust = 0;
 723                 }
 724         }
 725
 726         spin_unlock(&scd->scd_lock);
 727         LIBCFS_FREE(buf, sizeof(*buf));
 728         spin_lock(&scd->scd_lock);
 729 }
 730
 731 void
 732 srpc_abort_service(struct srpc_service *sv)
 733 {
 734         struct srpc_service_cd *scd;
 735         struct srpc_server_rpc *rpc;
 736         int i;
 737
 738         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 739                sv->sv_id, sv->sv_name);
 740
 741         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 742                 spin_lock(&scd->scd_lock);
 743
 744                 /* schedule in-flight RPCs to notice the abort, NB:
 745                  * racing with incoming RPCs; complete fix should make test
 746                  * RPCs carry session ID in its headers
 747                  */
 748                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 749                         rpc->srpc_aborted = 1;
 750                         swi_schedule_workitem(&rpc->srpc_wi);
 751                 }
 752
 753                 spin_unlock(&scd->scd_lock);
 754         }
 755 }
 756
 757 void
 758 srpc_shutdown_service(struct srpc_service *sv)
 759 {
 760         struct srpc_service_cd *scd;
 761         struct srpc_server_rpc *rpc;
 762         struct srpc_buffer *buf;
 763         int i;
 764
 765         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 766                sv->sv_id, sv->sv_name);
 767
 768         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 769                 spin_lock(&scd->scd_lock);
 770
 771         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 772
 773         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 774                 spin_unlock(&scd->scd_lock);
 775
 776         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 777                 spin_lock(&scd->scd_lock);
 778
 779                 /* schedule in-flight RPCs to notice the shutdown */
 780                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 781                         swi_schedule_workitem(&rpc->srpc_wi);
 782
 783                 spin_unlock(&scd->scd_lock);
 784
 785                 /* OK to traverse scd_buf_posted without lock, since no one
 786                  * touches scd_buf_posted now
 787                  */
 788                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 789                         LNetMDUnlink(buf->buf_mdh);
 790         }
 791 }
 792
 793 static int
 794 srpc_send_request(struct srpc_client_rpc *rpc)
 795 {
 796         struct srpc_event *ev = &rpc->crpc_reqstev;
 797         int rc;
 798
 799         ev->ev_fired = 0;
 800         ev->ev_data  = rpc;
 801         ev->ev_type  = SRPC_REQUEST_SENT;
 802
 803         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 804                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 805                                    sizeof(struct srpc_msg), LNET_MD_OP_PUT,
 806                                    rpc->crpc_dest, LNET_NID_ANY,
 807                                    &rpc->crpc_reqstmdh, ev);
 808         if (rc != 0) {
 809                 LASSERT(rc == -ENOMEM);
 810                 ev->ev_fired = 1;  /* no more event expected */
 811         }
 812         return rc;
 813 }
 814
 815 static int
 816 srpc_prepare_reply(struct srpc_client_rpc *rpc)
 817 {
 818         struct srpc_event *ev = &rpc->crpc_replyev;
 819         u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 820         int rc;
 821
 822         ev->ev_fired = 0;
 823         ev->ev_data  = rpc;
 824         ev->ev_type  = SRPC_REPLY_RCVD;
 825
 826         *id = srpc_next_id();
 827
 828         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 829                                     &rpc->crpc_replymsg,
 830                                     sizeof(struct srpc_msg),
 831                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 832                                     &rpc->crpc_replymdh, ev);
 833         if (rc != 0) {
 834                 LASSERT(rc == -ENOMEM);
 835                 ev->ev_fired = 1;  /* no more event expected */
 836         }
 837         return rc;
 838 }
 839
 840 static int
 841 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 842 {
 843         struct srpc_bulk *bk = &rpc->crpc_bulk;
 844         struct srpc_event *ev = &rpc->crpc_bulkev;
 845         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 846         int rc;
 847         int opt;
 848
 849         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 850
 851         /* nothing to do */
 852         if (bk->bk_niov == 0)
 853                 return 0;
 854
 855         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 856         opt |= LNET_MD_KIOV;
 857
 858         ev->ev_fired = 0;
 859         ev->ev_data  = rpc;
