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 enum rpc_counter_32 {
  53         SRPC_ERROR,
  54         SRPC_RPC_SENT,
  55         SRPC_RPC_RCVD,
  56         SRPC_RPC_DROP,
  57         SRPC_RPC_EXPIRED,
  58         SRPC_COUNTER32_MAX,
  59 };
  60
  61 enum rpc_counter_64 {
  62         SRPC_BULK_GET,
  63         SRPC_BULK_PUT,
  64         SRPC_COUNTER64_MAX,
  65 };
  66
  67 static struct smoketest_rpc {
  68         spinlock_t       rpc_glock;     /* global lock */
  69         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  70         lnet_handler_t           rpc_lnet_handler;/* _the_ LNet event handler */
  71         enum srpc_state          rpc_state;
  72         atomic_t                 rpc_counters32[SRPC_COUNTER32_MAX];
  73         atomic64_t               rpc_counters64[SRPC_COUNTER64_MAX];
  74         atomic64_t               rpc_matchbits; /* matchbits counter */
  75
  76 } srpc_data;
  77
  78 #define RPC_STAT32(a) \
  79         srpc_data.rpc_counters32[(a)]
  80
  81 #define GET_RPC_STAT32(a) \
  82         atomic_read(&srpc_data.rpc_counters32[(a)])
  83
  84 #define GET_RPC_STAT64(a) \
  85         atomic64_read(&srpc_data.rpc_counters64[(a)])
  86
  87 static inline int
  88 srpc_serv_portal(int svc_id)
  89 {
  90         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  91                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  92 }
  93
  94 /* forward ref's */
  95 static int srpc_handle_rpc(struct swi_workitem *wi);
  96
  97
  98 void srpc_get_counters(struct srpc_counters *cnt)
  99 {
 100         cnt->errors = GET_RPC_STAT32(SRPC_ERROR);
 101         cnt->rpcs_sent = GET_RPC_STAT32(SRPC_RPC_SENT);
 102         cnt->rpcs_rcvd = GET_RPC_STAT32(SRPC_RPC_RCVD);
 103         cnt->rpcs_dropped = GET_RPC_STAT32(SRPC_RPC_DROP);
 104         cnt->rpcs_expired = GET_RPC_STAT32(SRPC_RPC_EXPIRED);
 105
 106         cnt->bulk_get = GET_RPC_STAT64(SRPC_BULK_GET);
 107         cnt->bulk_put = GET_RPC_STAT64(SRPC_BULK_PUT);
 108 }
 109
 110 static int
 111 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
 112                    int nob)
 113 {
 114         LASSERT(off < PAGE_SIZE);
 115         LASSERT(nob > 0 && nob <= PAGE_SIZE);
 116
 117         bk->bk_iovs[i].bv_offset = off;
 118         bk->bk_iovs[i].bv_page   = pg;
 119         bk->bk_iovs[i].bv_len    = nob;
 120         return nob;
 121 }
 122
 123 void
 124 srpc_free_bulk(struct srpc_bulk *bk)
 125 {
 126         int i;
 127         struct page *pg;
 128
 129         LASSERT(bk != NULL);
 130
 131         for (i = 0; i < bk->bk_niov; i++) {
 132                 pg = bk->bk_iovs[i].bv_page;
 133                 if (pg == NULL)
 134                         break;
 135
 136                 __free_page(pg);
 137         }
 138
 139         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
 140 }
 141
 142 struct srpc_bulk *
 143 srpc_alloc_bulk(int cpt, unsigned bulk_off, unsigned bulk_npg,
 144                 unsigned bulk_len, int sink)
 145 {
 146         struct srpc_bulk *bk;
 147         int i;
 148
 149         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 150
 151         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 152                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 153         if (bk == NULL) {
 154                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 155                 return NULL;
 156         }
 157
 158         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 159         bk->bk_sink   = sink;
 160         bk->bk_len    = bulk_len;
 161         bk->bk_niov   = bulk_npg;
 162
 163         for (i = 0; i < bulk_npg; i++) {
 164                 struct page *pg;
 165                 int nob;
 166
 167                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 168                 if (pg == NULL) {
 169                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 170                         srpc_free_bulk(bk);
 171                         return NULL;
 172                 }
 173
 174                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 175                       bulk_off;
 176
 177                 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
 178                 bulk_len -= nob;
 179                 bulk_off = 0;
 180         }
 181
 182         return bk;
 183 }
 184
 185 static inline __u64
 186 srpc_next_id (void)
 187 {
 188         return atomic64_inc_return(&srpc_data.rpc_matchbits);
 189 }
 190
 191 static void
 192 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 193                      struct srpc_service_cd *scd,
 194                      struct srpc_buffer *buffer)
 195 {
 196         memset(rpc, 0, sizeof(*rpc));
 197         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 198                           srpc_serv_is_framework(scd->scd_svc) ?
 199                           lst_sched_serial : lst_sched_test[scd->scd_cpt]);
 200
 201         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 202
 203         rpc->srpc_scd      = scd;
 204         rpc->srpc_reqstbuf = buffer;
 205         rpc->srpc_peer     = buffer->buf_peer;
 206         rpc->srpc_self     = buffer->buf_self;
 207         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 208 }
 209
 210 static void
 211 srpc_service_fini(struct srpc_service *svc)
 212 {
 213         struct srpc_service_cd *scd;
 214         struct srpc_server_rpc *rpc;
 215         struct srpc_buffer *buf;
 216         struct list_head *q;
 217         int i;
 218
 219         if (svc->sv_cpt_data == NULL)
 220                 return;
 221
 222         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 223                 while (1) {
 224                         if (!list_empty(&scd->scd_buf_posted))
 225                                 q = &scd->scd_buf_posted;
 226                         else if (!list_empty(&scd->scd_buf_blocked))
 227                                 q = &scd->scd_buf_blocked;
 228                         else
 229                                 break;
 230
 231                         while (!list_empty(q)) {
 232                                 buf = list_entry(q->next,
 233                                                  struct srpc_buffer,
 234                                                  buf_list);
 235                                 list_del(&buf->buf_list);
 236                                 LIBCFS_FREE(buf, sizeof(*buf));
 237                         }
 238                 }
 239
 240                 LASSERT(list_empty(&scd->scd_rpc_active));
 241
 242                 while (!list_empty(&scd->scd_rpc_free)) {
 243                         rpc = list_entry(scd->scd_rpc_free.next,
 244                                          struct srpc_server_rpc,
 245                                          srpc_list);
 246                         list_del(&rpc->srpc_list);
 247                         LIBCFS_FREE(rpc, sizeof(*rpc));
 248                 }
 249         }
 250
 251         cfs_percpt_free(svc->sv_cpt_data);
 252         svc->sv_cpt_data = NULL;
 253 }
 254
 255 static int
 256 srpc_service_nrpcs(struct srpc_service *svc)
 257 {
 258         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 259
 260         return srpc_serv_is_framework(svc) ?
