lnet/selftest/rpc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   5  * Use is subject to license terms.
   6  *
   7  * Copyright (c) 2012, 2017, Intel Corporation.
   8  */
   9
  10 /*
  11  * This file is part of Lustre, http://www.lustre.org/
  12  *
  13  * Author: Isaac Huang <isaac@clusterfs.com>
  14  *
  15  * 2012-05-13: Liang Zhen <liang@whamcloud.com>
  16  * - percpt data for service to improve smp performance
  17  * - code cleanup
  18  */
  19
  20 #define DEBUG_SUBSYSTEM S_LNET
  21
  22 #include "selftest.h"
  23
  24 enum srpc_state {
  25         SRPC_STATE_NONE,
  26         SRPC_STATE_NI_INIT,
  27         SRPC_STATE_EQ_INIT,
  28         SRPC_STATE_RUNNING,
  29         SRPC_STATE_STOPPING,
  30 };
  31
  32 enum rpc_counter_32 {
  33         SRPC_ERROR,
  34         SRPC_RPC_SENT,
  35         SRPC_RPC_RCVD,
  36         SRPC_RPC_DROP,
  37         SRPC_RPC_EXPIRED,
  38         SRPC_COUNTER32_MAX,
  39 };
  40
  41 enum rpc_counter_64 {
  42         SRPC_BULK_GET,
  43         SRPC_BULK_PUT,
  44         SRPC_COUNTER64_MAX,
  45 };
  46
  47 static struct smoketest_rpc {
  48         spinlock_t       rpc_glock;     /* global lock */
  49         struct srpc_service     *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  50         lnet_handler_t           rpc_lnet_handler;/* _the_ LNet event handler */
  51         enum srpc_state          rpc_state;
  52         atomic_t                 rpc_counters32[SRPC_COUNTER32_MAX];
  53         atomic64_t               rpc_counters64[SRPC_COUNTER64_MAX];
  54         atomic64_t               rpc_matchbits; /* matchbits counter */
  55
  56 } srpc_data;
  57
  58 #define RPC_STAT32(a) \
  59         srpc_data.rpc_counters32[(a)]
  60
  61 #define GET_RPC_STAT32(a) \
  62         atomic_read(&srpc_data.rpc_counters32[(a)])
  63
  64 #define GET_RPC_STAT64(a) \
  65         atomic64_read(&srpc_data.rpc_counters64[(a)])
  66
  67 static inline int
  68 srpc_serv_portal(int svc_id)
  69 {
  70         return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
  71                SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
  72 }
  73
  74 /* forward ref's */
  75 static void srpc_handle_rpc(struct swi_workitem *wi);
  76
  77
  78 void srpc_get_counters(struct srpc_counters *cnt)
  79 {
  80         cnt->errors = GET_RPC_STAT32(SRPC_ERROR);
  81         cnt->rpcs_sent = GET_RPC_STAT32(SRPC_RPC_SENT);
  82         cnt->rpcs_rcvd = GET_RPC_STAT32(SRPC_RPC_RCVD);
  83         cnt->rpcs_dropped = GET_RPC_STAT32(SRPC_RPC_DROP);
  84         cnt->rpcs_expired = GET_RPC_STAT32(SRPC_RPC_EXPIRED);
  85
  86         cnt->bulk_get = GET_RPC_STAT64(SRPC_BULK_GET);
  87         cnt->bulk_put = GET_RPC_STAT64(SRPC_BULK_PUT);
  88 }
  89
  90 static int
  91 srpc_init_bulk_page(struct srpc_bulk *bk, int i, int off, int nob)
  92 {
  93         LASSERT(off < PAGE_SIZE);
  94         LASSERT(nob > 0 && nob <= PAGE_SIZE);
  95
  96         bk->bk_iovs[i].bv_offset = off;
  97         bk->bk_iovs[i].bv_len    = nob;
  98         return nob;
  99 }
 100
 101 void
 102 srpc_free_bulk(struct srpc_bulk *bk)
 103 {
 104         int i;
 105         struct page *pg;
 106
 107         LASSERT(bk != NULL);
 108
 109         for (i = 0; i < bk->bk_alloc; i++) {
 110                 pg = bk->bk_iovs[i].bv_page;
 111                 if (pg == NULL)
 112                         break;
 113
 114                 __free_page(pg);
 115         }
 116
 117         LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_alloc]));
 118 }
 119
 120 struct srpc_bulk *
 121 srpc_alloc_bulk(int cpt, unsigned int bulk_len)
 122 {
 123         struct srpc_bulk *bk;
 124         int i;
 125         int bulk_npg = (bulk_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 126
 127         LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
 128
 129         LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
 130                          offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 131         if (bk == NULL) {
 132                 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
 133                 return NULL;
 134         }
 135
 136         memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
 137         bk->bk_alloc   = bulk_npg;
 138         LASSERTF(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV, "b: %u\n", bulk_npg);
 139
 140         for (i = 0; i < bulk_npg; i++) {
 141                 struct page *pg;
 142