 860         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 861
 862         *id = srpc_next_id();
 863
 864         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 865                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 866                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 867         if (rc != 0) {
 868                 LASSERT(rc == -ENOMEM);
 869                 ev->ev_fired = 1;  /* no more event expected */
 870         }
 871         return rc;
 872 }
 873
 874 static int
 875 srpc_do_bulk(struct srpc_server_rpc *rpc)
 876 {
 877         struct srpc_event *ev = &rpc->srpc_ev;
 878         struct srpc_bulk *bk = rpc->srpc_bulk;
 879         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 880         int rc;
 881         int opt;
 882
 883         LASSERT(bk != NULL);
 884
 885         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 886         opt |= LNET_MD_KIOV;
 887
 888         ev->ev_fired = 0;
 889         ev->ev_data  = rpc;
 890         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 891
 892         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 893                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 894                                    rpc->srpc_peer, rpc->srpc_self,
 895                                    &bk->bk_mdh, ev);
 896         if (rc != 0)
 897                 ev->ev_fired = 1;  /* no more event expected */
 898         return rc;
 899 }
 900
 901 /* only called from srpc_handle_rpc */
 902 static void
 903 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 904 {
 905         struct srpc_service_cd  *scd = rpc->srpc_scd;
 906         struct srpc_service     *sv  = scd->scd_svc;
 907         struct srpc_buffer *buffer;
 908
 909         LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 910
 911         rpc->srpc_status = status;
 912
 913         CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
 914                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 915                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 916                      swi_state2str(rpc->srpc_wi.swi_state), status);
 917
 918         if (status != 0) {
 919                 spin_lock(&srpc_data.rpc_glock);
 920                 srpc_data.rpc_counters.rpcs_dropped++;
 921                 spin_unlock(&srpc_data.rpc_glock);
 922         }
 923
 924         if (rpc->srpc_done != NULL)
 925                 (*rpc->srpc_done) (rpc);
 926         LASSERT(rpc->srpc_bulk == NULL);
 927
 928         spin_lock(&scd->scd_lock);
 929
 930         if (rpc->srpc_reqstbuf != NULL) {
 931                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 932                  * sv won't go away for scd_rpc_active must not be empty
 933                  */
 934                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 935                 rpc->srpc_reqstbuf = NULL;
 936         }
 937
 938         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 939
 940         /*
 941          * No one can schedule me now since:
 942          * - I'm not on scd_rpc_active.
 943          * - all LNet events have been fired.
 944          * Cancel pending schedules and prevent future schedule attempts:
 945          */
 946         LASSERT(rpc->srpc_ev.ev_fired);
 947         swi_exit_workitem(&rpc->srpc_wi);
 948
 949         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 950                 buffer = list_entry(scd->scd_buf_blocked.next,
 951                                     struct srpc_buffer, buf_list);
 952                 list_del(&buffer->buf_list);
 953
 954                 srpc_init_server_rpc(rpc, scd, buffer);
 955                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 956                 swi_schedule_workitem(&rpc->srpc_wi);
 957         } else {
 958                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 959         }
 960
 961         spin_unlock(&scd->scd_lock);
 962 }
 963
 964 /* handles an incoming RPC */
 965 static int srpc_handle_rpc(struct swi_workitem *wi)
 966 {
 967         struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc,
 968                                                    srpc_wi);
 969         struct srpc_service_cd *scd = rpc->srpc_scd;
 970         struct srpc_service *sv = scd->scd_svc;
 971         struct srpc_event *ev = &rpc->srpc_ev;
 972         int rc = 0;
 973
 974         LASSERT(wi == &rpc->srpc_wi);
 975
 976         spin_lock(&scd->scd_lock);
 977
 978         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 979                 spin_unlock(&scd->scd_lock);
 980
 981                 if (rpc->srpc_bulk != NULL)
 982                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 983                 LNetMDUnlink(rpc->srpc_replymdh);
 984
 985                 if (ev->ev_fired) { /* no more event, OK to finish */