 261                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 262 }
 263
 264 int srpc_add_buffer(struct swi_workitem *wi);
 265
 266 static int
 267 srpc_service_init(struct srpc_service *svc)
 268 {
 269         struct srpc_service_cd *scd;
 270         struct srpc_server_rpc *rpc;
 271         int nrpcs;
 272         int i;
 273         int j;
 274
 275         svc->sv_shuttingdown = 0;
 276
 277         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 278                                             sizeof(struct srpc_service_cd));
 279         if (svc->sv_cpt_data == NULL)
 280                 return -ENOMEM;
 281
 282         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 283                         1 : cfs_cpt_number(lnet_cpt_table());
 284         nrpcs = srpc_service_nrpcs(svc);
 285
 286         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 287                 scd->scd_cpt = i;
 288                 scd->scd_svc = svc;
 289                 spin_lock_init(&scd->scd_lock);
 290                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 291                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 292                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 293                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 294
 295                 scd->scd_ev.ev_data = scd;
 296                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 297
 298                 /* NB: don't use lst_sched_serial for adding buffer,
 299                  * see details in srpc_service_add_buffers() */
 300                 swi_init_workitem(&scd->scd_buf_wi,
 301                                   srpc_add_buffer, lst_sched_test[i]);
 302
 303                 if (i != 0 && srpc_serv_is_framework(svc)) {
 304                         /* NB: framework service only needs srpc_service_cd for
 305                          * one partition, but we allocate for all to make
 306                          * it easier to implement, it will waste a little
 307                          * memory but nobody should care about this */
 308                         continue;
 309                 }
 310
 311                 for (j = 0; j < nrpcs; j++) {
 312                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 313                                          i, sizeof(*rpc));
 314                         if (rpc == NULL) {
 315                                 srpc_service_fini(svc);
 316                                 return -ENOMEM;
 317                         }
 318                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 319                 }
 320         }
 321
 322         return 0;
 323 }
 324
 325 int
 326 srpc_add_service(struct srpc_service *sv)
 327 {
 328         int id = sv->sv_id;
 329
 330         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 331
 332         if (srpc_service_init(sv) != 0)
 333                 return -ENOMEM;
 334
 335         spin_lock(&srpc_data.rpc_glock);
 336
 337         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 338
 339         if (srpc_data.rpc_services[id] != NULL) {
 340                 spin_unlock(&srpc_data.rpc_glock);
 341                 goto failed;
 342         }
 343
 344         srpc_data.rpc_services[id] = sv;
 345         spin_unlock(&srpc_data.rpc_glock);
 346
 347         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 348         return 0;
 349
 350 failed:
 351         srpc_service_fini(sv);
 352         return -EBUSY;
 353 }
 354
 355 int
 356 srpc_remove_service(struct srpc_service *sv)
 357 {
 358         int id = sv->sv_id;
 359
 360         spin_lock(&srpc_data.rpc_glock);
 361
 362         if (srpc_data.rpc_services[id] != sv) {
 363                 spin_unlock(&srpc_data.rpc_glock);
 364                 return -ENOENT;
 365         }
 366
 367         srpc_data.rpc_services[id] = NULL;
 368         spin_unlock(&srpc_data.rpc_glock);
 369         return 0;
 370 }
 371
 372 static int
 373 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 374                        int len, int options, struct lnet_process_id peer4,
 375                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 376 {
 377         int rc;
 378         struct lnet_md md;
 379         struct lnet_me *me;
 380         struct lnet_processid peer;
 381
 382         peer.pid = peer4.pid;
 383         lnet_nid4_to_nid(peer4.nid, &peer.nid);
 384
 385         me = LNetMEAttach(portal, &peer, matchbits, 0, LNET_UNLINK,
 386                           local ? LNET_INS_LOCAL : LNET_INS_AFTER);
 387         if (IS_ERR(me)) {
 388                 rc = PTR_ERR(me);
 389                 CERROR("LNetMEAttach failed: %d\n", rc);
 390                 LASSERT(rc == -ENOMEM);
 391                 return -ENOMEM;
 392         }
 393
 394         md.threshold = 1;
 395         md.user_ptr  = ev;
 396         md.start     = buf;
 397         md.length    = len;
 398         md.options   = options;
 399         md.handler   = srpc_data.rpc_lnet_handler;
 400
 401         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdh);
 402         if (rc != 0) {
 403                 CERROR("LNetMDAttach failed: %d\n", rc);
 404                 LASSERT(rc == -ENOMEM);
 405
 406                 return -ENOMEM;
 407         }
 408
 409         CDEBUG(D_NET,
 410                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 411                libcfs_id2str(peer4), portal, matchbits);
 412         return 0;
 413 }
 414
 415 static int
 416 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 417                       int options, struct lnet_process_id peer4,
 418                       lnet_nid_t self4, struct lnet_handle_md *mdh,
 419                       struct srpc_event *ev)
 420 {
 421         int rc;
 422         struct lnet_md md;
 423         struct lnet_nid self;
 424         struct lnet_processid peer;
 425
 426         lnet_nid4_to_nid(self4, &self);
 427         lnet_pid4_to_pid(peer4, &peer);
 428
 429         md.user_ptr  = ev;
 430         md.start     = buf;
 431         md.length    = len;
 432         md.handler   = srpc_data.rpc_lnet_handler;
 433         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 434         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 435
 436         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
 437         if (rc != 0) {
 438                 CERROR("LNetMDBind failed: %d\n", rc);
 439                 LASSERT(rc == -ENOMEM);
 440                 return -ENOMEM;
 441         }
 442
 443         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 444          * they're only meaningful for MDs attached to an ME (i.e. passive
 445          * buffers...
 446          */
 447         if ((options & LNET_MD_OP_PUT) != 0) {
 448                 rc = LNetPut(&self, *mdh, LNET_NOACK_REQ, &peer,
 449                              portal, matchbits, 0, 0);
 450         } else {
 451                 LASSERT((options & LNET_MD_OP_GET) != 0);
 452
 453                 rc = LNetGet(&self, *mdh, &peer, portal, matchbits, 0, false);
 454         }
 455
 456         if (rc != 0) {
 457                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 458                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 459                        libcfs_id2str(peer4), portal, matchbits, rc);
 460
 461                 /* The forthcoming unlink event will complete this operation
 462                  * with failure, so fall through and return success here.