 143                 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
 144                 if (pg == NULL) {
 145                         CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
 146                         srpc_free_bulk(bk);
 147                         return NULL;
 148                 }
 149                 bk->bk_iovs[i].bv_page   = pg;
 150         }
 151
 152         return bk;
 153 }
 154
 155 void
 156 srpc_init_bulk(struct srpc_bulk *bk, unsigned int bulk_off,
 157                 unsigned int bulk_len, int sink)
 158 {
 159         int i;
 160         ENTRY;
 161
 162         CDEBUG(D_INFO, "bulk %p o %u l %u s %u\n",
 163                 bk, bulk_off, bulk_len, sink);
 164
 165         LASSERT(bk != NULL);
 166
 167         bk->bk_sink   = sink;
 168         bk->bk_len    = bulk_len;
 169
 170         for (i = 0; bulk_len > 0; i++) {
 171                 int nob;
 172
 173                 LASSERT(bk->bk_iovs[i].bv_page != NULL);
 174
 175                 nob = min_t(unsigned, bulk_off + bulk_len, PAGE_SIZE) -
 176                       bulk_off;
 177
 178                 srpc_init_bulk_page(bk, i, bulk_off, nob);
 179                 bulk_len -= nob;
 180                 bulk_off = 0;
 181         }
 182         bk->bk_niov = i;
 183         LASSERTF(bk->bk_niov >= 0 && bk->bk_niov <= bk->bk_alloc,
 184                 "bk %p - n: %u/%u\n", bk, bk->bk_niov, bk->bk_alloc);
 185 }
 186
 187 static inline __u64
 188 srpc_next_id (void)
 189 {
 190         return atomic64_inc_return(&srpc_data.rpc_matchbits);
 191 }
 192
 193 static void
 194 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
 195                      struct srpc_service_cd *scd,
 196                      struct srpc_buffer *buffer)
 197 {
 198         swi_init_workitem(&rpc->srpc_wi, srpc_handle_rpc,
 199                           srpc_serv_is_framework(scd->scd_svc) ?
 200                           lst_serial_wq : lst_test_wq[scd->scd_cpt]);
 201
 202         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 203
 204         rpc->srpc_scd      = scd;
 205         rpc->srpc_reqstbuf = buffer;
 206         rpc->srpc_peer     = buffer->buf_peer;
 207         rpc->srpc_self     = buffer->buf_self;
 208         LNetInvalidateMDHandle(&rpc->srpc_replymdh);
 209
 210         rpc->srpc_aborted  = 0;
 211         rpc->srpc_status   = 0;
 212 }
 213
 214 static void
 215 srpc_service_fini(struct srpc_service *svc)
 216 {
 217         struct srpc_service_cd *scd;
 218         struct srpc_server_rpc *rpc;
 219         struct srpc_buffer *buf;
 220         struct list_head *q;
 221         int i;
 222
 223         if (svc->sv_cpt_data == NULL)
 224                 return;
 225
 226         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 227                 while (1) {
 228                         if (!list_empty(&scd->scd_buf_posted))
 229                                 q = &scd->scd_buf_posted;
 230                         else if (!list_empty(&scd->scd_buf_blocked))
 231                                 q = &scd->scd_buf_blocked;
 232                         else
 233                                 break;
 234
 235                         while (!list_empty(q)) {
 236                                 buf = list_first_entry(q,
 237                                                        struct srpc_buffer,
 238                                                        buf_list);
 239                                 list_del(&buf->buf_list);
 240                                 LIBCFS_FREE(buf, sizeof(*buf));
 241                         }
 242                 }
 243
 244                 LASSERT(list_empty(&scd->scd_rpc_active));
 245
 246                 while (!list_empty(&scd->scd_rpc_free)) {
 247                         rpc = list_first_entry(&scd->scd_rpc_free,
 248                                                struct srpc_server_rpc,
 249                                                srpc_list);
 250                         list_del(&rpc->srpc_list);
 251                         if (svc->sv_srpc_fini)
 252                                 svc->sv_srpc_fini(rpc);
 253                         LIBCFS_FREE(rpc, sizeof(*rpc));
 254                 }
 255         }
 256
 257         cfs_percpt_free(svc->sv_cpt_data);
 258         svc->sv_cpt_data = NULL;
 259 }
 260
 261 static int
 262 srpc_service_nrpcs(struct srpc_service *svc)
 263 {
 264         int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
 265
 266         return srpc_serv_is_framework(svc) ?