 986                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 987                         return 1;
 988                 }
 989                 return 0;
 990         }
 991
 992         spin_unlock(&scd->scd_lock);
 993
 994         switch (wi->swi_state) {
 995         default:
 996                 LBUG();
 997                 fallthrough;
 998         case SWI_STATE_NEWBORN: {
 999                 struct srpc_msg *msg;
1000                 struct srpc_generic_reply *reply;
1001
1002                 msg = &rpc->srpc_reqstbuf->buf_msg;
1003                 reply = &rpc->srpc_replymsg.msg_body.reply;
1004
1005                 if (msg->msg_magic == 0) {
1006                         /* moaned already in srpc_lnet_ev_handler */
1007                         srpc_server_rpc_done(rpc, EBADMSG);
1008                         return 1;
1009                 }
1010
1011                 srpc_unpack_msg_hdr(msg);
1012                 if (msg->msg_version != SRPC_MSG_VERSION) {
1013                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1014                               msg->msg_version, SRPC_MSG_VERSION,
1015                               libcfs_id2str(rpc->srpc_peer));
1016                         reply->status = EPROTO;
1017                         /* drop through and send reply */
1018                 } else {
1019                         reply->status = 0;
1020                         rc = (*sv->sv_handler)(rpc);
1021                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1022                         if (rc != 0) {
1023                                 srpc_server_rpc_done(rpc, rc);
1024                                 return 1;
1025                         }
1026                 }
1027
1028                 wi->swi_state = SWI_STATE_BULK_STARTED;
1029
1030                 if (rpc->srpc_bulk != NULL) {
1031                         rc = srpc_do_bulk(rpc);
1032                         if (rc == 0)
1033                                 return 0; /* wait for bulk */
1034
1035                         LASSERT(ev->ev_fired);
1036                         ev->ev_status = rc;
1037                 }
1038         }
1039         fallthrough;
1040         case SWI_STATE_BULK_STARTED:
1041                 LASSERT(rpc->srpc_bulk == NULL || ev->ev_fired);
1042
1043                 if (rpc->srpc_bulk != NULL) {
1044                         rc = ev->ev_status;
1045
1046                         if (sv->sv_bulk_ready != NULL)
1047                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1048
1049                         if (rc != 0) {
1050                                 srpc_server_rpc_done(rpc, rc);
1051                                 return 1;
1052                         }
1053                 }
1054
1055                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1056                 rc = srpc_send_reply(rpc);
1057                 if (rc == 0)
1058                         return 0; /* wait for reply */
1059                 srpc_server_rpc_done(rpc, rc);
1060                 return 1;
1061
1062         case SWI_STATE_REPLY_SUBMITTED:
1063                 if (!ev->ev_fired) {
1064                         CERROR("RPC %p: bulk %p, service %d\n",
1065                                rpc, rpc->srpc_bulk, sv->sv_id);
1066                         CERROR("Event: status %d, type %d, lnet %d\n",
1067                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1068                         LASSERT(ev->ev_fired);
1069                 }
1070
1071                 wi->swi_state = SWI_STATE_DONE;
1072                 srpc_server_rpc_done(rpc, ev->ev_status);
1073                 return 1;
1074         }
1075
1076         return 0;
1077 }
1078
1079 static void
1080 srpc_client_rpc_expired (void *data)
1081 {
1082         struct srpc_client_rpc *rpc = data;
1083
1084         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1085               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1086               rpc->crpc_timeout);
1087
1088         spin_lock(&rpc->crpc_lock);
1089
1090         rpc->crpc_timeout = 0;
1091         srpc_abort_rpc(rpc, -ETIMEDOUT);
1092
1093         spin_unlock(&rpc->crpc_lock);
1094
1095         spin_lock(&srpc_data.rpc_glock);
1096         srpc_data.rpc_counters.rpcs_expired++;
1097         spin_unlock(&srpc_data.rpc_glock);
1098 }
1099
1100 static void
1101 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1102 {
1103         struct stt_timer *timer = &rpc->crpc_timer;
1104
1105         if (rpc->crpc_timeout == 0)
1106                 return;
1107
1108         INIT_LIST_HEAD(&timer->stt_list);
1109         timer->stt_data    = rpc;
1110         timer->stt_func    = srpc_client_rpc_expired;
1111         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1112         stt_add_timer(timer);
1113 }
1114
1115 /*
1116  * Called with rpc->crpc_lock held.