 463                  */
 464                 rc = LNetMDUnlink(*mdh);
 465                 LASSERT(rc == 0);
 466         } else {
 467                 CDEBUG(D_NET,
 468                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 469                        libcfs_id2str(peer4), portal, matchbits);
 470         }
 471         return 0;
 472 }
 473
 474 static int
 475 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 476                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 477 {
 478         struct lnet_process_id any = {0};
 479
 480         any.nid = LNET_NID_ANY;
 481         any.pid = LNET_PID_ANY;
 482
 483         return srpc_post_passive_rdma(srpc_serv_portal(service),
 484                                       local, service, buf, len,
 485                                       LNET_MD_OP_PUT, any, mdh, ev);
 486 }
 487
 488 static int
 489 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 490 __must_hold(&scd->scd_lock)
 491 {
 492         struct srpc_service *sv = scd->scd_svc;
 493         struct srpc_msg *msg = &buf->buf_msg;
 494         int rc;
 495
 496         LNetInvalidateMDHandle(&buf->buf_mdh);
 497         list_add(&buf->buf_list, &scd->scd_buf_posted);
 498         scd->scd_buf_nposted++;
 499         spin_unlock(&scd->scd_lock);
 500
 501         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 502                                       !srpc_serv_is_framework(sv),
 503                                       msg, sizeof(*msg), &buf->buf_mdh,
 504                                       &scd->scd_ev);
 505
 506         /* At this point, a RPC (new or delayed) may have arrived in
 507          * msg and its event handler has been called. So we must add
 508          * buf to scd_buf_posted _before_ dropping scd_lock */
 509
 510         spin_lock(&scd->scd_lock);
 511
 512         if (rc == 0) {
 513                 if (!sv->sv_shuttingdown)
 514                         return 0;
 515
 516                 spin_unlock(&scd->scd_lock);
 517                 /* srpc_shutdown_service might have tried to unlink me
 518                  * when my buf_mdh was still invalid */
 519                 LNetMDUnlink(buf->buf_mdh);
 520                 spin_lock(&scd->scd_lock);
 521                 return 0;
 522         }
 523
 524         scd->scd_buf_nposted--;
 525         if (sv->sv_shuttingdown)
 526                 return rc; /* don't allow to change scd_buf_posted */
 527
 528         list_del(&buf->buf_list);
 529         spin_unlock(&scd->scd_lock);
 530
 531         LIBCFS_FREE(buf, sizeof(*buf));
 532
 533         spin_lock(&scd->scd_lock);
 534         return rc;
 535 }
 536
 537 int
 538 srpc_add_buffer(struct swi_workitem *wi)
 539 {
 540         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 541                                                    scd_buf_wi);
 542         struct srpc_buffer *buf;
 543         int rc = 0;
 544
 545         /* it's called by workitem scheduler threads, these threads
 546          * should have been set CPT affinity, so buffers will be posted
 547          * on CPT local list of Portal */
 548         spin_lock(&scd->scd_lock);
 549
 550         while (scd->scd_buf_adjust > 0 &&
 551                !scd->scd_svc->sv_shuttingdown) {
 552                 scd->scd_buf_adjust--; /* consume it */
 553                 scd->scd_buf_posting++;
 554
 555                 spin_unlock(&scd->scd_lock);
 556
 557                 LIBCFS_ALLOC(buf, sizeof(*buf));
 558                 if (buf == NULL) {
 559                         CERROR("Failed to add new buf to service: %s\n",
 560                                scd->scd_svc->sv_name);
 561                         spin_lock(&scd->scd_lock);
 562                         rc = -ENOMEM;
 563                         break;
 564                 }
 565
 566                 spin_lock(&scd->scd_lock);
 567                 if (scd->scd_svc->sv_shuttingdown) {
 568                         spin_unlock(&scd->scd_lock);
 569                         LIBCFS_FREE(buf, sizeof(*buf));
 570
 571                         spin_lock(&scd->scd_lock);
 572                         rc = -ESHUTDOWN;
 573                         break;
 574                 }
 575
 576                 rc = srpc_service_post_buffer(scd, buf);
 577                 if (rc != 0)
 578                         break; /* buf has been freed inside */
 579
 580                 LASSERT(scd->scd_buf_posting > 0);
 581                 scd->scd_buf_posting--;
 582                 scd->scd_buf_total++;
 583                 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
 584         }
 585
 586         if (rc != 0) {
 587                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 588                 scd->scd_buf_err = rc;
 589
 590                 LASSERT(scd->scd_buf_posting > 0);
 591                 scd->scd_buf_posting--;
 592         }
 593
 594         spin_unlock(&scd->scd_lock);
 595         return 0;
 596 }
 597
 598 int
 599 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 600 {
 601         struct srpc_service_cd *scd;
 602         int rc = 0;
 603         int i;
 604
 605         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 606
 607         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 608                 spin_lock(&scd->scd_lock);
 609
 610                 scd->scd_buf_err = 0;
 611                 scd->scd_buf_err_stamp = 0;
 612                 scd->scd_buf_posting = 0;
 613                 scd->scd_buf_adjust = nbuffer;
 614                 /* start to post buffers */
 615                 swi_schedule_workitem(&scd->scd_buf_wi);
 616                 spin_unlock(&scd->scd_lock);
 617
 618                 /* framework service only post buffer for one partition  */
 619                 if (srpc_serv_is_framework(sv))
 620                         break;
 621         }
 622
 623         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 624                 spin_lock(&scd->scd_lock);
 625                 /*
 626                  * NB: srpc_service_add_buffers() can be called inside
 627                  * thread context of lst_sched_serial, and we don't normally
 628                  * allow to sleep inside thread context of WI scheduler
 629                  * because it will block current scheduler thread from doing
 630                  * anything else, even worse, it could deadlock if it's
 631                  * waiting on result from another WI of the same scheduler.
 632                  * However, it's safe at here because scd_buf_wi is scheduled
 633                  * by thread in a different WI scheduler (lst_sched_test),
 634                  * so we don't have any risk of deadlock, though this could
 635                  * block all WIs pending on lst_sched_serial for a moment
 636                  * which is not good but not fatal.