 267                max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
 268 }
 269
 270 void srpc_add_buffer(struct swi_workitem *wi);
 271
 272 static int
 273 srpc_service_init(struct srpc_service *svc)
 274 {
 275         struct srpc_service_cd *scd;
 276         struct srpc_server_rpc *rpc;
 277         int nrpcs;
 278         int i;
 279         int j;
 280
 281         svc->sv_shuttingdown = 0;
 282
 283         svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
 284                                             sizeof(struct srpc_service_cd));
 285         if (svc->sv_cpt_data == NULL)
 286                 return -ENOMEM;
 287
 288         svc->sv_ncpts = srpc_serv_is_framework(svc) ?
 289                         1 : cfs_cpt_number(lnet_cpt_table());
 290         nrpcs = srpc_service_nrpcs(svc);
 291
 292         cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
 293                 scd->scd_cpt = i;
 294                 scd->scd_svc = svc;
 295                 spin_lock_init(&scd->scd_lock);
 296                 INIT_LIST_HEAD(&scd->scd_rpc_free);
 297                 INIT_LIST_HEAD(&scd->scd_rpc_active);
 298                 INIT_LIST_HEAD(&scd->scd_buf_posted);
 299                 INIT_LIST_HEAD(&scd->scd_buf_blocked);
 300
 301                 scd->scd_ev.ev_data = scd;
 302                 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
 303
 304                 /* NB: don't use lst_serial_wq for adding buffer,
 305                  * see details in srpc_service_add_buffers() */
 306                 swi_init_workitem(&scd->scd_buf_wi,
 307                                   srpc_add_buffer, lst_test_wq[i]);
 308
 309                 if (i != 0 && srpc_serv_is_framework(svc)) {
 310                         /* NB: framework service only needs srpc_service_cd for
 311                          * one partition, but we allocate for all to make
 312                          * it easier to implement, it will waste a little
 313                          * memory but nobody should care about this */
 314                         continue;
 315                 }
 316
 317                 for (j = 0; j < nrpcs; j++) {
 318                         LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
 319                                          i, sizeof(*rpc));
 320                         if (rpc == NULL ||
 321                            (svc->sv_srpc_init && svc->sv_srpc_init(rpc, i))) {
 322                                 srpc_service_fini(svc);
 323                                 return -ENOMEM;
 324                         }
 325                         list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 326                 }
 327         }
 328
 329         return 0;
 330 }
 331
 332 int
 333 srpc_add_service(struct srpc_service *sv)
 334 {
 335         int id = sv->sv_id;
 336
 337         LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
 338
 339         if (srpc_service_init(sv) != 0)
 340                 return -ENOMEM;
 341
 342         spin_lock(&srpc_data.rpc_glock);
 343
 344         LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
 345
 346         if (srpc_data.rpc_services[id] != NULL) {
 347                 spin_unlock(&srpc_data.rpc_glock);
 348                 goto failed;
 349         }
 350
 351         srpc_data.rpc_services[id] = sv;
 352         spin_unlock(&srpc_data.rpc_glock);
 353
 354         CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
 355         return 0;
 356
 357 failed:
 358         srpc_service_fini(sv);
 359         return -EBUSY;
 360 }
 361
 362 int
 363 srpc_remove_service(struct srpc_service *sv)
 364 {
 365         int id = sv->sv_id;
 366
 367         spin_lock(&srpc_data.rpc_glock);
 368
 369         if (srpc_data.rpc_services[id] != sv) {
 370                 spin_unlock(&srpc_data.rpc_glock);
 371                 return -ENOENT;
 372         }
 373
 374         srpc_data.rpc_services[id] = NULL;
 375         spin_unlock(&srpc_data.rpc_glock);
 376         return 0;
 377 }
 378
 379 static int
 380 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
 381                        int len, int options, struct lnet_process_id peer4,
 382                        struct lnet_handle_md *mdh, struct srpc_event *ev)
 383 {
 384         int rc;
 385         struct lnet_md md;
 386         struct lnet_me *me;
 387         struct lnet_processid peer;
 388
 389         peer.pid = peer4.pid;
 390         lnet_nid4_to_nid(peer4.nid, &peer.nid);
 391
 392         me = LNetMEAttach(portal, &peer, matchbits, 0, LNET_UNLINK,
 393                           local ? LNET_INS_LOCAL : LNET_INS_AFTER);
 394         if (IS_ERR(me)) {
 395                 rc = PTR_ERR(me);
 396                 CERROR("LNetMEAttach failed: %d\n", rc);
 397                 LASSERT(rc == -ENOMEM);