1117  *
1118  * Upon exit the RPC expiry timer is not queued and the handler is not
1119  * running on any CPU.
1120  */
1121 static void
1122 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1123 {
1124         /* timer not planted or already exploded */
1125         if (rpc->crpc_timeout == 0)
1126                 return;
1127
1128         /* timer successfully defused */
1129         if (stt_del_timer(&rpc->crpc_timer))
1130                 return;
1131
1132         /* timer detonated, wait for it to explode */
1133         while (rpc->crpc_timeout != 0) {
1134                 spin_unlock(&rpc->crpc_lock);
1135
1136                 schedule();
1137
1138                 spin_lock(&rpc->crpc_lock);
1139         }
1140 }
1141
1142 static void
1143 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1144 {
1145         struct swi_workitem *wi = &rpc->crpc_wi;
1146
1147         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1148
1149         spin_lock(&rpc->crpc_lock);
1150
1151         rpc->crpc_closed = 1;
1152         if (rpc->crpc_status == 0)
1153                 rpc->crpc_status = status;
1154
1155         srpc_del_client_rpc_timer(rpc);
1156
1157         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1158                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1159                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1160                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1161
1162         /*
1163          * No one can schedule me now since:
1164          * - RPC timer has been defused.
1165          * - all LNet events have been fired.
1166          * - crpc_closed has been set, preventing srpc_abort_rpc from
1167          *   scheduling me.
1168          * Cancel pending schedules and prevent future schedule attempts:
1169          */
1170         LASSERT(!srpc_event_pending(rpc));
1171         swi_exit_workitem(wi);
1172
1173         spin_unlock(&rpc->crpc_lock);
1174
1175         (*rpc->crpc_done)(rpc);
1176 }
1177
1178 /* sends an outgoing RPC */
1179 int
1180 srpc_send_rpc(struct swi_workitem *wi)
1181 {
1182         int rc = 0;
1183         struct srpc_client_rpc *rpc;
1184         struct srpc_msg *reply;
1185         int do_bulk;
1186
1187         LASSERT(wi != NULL);
1188
1189         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1190
1191         LASSERT(rpc != NULL);
1192         LASSERT(wi == &rpc->crpc_wi);
1193
1194         reply = &rpc->crpc_replymsg;
1195         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1196
1197         spin_lock(&rpc->crpc_lock);
1198
1199         if (rpc->crpc_aborted) {
1200                 spin_unlock(&rpc->crpc_lock);
1201                 goto abort;
1202         }
1203
1204         spin_unlock(&rpc->crpc_lock);
1205
1206         switch (wi->swi_state) {
1207         default:
1208                 LBUG();
1209         case SWI_STATE_NEWBORN:
1210                 LASSERT(!srpc_event_pending(rpc));
1211
1212                 rc = srpc_prepare_reply(rpc);
1213                 if (rc != 0) {
1214                         srpc_client_rpc_done(rpc, rc);
1215                         return 1;
1216                 }
1217
1218                 rc = srpc_prepare_bulk(rpc);
1219                 if (rc != 0)
1220                         break;
1221
1222                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1223                 rc = srpc_send_request(rpc);
1224                 break;
1225
1226         case SWI_STATE_REQUEST_SUBMITTED:
1227                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1228                  * order; however, they're processed in a strict order:
1229                  * rqt, rpy, and bulk.