 637                  */
 638                 lst_wait_until(scd->scd_buf_err != 0 ||
 639                                (scd->scd_buf_adjust == 0 &&
 640                                 scd->scd_buf_posting == 0),
 641                                scd->scd_lock, "waiting for adding buffer\n");
 642
 643                 if (scd->scd_buf_err != 0 && rc == 0)
 644                         rc = scd->scd_buf_err;
 645
 646                 spin_unlock(&scd->scd_lock);
 647         }
 648
 649         return rc;
 650 }
 651
 652 void
 653 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 654 {
 655         struct srpc_service_cd *scd;
 656         int num;
 657         int i;
 658
 659         LASSERT(!sv->sv_shuttingdown);
 660
 661         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 662                 spin_lock(&scd->scd_lock);
 663
 664                 num = scd->scd_buf_total + scd->scd_buf_posting;
 665                 scd->scd_buf_adjust -= min(nbuffer, num);
 666
 667                 spin_unlock(&scd->scd_lock);
 668         }
 669 }
 670
 671 /* returns 1 if sv has finished, otherwise 0 */
 672 int
 673 srpc_finish_service(struct srpc_service *sv)
 674 {
 675         struct srpc_service_cd *scd;
 676         struct srpc_server_rpc *rpc;
 677         int i;
 678
 679         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 680
 681         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 682                 spin_lock(&scd->scd_lock);
 683                 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
 684                         spin_unlock(&scd->scd_lock);
 685                         return 0;
 686                 }
 687
 688                 if (scd->scd_buf_nposted > 0) {
 689                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 690                                scd->scd_buf_nposted);
 691                         spin_unlock(&scd->scd_lock);
 692                         return 0;
 693                 }
 694
 695                 if (list_empty(&scd->scd_rpc_active)) {
 696                         spin_unlock(&scd->scd_lock);
 697                         continue;
 698                 }
 699
 700                 rpc = list_entry(scd->scd_rpc_active.next,
 701                                  struct srpc_server_rpc, srpc_list);
 702                 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",
 703                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 704                         swi_state2str(rpc->srpc_wi.swi_state),
 705                         rpc->srpc_wi.swi_workitem.wi_scheduled,
 706                         rpc->srpc_wi.swi_workitem.wi_running,
 707                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 708                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 709                 spin_unlock(&scd->scd_lock);
 710                 return 0;
 711         }
 712
 713         /* no lock needed from now on */
 714         srpc_service_fini(sv);
 715         return 1;
 716 }
 717
 718 /* called with sv->sv_lock held */
 719 static void
 720 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 721                             struct srpc_buffer *buf)
 722 __must_hold(&scd->scd_lock)
 723 {
 724         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 725                 if (srpc_service_post_buffer(scd, buf) != 0) {
 726                         CWARN("Failed to post %s buffer\n",
 727                               scd->scd_svc->sv_name);
 728                 }
 729                 return;
 730         }
 731
 732         /* service is shutting down, or we want to recycle some buffers */
 733         scd->scd_buf_total--;
 734
 735         if (scd->scd_buf_adjust < 0) {
 736                 scd->scd_buf_adjust++;
 737                 if (scd->scd_buf_adjust < 0 &&
 738                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 739                         CDEBUG(D_INFO,
 740                                "Try to recyle %d buffers but nothing left\n",
 741                                scd->scd_buf_adjust);
 742                         scd->scd_buf_adjust = 0;
 743                 }
 744         }
 745
 746         spin_unlock(&scd->scd_lock);
 747         LIBCFS_FREE(buf, sizeof(*buf));
 748         spin_lock(&scd->scd_lock);
 749 }
 750
 751 void
 752 srpc_abort_service(struct srpc_service *sv)
 753 {
 754         struct srpc_service_cd *scd;
 755         struct srpc_server_rpc *rpc;
 756         int i;
 757
 758         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 759                sv->sv_id, sv->sv_name);
 760
 761         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 762                 spin_lock(&scd->scd_lock);
 763
 764                 /* schedule in-flight RPCs to notice the abort, NB:
 765                  * racing with incoming RPCs; complete fix should make test
 766                  * RPCs carry session ID in its headers
 767                  */
 768                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 769                         rpc->srpc_aborted = 1;
 770                         swi_schedule_workitem(&rpc->srpc_wi);
 771                 }
 772
 773                 spin_unlock(&scd->scd_lock);
 774         }
 775 }
 776
 777 void
 778 srpc_shutdown_service(struct srpc_service *sv)
 779 {
 780         struct srpc_service_cd *scd;
 781         struct srpc_server_rpc *rpc;
 782         struct srpc_buffer *buf;
 783         int i;
 784
 785         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 786                sv->sv_id, sv->sv_name);
 787
 788         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 789                 spin_lock(&scd->scd_lock);
 790
 791         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 792
 793         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 794                 spin_unlock(&scd->scd_lock);
 795
 796         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 797                 spin_lock(&scd->scd_lock);
 798
 799                 /* schedule in-flight RPCs to notice the shutdown */
 800                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 801                         swi_schedule_workitem(&rpc->srpc_wi);
 802
 803                 spin_unlock(&scd->scd_lock);
 804
 805                 /* OK to traverse scd_buf_posted without lock, since no one
 806                  * touches scd_buf_posted now
 807                  */
 808                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 809                         LNetMDUnlink(buf->buf_mdh);
 810         }
 811 }
 812
 813 static int
 814 srpc_send_request(struct srpc_client_rpc *rpc)
 815 {
 816         struct srpc_event *ev = &rpc->crpc_reqstev;
 817         int rc;
 818
 819         ev->ev_fired = 0;
 820         ev->ev_data  = rpc;
 821         ev->ev_type  = SRPC_REQUEST_SENT;
 822
 823         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 824                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 825                                    sizeof(struct srpc_msg), LNET_MD_OP_PUT,
 826                                    rpc->crpc_dest, LNET_NID_ANY,
 827                                    &rpc->crpc_reqstmdh, ev);
 828         if (rc != 0) {
 829                 LASSERT(rc == -ENOMEM);
 830                 ev->ev_fired = 1;  /* no more event expected */
 831         }
 832         return rc;
 833 }
 834
 835 static int
 836 srpc_prepare_reply(struct srpc_client_rpc *rpc)
 837 {
 838         struct srpc_event *ev = &rpc->crpc_replyev;
 839         u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 840         int rc;
 841
 842         ev->ev_fired = 0;
 843         ev->ev_data  = rpc;
 844         ev->ev_type  = SRPC_REPLY_RCVD;
 845
 846         *id = srpc_next_id();
 847
 848         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 849                                     &rpc->crpc_replymsg,
 850                                     sizeof(struct srpc_msg),
 851                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 852                                     &rpc->crpc_replymdh, ev);
 853         if (rc != 0) {
 854                 LASSERT(rc == -ENOMEM);
 855                 ev->ev_fired = 1;  /* no more event expected */
 856         }
 857         return rc;
 858 }
 859
 860 static int
 861 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 862 {
 863         struct srpc_bulk *bk = &rpc->crpc_bulk;
 864         struct srpc_event *ev = &rpc->crpc_bulkev;
 865         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 866         int rc;
 867         int opt;
 868
 869         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 870
 871         /* nothing to do */
 872         if (bk->bk_niov == 0)
 873                 return 0;
 874
 875         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 876         opt |= LNET_MD_KIOV;
 877
 878         ev->ev_fired = 0;
 879         ev->ev_data  = rpc;
 880         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 881
 882         *id = srpc_next_id();
 883
 884         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 885                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 886                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 887         if (rc != 0) {