 398                 return -ENOMEM;
 399         }
 400
 401         md.threshold = 1;
 402         md.user_ptr  = ev;
 403         md.start     = buf;
 404         md.length    = len;
 405         md.options   = options;
 406         md.handler   = srpc_data.rpc_lnet_handler;
 407
 408         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdh);
 409         if (rc != 0) {
 410                 CERROR("LNetMDAttach failed: %d\n", rc);
 411                 LASSERT(rc == -ENOMEM);
 412
 413                 return -ENOMEM;
 414         }
 415
 416         CDEBUG(D_NET,
 417                "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
 418                libcfs_id2str(peer4), portal, matchbits);
 419         return 0;
 420 }
 421
 422 static int
 423 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 424                       int options, struct lnet_process_id peer4,
 425                       lnet_nid_t self4, struct lnet_handle_md *mdh,
 426                       struct srpc_event *ev)
 427 {
 428         int rc;
 429         struct lnet_md md;
 430         struct lnet_nid self;
 431         struct lnet_processid peer;
 432
 433         lnet_nid4_to_nid(self4, &self);
 434         lnet_pid4_to_pid(peer4, &peer);
 435
 436         md.user_ptr  = ev;
 437         md.start     = buf;
 438         md.length    = len;
 439         md.handler   = srpc_data.rpc_lnet_handler;
 440         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 441         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 442
 443         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
 444         if (rc != 0) {
 445                 CERROR("LNetMDBind failed: %d\n", rc);
 446                 LASSERT(rc == -ENOMEM);
 447                 return -ENOMEM;
 448         }
 449
 450         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 451          * they're only meaningful for MDs attached to an ME (i.e. passive
 452          * buffers...
 453          */
 454         if ((options & LNET_MD_OP_PUT) != 0) {
 455                 rc = LNetPut(&self, *mdh, LNET_NOACK_REQ, &peer,
 456                              portal, matchbits, 0, 0);
 457         } else {
 458                 LASSERT((options & LNET_MD_OP_GET) != 0);
 459
 460                 rc = LNetGet(&self, *mdh, &peer, portal, matchbits, 0, false);
 461         }
 462
 463         if (rc != 0) {
 464                 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
 465                        ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 466                        libcfs_id2str(peer4), portal, matchbits, rc);
 467
 468                 /* The forthcoming unlink event will complete this operation
 469                  * with failure, so fall through and return success here.
 470                  */
 471                 rc = LNetMDUnlink(*mdh);
 472                 LASSERT(rc == 0);
 473         } else {
 474                 CDEBUG(D_NET,
 475                        "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
 476                        libcfs_id2str(peer4), portal, matchbits);
 477         }
 478         return 0;
 479 }
 480
 481 static int
 482 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
 483                          struct lnet_handle_md *mdh, struct srpc_event *ev)
 484 {
 485         struct lnet_process_id any = {0};
 486
 487         any.nid = LNET_NID_ANY;
 488         any.pid = LNET_PID_ANY;
 489
 490         return srpc_post_passive_rdma(srpc_serv_portal(service),
 491                                       local, service, buf, len,
 492                                       LNET_MD_OP_PUT, any, mdh, ev);
 493 }
 494
 495 static int
 496 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
 497 __must_hold(&scd->scd_lock)
 498 {
 499         struct srpc_service *sv = scd->scd_svc;
 500         struct srpc_msg *msg = &buf->buf_msg;
 501         int rc;
 502
 503         LNetInvalidateMDHandle(&buf->buf_mdh);
 504         list_add(&buf->buf_list, &scd->scd_buf_posted);
 505         scd->scd_buf_nposted++;
 506         spin_unlock(&scd->scd_lock);
 507
 508         rc = srpc_post_passive_rqtbuf(sv->sv_id,
 509                                       !srpc_serv_is_framework(sv),
 510                                       msg, sizeof(*msg), &buf->buf_mdh,
 511                                       &scd->scd_ev);
 512
 513         /* At this point, a RPC (new or delayed) may have arrived in
 514          * msg and its event handler has been called. So we must add
 515          * buf to scd_buf_posted _before_ dropping scd_lock */
 516
 517         spin_lock(&scd->scd_lock);
 518
 519         if (rc == 0) {
 520                 if (!sv->sv_shuttingdown)
 521                         return 0;
 522
 523                 spin_unlock(&scd->scd_lock);
 524                 /* srpc_shutdown_service might have tried to unlink me