1230                  */
1231                 if (!rpc->crpc_reqstev.ev_fired)
1232                         break;
1233
1234                 rc = rpc->crpc_reqstev.ev_status;
1235                 if (rc != 0)
1236                         break;
1237
1238                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1239                 fallthrough;
1240         case SWI_STATE_REQUEST_SENT: {
1241                 enum srpc_msg_type type;
1242
1243                 type = srpc_service2reply(rpc->crpc_service);
1244
1245                 if (!rpc->crpc_replyev.ev_fired)
1246                         break;
1247
1248                 rc = rpc->crpc_replyev.ev_status;
1249                 if (rc != 0)
1250                         break;
1251
1252                 srpc_unpack_msg_hdr(reply);
1253                 if (reply->msg_type != type ||
1254                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1255                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1256                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1257                               libcfs_id2str(rpc->crpc_dest),
1258                               reply->msg_type, type,
1259                               reply->msg_magic, SRPC_MSG_MAGIC);
1260                         rc = -EBADMSG;
1261                         break;
1262                 }
1263
1264                 if (do_bulk && reply->msg_body.reply.status != 0) {
1265                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1266                               reply->msg_body.reply.status,
1267                               libcfs_id2str(rpc->crpc_dest));
1268                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1269                 }
1270
1271                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1272         }
1273         fallthrough;
1274         case SWI_STATE_REPLY_RECEIVED:
1275                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1276                         break;
1277
1278                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1279
1280                 /* Bulk buffer was unlinked due to remote error. Clear error
1281                  * since reply buffer still contains valid data.
1282                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1283                  * remote error.
1284                  */
1285                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1286                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1287                         rc = 0;
1288
1289                 wi->swi_state = SWI_STATE_DONE;
1290                 srpc_client_rpc_done(rpc, rc);
1291                 return 1;
1292         }
1293
1294         if (rc != 0) {
1295                 spin_lock(&rpc->crpc_lock);
1296                 srpc_abort_rpc(rpc, rc);
1297                 spin_unlock(&rpc->crpc_lock);
1298         }
1299
1300 abort:
1301         if (rpc->crpc_aborted) {
1302                 LNetMDUnlink(rpc->crpc_reqstmdh);
1303                 LNetMDUnlink(rpc->crpc_replymdh);
1304                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1305
1306                 if (!srpc_event_pending(rpc)) {
1307                         srpc_client_rpc_done(rpc, -EINTR);
1308                         return 1;
1309                 }
1310         }
1311         return 0;
1312 }
1313
1314 struct srpc_client_rpc *
1315 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1316                        int nbulkiov, int bulklen,
1317                        void (*rpc_done)(struct srpc_client_rpc *),
1318                        void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1319 {
1320         struct srpc_client_rpc *rpc;
1321
1322         LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1323                                    crpc_bulk.bk_iovs[nbulkiov]));
1324         if (rpc == NULL)
1325                 return NULL;
1326
1327         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1328                              bulklen, rpc_done, rpc_fini, priv);
1329         return rpc;
1330 }
1331
1332 /* called with rpc->crpc_lock held */
1333 void
1334 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1335 {
1336         LASSERT(why != 0);
1337
1338         if (rpc->crpc_aborted || /* already aborted */
1339             rpc->crpc_closed)    /* callback imminent */
1340                 return;
1341
1342         CDEBUG(D_NET,
1343                "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1344                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1345                swi_state2str(rpc->crpc_wi.swi_state), why);