 888                 LASSERT(rc == -ENOMEM);
 889                 ev->ev_fired = 1;  /* no more event expected */
 890         }
 891         return rc;
 892 }
 893
 894 static int
 895 srpc_do_bulk(struct srpc_server_rpc *rpc)
 896 {
 897         struct srpc_event *ev = &rpc->srpc_ev;
 898         struct srpc_bulk *bk = rpc->srpc_bulk;
 899         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 900         int rc;
 901         int opt;
 902
 903         LASSERT(bk != NULL);
 904
 905         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 906         opt |= LNET_MD_KIOV;
 907
 908         ev->ev_fired = 0;
 909         ev->ev_data  = rpc;
 910         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 911
 912         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 913                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 914                                    rpc->srpc_peer, rpc->srpc_self,
 915                                    &bk->bk_mdh, ev);
 916         if (rc != 0)
 917                 ev->ev_fired = 1;  /* no more event expected */
 918         return rc;
 919 }
 920
 921 /* only called from srpc_handle_rpc */
 922 static void
 923 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 924 {
 925         struct srpc_service_cd  *scd = rpc->srpc_scd;
 926         struct srpc_service     *sv  = scd->scd_svc;
 927         struct srpc_buffer *buffer;
 928
 929         LASSERT(status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
 930
 931         rpc->srpc_status = status;
 932
 933         CDEBUG_LIMIT(status == 0 ? D_NET : D_NETERROR,
 934                      "Server RPC %p done: service %s, peer %s, status %s:%d\n",
 935                      rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 936                      swi_state2str(rpc->srpc_wi.swi_state), status);
 937
 938         if (status != 0)
 939                 atomic_inc(&RPC_STAT32(SRPC_RPC_DROP));
 940
 941         if (rpc->srpc_done != NULL)
 942                 (*rpc->srpc_done) (rpc);
 943         LASSERT(rpc->srpc_bulk == NULL);
 944
 945         spin_lock(&scd->scd_lock);
 946
 947         if (rpc->srpc_reqstbuf != NULL) {
 948                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 949                  * sv won't go away for scd_rpc_active must not be empty
 950                  */
 951                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 952                 rpc->srpc_reqstbuf = NULL;
 953         }
 954
 955         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 956
 957         /*
 958          * No one can schedule me now since:
 959          * - I'm not on scd_rpc_active.
 960          * - all LNet events have been fired.
 961          * Cancel pending schedules and prevent future schedule attempts:
 962          */
 963         LASSERT(rpc->srpc_ev.ev_fired);
 964         swi_exit_workitem(&rpc->srpc_wi);
 965
 966         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 967                 buffer = list_entry(scd->scd_buf_blocked.next,
 968                                     struct srpc_buffer, buf_list);
 969                 list_del(&buffer->buf_list);
 970
 971                 srpc_init_server_rpc(rpc, scd, buffer);
 972                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 973                 swi_schedule_workitem(&rpc->srpc_wi);
 974         } else {
 975                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 976         }
 977
 978         spin_unlock(&scd->scd_lock);
 979 }
 980
 981 /* handles an incoming RPC */
 982 static int srpc_handle_rpc(struct swi_workitem *wi)
 983 {
 984         struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc,
 985                                                    srpc_wi);
 986         struct srpc_service_cd *scd = rpc->srpc_scd;
 987         struct srpc_service *sv = scd->scd_svc;
 988         struct srpc_event *ev = &rpc->srpc_ev;
 989         int rc = 0;
 990
 991         LASSERT(wi == &rpc->srpc_wi);
 992
 993         spin_lock(&scd->scd_lock);
 994
 995         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 996                 spin_unlock(&scd->scd_lock);
 997
 998                 if (rpc->srpc_bulk != NULL)
 999                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
1000                 LNetMDUnlink(rpc->srpc_replymdh);
1001
1002                 if (ev->ev_fired) { /* no more event, OK to finish */
1003                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
1004                         return 1;
1005                 }
1006                 return 0;
1007         }
1008
1009         spin_unlock(&scd->scd_lock);
1010
1011         switch (wi->swi_state) {
1012         default:
1013                 LBUG();
1014                 fallthrough;
1015         case SWI_STATE_NEWBORN: {
1016                 struct srpc_msg *msg;
1017                 struct srpc_generic_reply *reply;
1018
1019                 msg = &rpc->srpc_reqstbuf->buf_msg;
1020                 reply = &rpc->srpc_replymsg.msg_body.reply;
1021
1022                 if (msg->msg_magic == 0) {
1023                         /* moaned already in srpc_lnet_ev_handler */
1024                         srpc_server_rpc_done(rpc, EBADMSG);
1025                         return 1;
1026                 }
1027
1028                 srpc_unpack_msg_hdr(msg);
1029                 if (msg->msg_version != SRPC_MSG_VERSION) {
1030                         CWARN("Version mismatch: %u, %u expected, from %s\n",
1031                               msg->msg_version, SRPC_MSG_VERSION,
1032                               libcfs_id2str(rpc->srpc_peer));
1033                         reply->status = EPROTO;
1034                         /* drop through and send reply */
1035                 } else {
1036                         reply->status = 0;
1037                         rc = (*sv->sv_handler)(rpc);
1038                         LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1039                         if (rc != 0) {
1040                                 srpc_server_rpc_done(rpc, rc);
1041                                 return 1;
1042                         }
1043                 }
1044
1045                 wi->swi_state = SWI_STATE_BULK_STARTED;
1046
1047                 if (rpc->srpc_bulk != NULL) {
1048                         rc = srpc_do_bulk(rpc);
1049                         if (rc == 0)
1050                                 return 0; /* wait for bulk */
1051
1052                         LASSERT(ev->ev_fired);
1053                         ev->ev_status = rc;
1054                 }
1055         }
1056         fallthrough;
1057         case SWI_STATE_BULK_STARTED:
1058                 LASSERT(rpc->srpc_bulk == NULL || ev->ev_fired);
1059
1060                 if (rpc->srpc_bulk != NULL) {
1061                         rc = ev->ev_status;
1062
1063                         if (sv->sv_bulk_ready != NULL)
1064                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
1065
1066                         if (rc != 0) {
1067                                 srpc_server_rpc_done(rpc, rc);
1068                                 return 1;
1069                         }
1070                 }
1071
1072                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1073                 rc = srpc_send_reply(rpc);
1074                 if (rc == 0)
1075                         return 0; /* wait for reply */
1076                 srpc_server_rpc_done(rpc, rc);
1077                 return 1;
1078
1079         case SWI_STATE_REPLY_SUBMITTED:
1080                 if (!ev->ev_fired) {
1081                         CERROR("RPC %p: bulk %p, service %d\n",
1082                                rpc, rpc->srpc_bulk, sv->sv_id);
1083                         CERROR("Event: status %d, type %d, lnet %d\n",
1084                                ev->ev_status, ev->ev_type, ev->ev_lnet);
1085                         LASSERT(ev->ev_fired);
1086                 }
1087
1088                 wi->swi_state = SWI_STATE_DONE;
1089                 srpc_server_rpc_done(rpc, ev->ev_status);
1090                 return 1;
1091         }
1092
1093         return 0;
1094 }
1095
1096 static void
1097 srpc_client_rpc_expired (void *data)
1098 {
1099         struct srpc_client_rpc *rpc = data;
1100
1101         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1102               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1103               rpc->crpc_timeout);
1104
1105         spin_lock(&rpc->crpc_lock);
1106
1107         rpc->crpc_timeout = 0;
1108         srpc_abort_rpc(rpc, -ETIMEDOUT);
1109
1110         spin_unlock(&rpc->crpc_lock);
1111
1112         atomic_inc(&RPC_STAT32(SRPC_RPC_EXPIRED));
1113 }
1114
1115 static void
1116 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1117 {
1118         struct stt_timer *timer = &rpc->crpc_timer;
1119
1120         if (rpc->crpc_timeout == 0)
1121                 return;
1122
1123         INIT_LIST_HEAD(&timer->stt_list);
1124         timer->stt_data    = rpc;
1125         timer->stt_func    = srpc_client_rpc_expired;
1126         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1127         stt_add_timer(timer);
1128 }
1129
1130 /*
1131  * Called with rpc->crpc_lock held.