 525                  * when my buf_mdh was still invalid */
 526                 LNetMDUnlink(buf->buf_mdh);
 527                 spin_lock(&scd->scd_lock);
 528                 return 0;
 529         }
 530
 531         scd->scd_buf_nposted--;
 532         if (sv->sv_shuttingdown)
 533                 return rc; /* don't allow to change scd_buf_posted */
 534
 535         list_del(&buf->buf_list);
 536         spin_unlock(&scd->scd_lock);
 537
 538         LIBCFS_FREE(buf, sizeof(*buf));
 539
 540         spin_lock(&scd->scd_lock);
 541         return rc;
 542 }
 543
 544 void
 545 srpc_add_buffer(struct swi_workitem *wi)
 546 {
 547         struct srpc_service_cd *scd = container_of(wi, struct srpc_service_cd,
 548                                                    scd_buf_wi);
 549         struct srpc_buffer *buf;
 550         int rc = 0;
 551
 552         /* it's called by workitem scheduler threads, these threads
 553          * should have been set CPT affinity, so buffers will be posted
 554          * on CPT local list of Portal */
 555         spin_lock(&scd->scd_lock);
 556
 557         while (scd->scd_buf_adjust > 0 &&
 558                !scd->scd_svc->sv_shuttingdown) {
 559                 scd->scd_buf_adjust--; /* consume it */
 560                 scd->scd_buf_posting++;
 561
 562                 spin_unlock(&scd->scd_lock);
 563
 564                 LIBCFS_ALLOC(buf, sizeof(*buf));
 565                 if (buf == NULL) {
 566                         CERROR("Failed to add new buf to service: %s\n",
 567                                scd->scd_svc->sv_name);
 568                         spin_lock(&scd->scd_lock);
 569                         rc = -ENOMEM;
 570                         break;
 571                 }
 572
 573                 spin_lock(&scd->scd_lock);
 574                 if (scd->scd_svc->sv_shuttingdown) {
 575                         spin_unlock(&scd->scd_lock);
 576                         LIBCFS_FREE(buf, sizeof(*buf));
 577
 578                         spin_lock(&scd->scd_lock);
 579                         rc = -ESHUTDOWN;
 580                         break;
 581                 }
 582
 583                 rc = srpc_service_post_buffer(scd, buf);
 584                 if (rc != 0)
 585                         break; /* buf has been freed inside */
 586
 587                 LASSERT(scd->scd_buf_posting > 0);
 588                 scd->scd_buf_posting--;
 589                 scd->scd_buf_total++;
 590                 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
 591         }
 592
 593         if (rc != 0) {
 594                 scd->scd_buf_err_stamp = ktime_get_real_seconds();
 595                 scd->scd_buf_err = rc;
 596
 597                 LASSERT(scd->scd_buf_posting > 0);
 598                 scd->scd_buf_posting--;
 599         }
 600
 601         wi->swi_state = SWI_STATE_RUNNING;
 602         spin_unlock(&scd->scd_lock);
 603 }
 604
 605 int
 606 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
 607 {
 608         struct srpc_service_cd *scd;
 609         int rc = 0;
 610         int i;
 611
 612         LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
 613
 614         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 615                 spin_lock(&scd->scd_lock);
 616
 617                 scd->scd_buf_err = 0;
 618                 scd->scd_buf_err_stamp = 0;
 619                 scd->scd_buf_posting = 0;
 620                 scd->scd_buf_adjust = nbuffer;
 621                 /* start to post buffers */
 622                 swi_schedule_workitem(&scd->scd_buf_wi);
 623                 spin_unlock(&scd->scd_lock);
 624
 625                 /* framework service only post buffer for one partition  */
 626                 if (srpc_serv_is_framework(sv))
 627                         break;
 628         }
 629
 630         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 631                 spin_lock(&scd->scd_lock);
 632                 /*
 633                  * NB: srpc_service_add_buffers() can be called inside
 634                  * thread context of lst_serial_wq, and we don't normally
 635                  * allow to sleep inside thread context of WI scheduler
 636                  * because it will block current scheduler thread from doing
 637                  * anything else, even worse, it could deadlock if it's
 638                  * waiting on result from another WI of the same scheduler.
 639                  * However, it's safe at here because scd_buf_wi is scheduled
 640                  * by thread in a different WI scheduler (lst_test_wq),
 641                  * so we don't have any risk of deadlock, though this could
 642                  * block all WIs pending on lst_serial_wq for a moment
 643                  * which is not good but not fatal.