1346
1347         rpc->crpc_aborted = 1;
1348         rpc->crpc_status  = why;
1349         swi_schedule_workitem(&rpc->crpc_wi);
1350 }
1351
1352 /* called with rpc->crpc_lock held */
1353 void
1354 srpc_post_rpc(struct srpc_client_rpc *rpc)
1355 {
1356         LASSERT(!rpc->crpc_aborted);
1357         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1358
1359         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1360                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1361                rpc->crpc_timeout);
1362
1363         srpc_add_client_rpc_timer(rpc);
1364         swi_schedule_workitem(&rpc->crpc_wi);
1365 }
1366
1367
1368 int
1369 srpc_send_reply(struct srpc_server_rpc *rpc)
1370 {
1371         struct srpc_event *ev = &rpc->srpc_ev;
1372         struct srpc_msg *msg = &rpc->srpc_replymsg;
1373         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1374         struct srpc_service_cd *scd = rpc->srpc_scd;
1375         struct srpc_service *sv = scd->scd_svc;
1376         __u64 rpyid;
1377         int rc;
1378
1379         LASSERT(buffer != NULL);
1380         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1381
1382         spin_lock(&scd->scd_lock);
1383
1384         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1385                 /* Repost buffer before replying since test client
1386                  * might send me another RPC once it gets the reply
1387                  */
1388                 if (srpc_service_post_buffer(scd, buffer) != 0)
1389                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1390                 rpc->srpc_reqstbuf = NULL;
1391         }
1392
1393         spin_unlock(&scd->scd_lock);
1394
1395         ev->ev_fired = 0;
1396         ev->ev_data  = rpc;
1397         ev->ev_type  = SRPC_REPLY_SENT;
1398
1399         msg->msg_magic   = SRPC_MSG_MAGIC;
1400         msg->msg_version = SRPC_MSG_VERSION;
1401         msg->msg_type    = srpc_service2reply(sv->sv_id);
1402
1403         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1404                                    sizeof(*msg), LNET_MD_OP_PUT,
1405                                    rpc->srpc_peer, rpc->srpc_self,
1406                                    &rpc->srpc_replymdh, ev);
1407         if (rc != 0)
1408                 ev->ev_fired = 1;  /* no more event expected */
1409         return rc;
1410 }
1411
1412 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1413 static void
1414 srpc_lnet_ev_handler(struct lnet_event *ev)
1415 {
1416         struct srpc_service_cd *scd;
1417         struct srpc_event *rpcev = ev->md_user_ptr;
1418         struct srpc_client_rpc *crpc;
1419         struct srpc_server_rpc *srpc;
1420         struct srpc_buffer *buffer;
1421         struct srpc_service *sv;
1422         struct srpc_msg *msg;
1423         enum srpc_msg_type type;
1424
1425         LASSERT(!in_interrupt());
1426
1427         if (ev->status != 0) {
1428                 __u32 errors;
1429
1430                 spin_lock(&srpc_data.rpc_glock);
1431                 if (ev->status != -ECANCELED) /* cancellation is not error */
1432                         srpc_data.rpc_counters.errors++;
1433                 errors = srpc_data.rpc_counters.errors;
1434                 spin_unlock(&srpc_data.rpc_glock);
1435
1436                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1437                         ev->status, ev->type, errors);
1438         }
1439
1440         rpcev->ev_lnet = ev->type;
1441
1442         switch (rpcev->ev_type) {
1443         default:
1444                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1445                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1446                 LBUG();
1447                 fallthrough;
1448         case SRPC_REQUEST_SENT:
1449                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1450                         spin_lock(&srpc_data.rpc_glock);
1451                         srpc_data.rpc_counters.rpcs_sent++;
1452                         spin_unlock(&srpc_data.rpc_glock);
1453                 }
1454                 fallthrough;
1455         case SRPC_REPLY_RCVD:
1456         case SRPC_BULK_REQ_RCVD:
1457                 crpc = rpcev->ev_data;
1458
1459                 if (rpcev != &crpc->crpc_reqstev &&
1460                     rpcev != &crpc->crpc_replyev &&
1461                     rpcev != &crpc->crpc_bulkev) {
1462                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1463                                rpcev, crpc, &crpc->crpc_reqstev,