1132  *
1133  * Upon exit the RPC expiry timer is not queued and the handler is not
1134  * running on any CPU.
1135  */
1136 static void
1137 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1138 {
1139         /* timer not planted or already exploded */
1140         if (rpc->crpc_timeout == 0)
1141                 return;
1142
1143         /* timer successfully defused */
1144         if (stt_del_timer(&rpc->crpc_timer))
1145                 return;
1146
1147         /* timer detonated, wait for it to explode */
1148         while (rpc->crpc_timeout != 0) {
1149                 spin_unlock(&rpc->crpc_lock);
1150
1151                 schedule();
1152
1153                 spin_lock(&rpc->crpc_lock);
1154         }
1155 }
1156
1157 static void
1158 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1159 {
1160         struct swi_workitem *wi = &rpc->crpc_wi;
1161
1162         LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1163
1164         spin_lock(&rpc->crpc_lock);
1165
1166         rpc->crpc_closed = 1;
1167         if (rpc->crpc_status == 0)
1168                 rpc->crpc_status = status;
1169
1170         srpc_del_client_rpc_timer(rpc);
1171
1172         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1173                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1174                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1175                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1176
1177         /*
1178          * No one can schedule me now since:
1179          * - RPC timer has been defused.
1180          * - all LNet events have been fired.
1181          * - crpc_closed has been set, preventing srpc_abort_rpc from
1182          *   scheduling me.
1183          * Cancel pending schedules and prevent future schedule attempts:
1184          */
1185         LASSERT(!srpc_event_pending(rpc));
1186         swi_exit_workitem(wi);
1187
1188         spin_unlock(&rpc->crpc_lock);
1189
1190         (*rpc->crpc_done)(rpc);
1191 }
1192
1193 /* sends an outgoing RPC */
1194 int
1195 srpc_send_rpc(struct swi_workitem *wi)
1196 {
1197         int rc = 0;
1198         struct srpc_client_rpc *rpc;
1199         struct srpc_msg *reply;
1200         int do_bulk;
1201
1202         LASSERT(wi != NULL);
1203
1204         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1205
1206         LASSERT(rpc != NULL);
1207         LASSERT(wi == &rpc->crpc_wi);
1208
1209         reply = &rpc->crpc_replymsg;
1210         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1211
1212         spin_lock(&rpc->crpc_lock);
1213
1214         if (rpc->crpc_aborted) {
1215                 spin_unlock(&rpc->crpc_lock);
1216                 goto abort;
1217         }
1218
1219         spin_unlock(&rpc->crpc_lock);
1220
1221         switch (wi->swi_state) {
1222         default:
1223                 LBUG();
1224         case SWI_STATE_NEWBORN:
1225                 LASSERT(!srpc_event_pending(rpc));
1226
1227                 rc = srpc_prepare_reply(rpc);
1228                 if (rc != 0) {
1229                         srpc_client_rpc_done(rpc, rc);
1230                         return 1;
1231                 }
1232
1233                 rc = srpc_prepare_bulk(rpc);
1234                 if (rc != 0)
1235                         break;
1236
1237                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1238                 rc = srpc_send_request(rpc);
1239                 break;
1240
1241         case SWI_STATE_REQUEST_SUBMITTED:
1242                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1243                  * order; however, they're processed in a strict order:
1244                  * rqt, rpy, and bulk.
1245                  */
1246                 if (!rpc->crpc_reqstev.ev_fired)
1247                         break;
1248
1249                 rc = rpc->crpc_reqstev.ev_status;
1250                 if (rc != 0)
1251                         break;
1252
1253                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1254                 fallthrough;
1255         case SWI_STATE_REQUEST_SENT: {
1256                 enum srpc_msg_type type;
1257
1258                 type = srpc_service2reply(rpc->crpc_service);
1259
1260                 if (!rpc->crpc_replyev.ev_fired)
1261                         break;
1262
1263                 rc = rpc->crpc_replyev.ev_status;
1264                 if (rc != 0)
1265                         break;
1266
1267                 srpc_unpack_msg_hdr(reply);
1268                 if (reply->msg_type != type ||
1269                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1270                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1271                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1272                               libcfs_id2str(rpc->crpc_dest),
1273                               reply->msg_type, type,
1274                               reply->msg_magic, SRPC_MSG_MAGIC);
1275                         rc = -EBADMSG;
1276                         break;
1277                 }
1278
1279                 if (do_bulk && reply->msg_body.reply.status != 0) {
1280                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1281                               reply->msg_body.reply.status,
1282                               libcfs_id2str(rpc->crpc_dest));
1283                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1284                 }
1285
1286                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1287         }
1288         fallthrough;
1289         case SWI_STATE_REPLY_RECEIVED:
1290                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1291                         break;
1292
1293                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1294
1295                 /* Bulk buffer was unlinked due to remote error. Clear error
1296                  * since reply buffer still contains valid data.
1297                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1298                  * remote error.