 644                  */
 645                 lst_wait_until(scd->scd_buf_err != 0 ||
 646                                (scd->scd_buf_adjust == 0 &&
 647                                 scd->scd_buf_posting == 0),
 648                                scd->scd_lock, "waiting for adding buffer\n");
 649
 650                 if (scd->scd_buf_err != 0 && rc == 0)
 651                         rc = scd->scd_buf_err;
 652
 653                 spin_unlock(&scd->scd_lock);
 654         }
 655
 656         return rc;
 657 }
 658
 659 void
 660 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
 661 {
 662         struct srpc_service_cd *scd;
 663         int num;
 664         int i;
 665
 666         LASSERT(!sv->sv_shuttingdown);
 667
 668         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 669                 spin_lock(&scd->scd_lock);
 670
 671                 num = scd->scd_buf_total + scd->scd_buf_posting;
 672                 scd->scd_buf_adjust -= min(nbuffer, num);
 673
 674                 spin_unlock(&scd->scd_lock);
 675         }
 676 }
 677
 678 /* returns 1 if sv has finished, otherwise 0 */
 679 int
 680 srpc_finish_service(struct srpc_service *sv)
 681 {
 682         struct srpc_service_cd *scd;
 683         struct srpc_server_rpc *rpc;
 684         int i;
 685
 686         LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
 687
 688         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 689                 swi_cancel_workitem(&scd->scd_buf_wi);
 690
 691                 spin_lock(&scd->scd_lock);
 692                 if (scd->scd_buf_nposted > 0) {
 693                         CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
 694                                scd->scd_buf_nposted);
 695                         spin_unlock(&scd->scd_lock);
 696                         return 0;
 697                 }
 698
 699                 if (list_empty(&scd->scd_rpc_active)) {
 700                         spin_unlock(&scd->scd_lock);
 701                         continue;
 702                 }
 703
 704                 rpc = list_first_entry(&scd->scd_rpc_active,
 705                                        struct srpc_server_rpc, srpc_list);
 706                 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, wi %s, ev fired %d type %d status %d lnet %d\n",
 707                         rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 708                         swi_state2str(rpc->srpc_wi.swi_state),
 709                         rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
 710                         rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
 711                 spin_unlock(&scd->scd_lock);
 712                 return 0;
 713         }
 714
 715         /* no lock needed from now on */
 716         srpc_service_fini(sv);
 717         return 1;
 718 }
 719
 720 /* called with sv->sv_lock held */
 721 static void
 722 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
 723                             struct srpc_buffer *buf)
 724 __must_hold(&scd->scd_lock)
 725 {
 726         if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
 727                 if (srpc_service_post_buffer(scd, buf) != 0) {
 728                         CWARN("Failed to post %s buffer\n",
 729                               scd->scd_svc->sv_name);
 730                 }
 731                 return;
 732         }
 733
 734         /* service is shutting down, or we want to recycle some buffers */
 735         scd->scd_buf_total--;
 736
 737         if (scd->scd_buf_adjust < 0) {
 738                 scd->scd_buf_adjust++;
 739                 if (scd->scd_buf_adjust < 0 &&
 740                     scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
 741                         CDEBUG(D_INFO,
 742                                "Try to recyle %d buffers but nothing left\n",
 743                                scd->scd_buf_adjust);
 744                         scd->scd_buf_adjust = 0;
 745                 }
 746         }
 747
 748         spin_unlock(&scd->scd_lock);
 749         LIBCFS_FREE(buf, sizeof(*buf));
 750         spin_lock(&scd->scd_lock);
 751 }
 752
 753 void
 754 srpc_abort_service(struct srpc_service *sv)
 755 {
 756         struct srpc_service_cd *scd;
 757         struct srpc_server_rpc *rpc;
 758         int i;
 759
 760         CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
 761                sv->sv_id, sv->sv_name);
 762
 763         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 764                 spin_lock(&scd->scd_lock);
 765
 766                 /* schedule in-flight RPCs to notice the abort, NB:
 767                  * racing with incoming RPCs; complete fix should make test
 768                  * RPCs carry session ID in its headers
 769                  */
 770                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
 771                         rpc->srpc_aborted = 1;
 772                         swi_schedule_workitem(&rpc->srpc_wi);
 773                 }
 774
 775                 spin_unlock(&scd->scd_lock);
 776         }
 777 }
 778
 779 void
 780 srpc_shutdown_service(struct srpc_service *sv)
 781 {
 782         struct srpc_service_cd *scd;
 783         struct srpc_server_rpc *rpc;
 784         struct srpc_buffer *buf;
 785         int i;
 786
 787         CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
 788                sv->sv_id, sv->sv_name);
 789
 790         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 791                 spin_lock(&scd->scd_lock);
 792
 793         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 794