1464                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1465                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1466                                rpcev->ev_status, rpcev->ev_type,
1467                                rpcev->ev_lnet);
1468                         LBUG();
1469                 }
1470
1471                 spin_lock(&crpc->crpc_lock);
1472
1473                 LASSERT(rpcev->ev_fired == 0);
1474                 rpcev->ev_fired  = 1;
1475                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1476                                    -EINTR : ev->status;
1477                 swi_schedule_workitem(&crpc->crpc_wi);
1478
1479                 spin_unlock(&crpc->crpc_lock);
1480                 break;
1481
1482         case SRPC_REQUEST_RCVD:
1483                 scd = rpcev->ev_data;
1484                 sv = scd->scd_svc;
1485
1486                 LASSERT(rpcev == &scd->scd_ev);
1487
1488                 spin_lock(&scd->scd_lock);
1489
1490                 LASSERT(ev->unlinked);
1491                 LASSERT(ev->type == LNET_EVENT_PUT ||
1492                         ev->type == LNET_EVENT_UNLINK);
1493                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1494                         sv->sv_shuttingdown);
1495
1496                 buffer = container_of(ev->md_start, struct srpc_buffer,
1497                                       buf_msg);
1498                 buffer->buf_peer = lnet_pid_to_pid4(&ev->source);
1499                 buffer->buf_self = lnet_nid_to_nid4(&ev->target.nid);
1500
1501                 LASSERT(scd->scd_buf_nposted > 0);
1502                 scd->scd_buf_nposted--;
1503
1504                 if (sv->sv_shuttingdown) {
1505                         /* Leave buffer on scd->scd_buf_nposted since
1506                          * srpc_finish_service needs to traverse it.
1507                          */
1508                         spin_unlock(&scd->scd_lock);
1509                         break;
1510                 }
1511
1512                 if (scd->scd_buf_err_stamp != 0 &&
1513                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1514                         /* re-enable adding buffer */
1515                         scd->scd_buf_err_stamp = 0;
1516                         scd->scd_buf_err = 0;
1517                 }
1518
1519                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1520                     scd->scd_buf_adjust == 0 &&
1521                     scd->scd_buf_nposted < scd->scd_buf_low) {
1522                         scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1523                                                   SFW_TEST_WI_MIN);
1524                         swi_schedule_workitem(&scd->scd_buf_wi);
1525                 }
1526
1527                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1528                 msg = &buffer->buf_msg;
1529                 type = srpc_service2request(sv->sv_id);
1530
1531                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1532                     (msg->msg_type != type &&
1533                      msg->msg_type != __swab32(type)) ||
1534                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1535                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1536                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1537                                sv->sv_name, libcfs_idstr(&ev->initiator),
1538                                ev->status, ev->mlength,
1539                                msg->msg_type, msg->msg_magic);
1540
1541                         /* NB can't call srpc_service_recycle_buffer here since
1542                          * it may call LNetM[DE]Attach. The invalid magic tells
1543                          * srpc_handle_rpc to drop this RPC
1544                          */
1545                         msg->msg_magic = 0;
1546                 }
1547
1548                 if (!list_empty(&scd->scd_rpc_free)) {
1549                         srpc = list_entry(scd->scd_rpc_free.next,
1550                                           struct srpc_server_rpc,
1551                                           srpc_list);
1552                         list_del(&srpc->srpc_list);
1553
1554                         srpc_init_server_rpc(srpc, scd, buffer);
1555                         list_add_tail(&srpc->srpc_list,
1556                                       &scd->scd_rpc_active);
1557                         swi_schedule_workitem(&srpc->srpc_wi);
1558                 } else {