1299                  */
1300                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1301                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1302                         rc = 0;
1303
1304                 wi->swi_state = SWI_STATE_DONE;
1305                 srpc_client_rpc_done(rpc, rc);
1306                 return 1;
1307         }
1308
1309         if (rc != 0) {
1310                 spin_lock(&rpc->crpc_lock);
1311                 srpc_abort_rpc(rpc, rc);
1312                 spin_unlock(&rpc->crpc_lock);
1313         }
1314
1315 abort:
1316         if (rpc->crpc_aborted) {
1317                 LNetMDUnlink(rpc->crpc_reqstmdh);
1318                 LNetMDUnlink(rpc->crpc_replymdh);
1319                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1320
1321                 if (!srpc_event_pending(rpc)) {
1322                         srpc_client_rpc_done(rpc, -EINTR);
1323                         return 1;
1324                 }
1325         }
1326         return 0;
1327 }
1328
1329 struct srpc_client_rpc *
1330 srpc_create_client_rpc(struct lnet_process_id peer, int service,
1331                        int nbulkiov, int bulklen,
1332                        void (*rpc_done)(struct srpc_client_rpc *),
1333                        void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1334 {
1335         struct srpc_client_rpc *rpc;
1336
1337         LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1338                                    crpc_bulk.bk_iovs[nbulkiov]));
1339         if (rpc == NULL)
1340                 return NULL;
1341
1342         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1343                              bulklen, rpc_done, rpc_fini, priv);
1344         return rpc;
1345 }
1346
1347 /* called with rpc->crpc_lock held */
1348 void
1349 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1350 {
1351         LASSERT(why != 0);
1352
1353         if (rpc->crpc_aborted || /* already aborted */
1354             rpc->crpc_closed)    /* callback imminent */
1355                 return;
1356
1357         CDEBUG(D_NET,
1358                "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1359                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1360                swi_state2str(rpc->crpc_wi.swi_state), why);
1361
1362         rpc->crpc_aborted = 1;
1363         rpc->crpc_status  = why;
1364         swi_schedule_workitem(&rpc->crpc_wi);
1365 }
1366
1367 /* called with rpc->crpc_lock held */
1368 void
1369 srpc_post_rpc(struct srpc_client_rpc *rpc)
1370 {
1371         LASSERT(!rpc->crpc_aborted);
1372         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1373
1374         CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1375                libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1376                rpc->crpc_timeout);
1377
1378         srpc_add_client_rpc_timer(rpc);
1379         swi_schedule_workitem(&rpc->crpc_wi);
1380 }
1381
1382
1383 int
1384 srpc_send_reply(struct srpc_server_rpc *rpc)
1385 {
1386         struct srpc_event *ev = &rpc->srpc_ev;
1387         struct srpc_msg *msg = &rpc->srpc_replymsg;
1388         struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1389         struct srpc_service_cd *scd = rpc->srpc_scd;
1390         struct srpc_service *sv = scd->scd_svc;
1391         __u64 rpyid;
1392         int rc;
1393
1394         LASSERT(buffer != NULL);
1395         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1396
1397         spin_lock(&scd->scd_lock);
1398
1399         if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1400                 /* Repost buffer before replying since test client
1401                  * might send me another RPC once it gets the reply
1402                  */
1403                 if (srpc_service_post_buffer(scd, buffer) != 0)
1404                         CWARN("Failed to repost %s buffer\n", sv->sv_name);
1405                 rpc->srpc_reqstbuf = NULL;
1406         }
1407
1408         spin_unlock(&scd->scd_lock);
1409
1410         ev->ev_fired = 0;
1411         ev->ev_data  = rpc;
1412         ev->ev_type  = SRPC_REPLY_SENT;
1413
1414         msg->msg_magic   = SRPC_MSG_MAGIC;
1415         msg->msg_version = SRPC_MSG_VERSION;
1416         msg->msg_type    = srpc_service2reply(sv->sv_id);
1417
1418         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1419                                    sizeof(*msg), LNET_MD_OP_PUT,
1420                                    rpc->srpc_peer, rpc->srpc_self,
1421                                    &rpc->srpc_replymdh, ev);
1422         if (rc != 0)
1423                 ev->ev_fired = 1;  /* no more event expected */
1424         return rc;
1425 }
1426
1427 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1428 static void
1429 srpc_lnet_ev_handler(struct lnet_event *ev)
1430 {
1431         struct srpc_service_cd *scd;
1432         struct srpc_event *rpcev = ev->md_user_ptr;
1433         struct srpc_client_rpc *crpc;
1434         struct srpc_server_rpc *srpc;
1435         struct srpc_buffer *buffer;
1436         struct srpc_service *sv;
1437         struct srpc_msg *msg;
1438         enum srpc_msg_type type;
1439
1440         LASSERT(!in_interrupt());
1441
1442         if (ev->status != 0) {
1443                 __u32 errors;
1444
1445
1446                 if (ev->status != -ECANCELED) /* cancellation is not error */
1447                         errors = atomic_inc_return(&RPC_STAT32(SRPC_ERROR));
1448                 else
1449                         errors = atomic_read(&RPC_STAT32(SRPC_ERROR));
1450
1451                 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1452                         ev->status, ev->type, errors);
1453         }
1454
1455         rpcev->ev_lnet = ev->type;
1456
1457         switch (rpcev->ev_type) {
1458         default:
1459                 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1460                        rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1461                 LBUG();
1462                 fallthrough;
1463         case SRPC_REQUEST_SENT:
1464                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK)
1465                         atomic_inc(&RPC_STAT32(SRPC_RPC_SENT));
1466
1467                 fallthrough;
1468         case SRPC_REPLY_RCVD:
1469         case SRPC_BULK_REQ_RCVD:
1470                 crpc = rpcev->ev_data;
1471
1472                 if (rpcev != &crpc->crpc_reqstev &&
1473                     rpcev != &crpc->crpc_replyev &&
1474                     rpcev != &crpc->crpc_bulkev) {
1475                         CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1476                                rpcev, crpc, &crpc->crpc_reqstev,
1477                                &crpc->crpc_replyev, &crpc->crpc_bulkev);
1478                         CERROR("Bad event: status %d, type %d, lnet %d\n",
1479                                rpcev->ev_status, rpcev->ev_type,
1480                                rpcev->ev_lnet);
1481                         LBUG();
1482                 }
1483
1484                 spin_lock(&crpc->crpc_lock);
1485
1486                 LASSERT(rpcev->ev_fired == 0);
1487                 rpcev->ev_fired  = 1;
1488                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1489                                    -EINTR : ev->status;
1490                 swi_schedule_workitem(&crpc->crpc_wi);
1491
1492                 spin_unlock(&crpc->crpc_lock);
1493                 break;
1494
1495         case SRPC_REQUEST_RCVD:
1496                 scd = rpcev->ev_data;
1497                 sv = scd->scd_svc;
1498
1499                 LASSERT(rpcev == &scd->scd_ev);
1500
1501                 spin_lock(&scd->scd_lock);
1502
1503                 LASSERT(ev->unlinked);
1504                 LASSERT(ev->type == LNET_EVENT_PUT ||
1505                         ev->type == LNET_EVENT_UNLINK);
1506                 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1507                         sv->sv_shuttingdown);
1508
1509                 buffer = container_of(ev->md_start, struct srpc_buffer,
1510                                       buf_msg);
1511                 buffer->buf_peer = lnet_pid_to_pid4(&ev->source);
1512                 buffer->buf_self = lnet_nid_to_nid4(&ev->target.nid);
1513
1514                 LASSERT(scd->scd_buf_nposted > 0);
1515                 scd->scd_buf_nposted--;
1516
1517                 if (sv->sv_shuttingdown) {
1518                         /* Leave buffer on scd->scd_buf_nposted since
1519                          * srpc_finish_service needs to traverse it.