 795         cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
 796                 spin_unlock(&scd->scd_lock);
 797
 798         cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
 799                 spin_lock(&scd->scd_lock);
 800
 801                 /* schedule in-flight RPCs to notice the shutdown */
 802                 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
 803                         swi_schedule_workitem(&rpc->srpc_wi);
 804
 805                 spin_unlock(&scd->scd_lock);
 806
 807                 /* OK to traverse scd_buf_posted without lock, since no one
 808                  * touches scd_buf_posted now
 809                  */
 810                 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
 811                         LNetMDUnlink(buf->buf_mdh);
 812         }
 813 }
 814
 815 static int
 816 srpc_send_request(struct srpc_client_rpc *rpc)
 817 {
 818         struct srpc_event *ev = &rpc->crpc_reqstev;
 819         int rc;
 820
 821         ev->ev_fired = 0;
 822         ev->ev_data  = rpc;
 823         ev->ev_type  = SRPC_REQUEST_SENT;
 824
 825         rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
 826                                    rpc->crpc_service, &rpc->crpc_reqstmsg,
 827                                    sizeof(struct srpc_msg), LNET_MD_OP_PUT,
 828                                    rpc->crpc_dest, LNET_NID_ANY,
 829                                    &rpc->crpc_reqstmdh, ev);
 830         if (rc != 0) {
 831                 LASSERT(rc == -ENOMEM);
 832                 ev->ev_fired = 1;  /* no more event expected */
 833         }
 834         return rc;
 835 }
 836
 837 static int
 838 srpc_prepare_reply(struct srpc_client_rpc *rpc)
 839 {
 840         struct srpc_event *ev = &rpc->crpc_replyev;
 841         u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 842         int rc;
 843
 844         ev->ev_fired = 0;
 845         ev->ev_data  = rpc;
 846         ev->ev_type  = SRPC_REPLY_RCVD;
 847
 848         *id = srpc_next_id();
 849
 850         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 851                                     &rpc->crpc_replymsg,
 852                                     sizeof(struct srpc_msg),
 853                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 854                                     &rpc->crpc_replymdh, ev);
 855         if (rc != 0) {
 856                 LASSERT(rc == -ENOMEM);
 857                 ev->ev_fired = 1;  /* no more event expected */
 858         }
 859         return rc;
 860 }
 861
 862 static int
 863 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
 864 {
 865         struct srpc_bulk *bk = &rpc->crpc_bulk;
 866         struct srpc_event *ev = &rpc->crpc_bulkev;
 867         __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 868         int rc;
 869         int opt;
 870
 871         LASSERT(bk->bk_niov <= LNET_MAX_IOV);
 872
 873         /* nothing to do */
 874         if (bk->bk_niov == 0)
 875                 return 0;
 876
 877         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 878         opt |= LNET_MD_KIOV;
 879
 880         ev->ev_fired = 0;
 881         ev->ev_data  = rpc;
 882         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 883
 884         *id = srpc_next_id();
 885
 886         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
 887                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 888                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 889         if (rc != 0) {
 890                 LASSERT(rc == -ENOMEM);
 891                 ev->ev_fired = 1;  /* no more event expected */
 892         }
 893         return rc;
 894 }
 895
 896 static int
 897 srpc_do_bulk(struct srpc_server_rpc *rpc)
 898 {
 899         struct srpc_event *ev = &rpc->srpc_ev;
 900         struct srpc_bulk *bk = rpc->srpc_bulk;
 901         __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 902         int rc;
 903         int opt;
 904
 905         LASSERT(bk != NULL);
 906
 907         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 908         opt |= LNET_MD_KIOV;
 909
 910         ev->ev_fired = 0;
 911         ev->ev_data  = rpc;
 912         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 913
 914         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 915                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 916                                    rpc->srpc_peer, rpc->srpc_self,
 917                                    &bk->bk_mdh, ev);
 918         if (rc != 0)
 919                 ev->ev_fired = 1;  /* no more event expected */
 920         return rc;
 921 }
 922
 923 /* only called from srpc_handle_rpc */
 924 static void
 925 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
 926 {
 927         struct srpc_service_cd  *scd = rpc->srpc_scd;
 928         struct srpc_service     *sv  = scd->scd_svc;
 929         struct srpc_buffer *buffer;
 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
 944         spin_lock(&scd->scd_lock);
 945
 946         if (rpc->srpc_reqstbuf != NULL) {
 947                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 948                  * sv won't go away for scd_rpc_active must not be empty
 949                  */
 950                 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
 951                 rpc->srpc_reqstbuf = NULL;
 952         }
 953
 954         list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
 955
 956         /*
 957          * No one can schedule me now since:
 958          * - I'm not on scd_rpc_active.
 959          * - all LNet events have been fired.