1559                         list_add_tail(&buffer->buf_list,
1560                                       &scd->scd_buf_blocked);
1561                 }
1562
1563                 spin_unlock(&scd->scd_lock);
1564
1565                 spin_lock(&srpc_data.rpc_glock);
1566                 srpc_data.rpc_counters.rpcs_rcvd++;
1567                 spin_unlock(&srpc_data.rpc_glock);
1568                 break;
1569
1570         case SRPC_BULK_GET_RPLD:
1571                 LASSERT(ev->type == LNET_EVENT_SEND ||
1572                         ev->type == LNET_EVENT_REPLY ||
1573                         ev->type == LNET_EVENT_UNLINK);
1574
1575                 if (!ev->unlinked)
1576                         break; /* wait for final event */
1577                 fallthrough;
1578         case SRPC_BULK_PUT_SENT:
1579                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1580                         spin_lock(&srpc_data.rpc_glock);
1581
1582                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1583                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1584                         else
1585                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1586
1587                         spin_unlock(&srpc_data.rpc_glock);
1588                 }
1589                 fallthrough;
1590         case SRPC_REPLY_SENT:
1591                 srpc = rpcev->ev_data;
1592                 scd  = srpc->srpc_scd;
1593
1594                 LASSERT(rpcev == &srpc->srpc_ev);
1595
1596                 spin_lock(&scd->scd_lock);
1597
1598                 rpcev->ev_fired  = 1;
1599                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1600                                    -EINTR : ev->status;
1601                 swi_schedule_workitem(&srpc->srpc_wi);
1602
1603                 spin_unlock(&scd->scd_lock);
1604                 break;
1605         }
1606 }
1607
1608
1609 int
1610 srpc_startup (void)
1611 {
1612         int rc;
1613
1614         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1615         spin_lock_init(&srpc_data.rpc_glock);
1616
1617         /* 1 second pause to avoid timestamp reuse */
1618         schedule_timeout_uninterruptible(cfs_time_seconds(1));
1619         srpc_data.rpc_matchbits = ((__u64) ktime_get_real_seconds()) << 48;
1620
1621         srpc_data.rpc_state = SRPC_STATE_NONE;
1622
1623         rc = LNetNIInit(LNET_PID_LUSTRE);
1624         if (rc < 0) {
1625                 CERROR("LNetNIInit() has failed: %d\n", rc);
1626                 return rc;
1627         }
1628
1629         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1630
1631         srpc_data.rpc_lnet_handler = srpc_lnet_ev_handler;
1632
1633         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1634         LASSERT(rc == 0);
1635         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1636         LASSERT(rc == 0);
1637
1638         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1639
1640         rc = stt_startup();
1641
1642         if (rc != 0)
1643                 srpc_shutdown();
1644         else
1645                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1646
1647         return rc;
1648 }
1649
1650 void
1651 srpc_shutdown (void)
1652 {
1653         int i;
1654         int rc;
1655         int state;
1656
1657         state = srpc_data.rpc_state;
1658         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1659
1660         switch (state) {
1661         default:
1662                 LBUG();
1663                 fallthrough;
1664         case SRPC_STATE_RUNNING:
1665                 spin_lock(&srpc_data.rpc_glock);
1666
1667                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1668                         struct srpc_service *sv = srpc_data.rpc_services[i];
1669
1670                         LASSERTF(sv == NULL,
1671                                  "service not empty: id %d, name %s\n",
1672                                  i, sv->sv_name);
1673                 }
1674
1675                 spin_unlock(&srpc_data.rpc_glock);
1676
1677                 stt_shutdown();
1678                 fallthrough;
1679
1680         case SRPC_STATE_EQ_INIT:
1681                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1682                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1683                 LASSERT(rc == 0);
1684                 lnet_assert_handler_unused(srpc_data.rpc_lnet_handler);
1685                 fallthrough;
1686
1687         case SRPC_STATE_NI_INIT:
1688                 LNetNIFini();
1689         }
1690 }