1520                          */
1521                         spin_unlock(&scd->scd_lock);
1522                         break;
1523                 }
1524
1525                 if (scd->scd_buf_err_stamp != 0 &&
1526                     scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1527                         /* re-enable adding buffer */
1528                         scd->scd_buf_err_stamp = 0;
1529                         scd->scd_buf_err = 0;
1530                 }
1531
1532                 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1533                     scd->scd_buf_adjust == 0 &&
1534                     scd->scd_buf_nposted < scd->scd_buf_low) {
1535                         scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1536                                                   SFW_TEST_WI_MIN);
1537                         swi_schedule_workitem(&scd->scd_buf_wi);
1538                 }
1539
1540                 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1541                 msg = &buffer->buf_msg;
1542                 type = srpc_service2request(sv->sv_id);
1543
1544                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1545                     (msg->msg_type != type &&
1546                      msg->msg_type != __swab32(type)) ||
1547                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1548                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1549                         CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1550                                sv->sv_name, libcfs_idstr(&ev->initiator),
1551                                ev->status, ev->mlength,
1552                                msg->msg_type, msg->msg_magic);
1553
1554                         /* NB can't call srpc_service_recycle_buffer here since
1555                          * it may call LNetM[DE]Attach. The invalid magic tells
1556                          * srpc_handle_rpc to drop this RPC
1557                          */
1558                         msg->msg_magic = 0;
1559                 }
1560
1561                 if (!list_empty(&scd->scd_rpc_free)) {
1562                         srpc = list_entry(scd->scd_rpc_free.next,
1563                                           struct srpc_server_rpc,
1564                                           srpc_list);
1565                         list_del(&srpc->srpc_list);
1566
1567                         srpc_init_server_rpc(srpc, scd, buffer);
1568                         list_add_tail(&srpc->srpc_list,
1569                                       &scd->scd_rpc_active);
1570                         swi_schedule_workitem(&srpc->srpc_wi);
1571                 } else {
1572                         list_add_tail(&buffer->buf_list,
1573                                       &scd->scd_buf_blocked);
1574                 }
1575
1576                 spin_unlock(&scd->scd_lock);
1577
1578                 atomic_inc(&RPC_STAT32(SRPC_RPC_RCVD));
1579                 break;
1580
1581         case SRPC_BULK_GET_RPLD:
1582                 LASSERT(ev->type == LNET_EVENT_SEND ||
1583                         ev->type == LNET_EVENT_REPLY ||
1584                         ev->type == LNET_EVENT_UNLINK);
1585
1586                 if (!ev->unlinked)
1587                         break; /* wait for final event */
1588                 fallthrough;
1589         case SRPC_BULK_PUT_SENT:
1590                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1591                         atomic64_t *data;
1592
1593                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1594                                 data = &srpc_data.rpc_counters64[SRPC_BULK_GET];
1595                         else
1596                                 data = &srpc_data.rpc_counters64[SRPC_BULK_PUT];
1597
1598                         atomic64_add(ev->mlength, data);
1599                 }
1600                 fallthrough;
1601         case SRPC_REPLY_SENT:
1602                 srpc = rpcev->ev_data;
1603                 scd  = srpc->srpc_scd;
1604
1605                 LASSERT(rpcev == &srpc->srpc_ev);
1606
1607                 spin_lock(&scd->scd_lock);
1608
1609                 rpcev->ev_fired  = 1;
1610                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1611                                    -EINTR : ev->status;
1612                 swi_schedule_workitem(&srpc->srpc_wi);
1613
1614                 spin_unlock(&scd->scd_lock);
1615                 break;
1616         }
1617 }
1618
1619
1620 int
1621 srpc_startup (void)
1622 {
1623         int rc;
1624
1625         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1626         spin_lock_init(&srpc_data.rpc_glock);
1627
1628         /* 1 second pause to avoid timestamp reuse */
1629         schedule_timeout_uninterruptible(cfs_time_seconds(1));
1630         atomic64_set(&srpc_data.rpc_matchbits,
1631                         ((__u64)ktime_get_real_seconds() << 48));
1632
1633         srpc_data.rpc_state = SRPC_STATE_NONE;
1634
1635         rc = LNetNIInit(LNET_PID_LUSTRE);
1636         if (rc < 0) {
1637                 CERROR("LNetNIInit() has failed: %d\n", rc);
1638                 return rc;
1639         }
1640
1641         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1642
1643         srpc_data.rpc_lnet_handler = srpc_lnet_ev_handler;
1644
1645         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1646         LASSERT(rc == 0);
1647         rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1648         LASSERT(rc == 0);
1649
1650         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1651
1652         rc = stt_startup();
1653
1654         if (rc != 0)
1655                 srpc_shutdown();
1656         else
1657                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1658
1659         return rc;
1660 }
1661
1662 void
1663 srpc_shutdown (void)
1664 {
1665         int i;
1666         int rc;
1667         int state;
1668
1669         state = srpc_data.rpc_state;
1670         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1671
1672         switch (state) {
1673         default:
1674                 LBUG();
1675                 fallthrough;
1676         case SRPC_STATE_RUNNING:
1677                 spin_lock(&srpc_data.rpc_glock);
1678
1679                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1680                         struct srpc_service *sv = srpc_data.rpc_services[i];
1681
1682                         LASSERTF(sv == NULL,
1683                                  "service not empty: id %d, name %s\n",
1684                                  i, sv->sv_name);
1685                 }
1686
1687                 spin_unlock(&srpc_data.rpc_glock);
1688
1689                 stt_shutdown();
1690                 fallthrough;
1691
1692         case SRPC_STATE_EQ_INIT:
1693                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1694                 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1695                 LASSERT(rc == 0);
1696                 lnet_assert_handler_unused(srpc_data.rpc_lnet_handler);
1697                 fallthrough;
1698
1699         case SRPC_STATE_NI_INIT:
1700                 LNetNIFini();
1701         }
1702 }