 960          * Cancel pending schedules and prevent future schedule attempts:
 961          */
 962         LASSERT(rpc->srpc_ev.ev_fired);
 963         rpc->srpc_wi.swi_state = SWI_STATE_DONE;
 964
 965         if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
 966                 buffer = list_first_entry(&scd->scd_buf_blocked,
 967                                           struct srpc_buffer, buf_list);
 968                 list_del(&buffer->buf_list);
 969
 970                 srpc_init_server_rpc(rpc, scd, buffer);
 971                 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
 972                 swi_schedule_workitem(&rpc->srpc_wi);
 973         } else {
 974                 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
 975         }
 976
 977         spin_unlock(&scd->scd_lock);
 978 }
 979
 980 /* handles an incoming RPC */
 981 static void srpc_handle_rpc(struct swi_workitem *wi)
 982 {
 983         struct srpc_server_rpc *rpc = container_of(wi, struct srpc_server_rpc,
 984                                                    srpc_wi);
 985         struct srpc_service_cd *scd = rpc->srpc_scd;
 986         struct srpc_service *sv = scd->scd_svc;
 987         struct srpc_event *ev = &rpc->srpc_ev;
 988         int rc = 0;
 989
 990         spin_lock(&scd->scd_lock);
 991         if (wi->swi_state == SWI_STATE_DONE) {
 992                 spin_unlock(&scd->scd_lock);
 993                 return;
 994         }
 995
 996         if (sv->sv_shuttingdown || rpc->srpc_aborted) {
 997                 wi->swi_state = SWI_STATE_DONE;
 998                 spin_unlock(&scd->scd_lock);
 999
1000                 if (rpc->srpc_bulk != NULL)
1001                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
1002                 LNetMDUnlink(rpc->srpc_replymdh);
1003
1004                 if (ev->ev_fired) /* no more event, OK to finish */
1005                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
1006                 return;
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;
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;
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; /* 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;
1069                         }
1070                 }
1071
1072                 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1073                 rc = srpc_send_reply(rpc);
1074                 if (rc == 0)
1075                         return; /* wait for reply */
1076                 srpc_server_rpc_done(rpc, rc);
1077                 return;
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;
1091         }
1092 }
1093
1094 static void
1095 srpc_client_rpc_expired (void *data)
1096 {
1097         struct srpc_client_rpc *rpc = data;
1098
1099         CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1100               rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1101               rpc->crpc_timeout);
1102
1103         LBUG();
1104         spin_lock(&rpc->crpc_lock);
1105
1106         rpc->crpc_timeout = 0;
1107         srpc_abort_rpc(rpc, -ETIMEDOUT);
1108
1109         spin_unlock(&rpc->crpc_lock);
1110
1111         atomic_inc(&RPC_STAT32(SRPC_RPC_EXPIRED));
1112 }
1113
1114 static void
1115 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1116 {
1117         struct stt_timer *timer = &rpc->crpc_timer;
1118
1119         if (rpc->crpc_timeout == 0)
1120                 return;
1121
1122         INIT_LIST_HEAD(&timer->stt_list);
1123         timer->stt_data    = rpc;
1124         timer->stt_func    = srpc_client_rpc_expired;
1125         timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1126         stt_add_timer(timer);
1127 }
1128
1129 /*
1130  * Called with rpc->crpc_lock held.
1131  *
1132  * Upon exit the RPC expiry timer is not queued and the handler is not
1133  * running on any CPU.
1134  */
1135 static void
1136 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1137 {
1138         /* timer not planted or already exploded */
1139         if (rpc->crpc_timeout == 0)
1140                 return;
1141
1142         /* timer successfully defused */
1143         if (stt_del_timer(&rpc->crpc_timer))
1144                 return;
1145
1146         /* timer detonated, wait for it to explode */
1147         while (rpc->crpc_timeout != 0) {
1148                 spin_unlock(&rpc->crpc_lock);
1149
1150                 schedule();
1151
1152                 spin_lock(&rpc->crpc_lock);
1153         }
1154 }
1155
1156 static void
1157 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1158 {
1159         struct swi_workitem *wi = &rpc->crpc_wi;
1160
1161         spin_lock(&rpc->crpc_lock);
1162
1163         rpc->crpc_closed = 1;
1164         if (rpc->crpc_status == 0)
1165                 rpc->crpc_status = status;
1166
1167         srpc_del_client_rpc_timer(rpc);
1168
1169         CDEBUG_LIMIT((status == 0) ? D_NET : D_NETERROR,
1170                      "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1171                      rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1172                      swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1173
1174         /*
1175          * No one can schedule me now since:
1176          * - RPC timer has been defused.
1177          * - all LNet events have been fired.
1178          * - crpc_closed has been set, preventing srpc_abort_rpc from
1179          *   scheduling me.
1180          * Cancel pending schedules and prevent future schedule attempts:
1181          */
1182         LASSERT(!srpc_event_pending(rpc));
1183         wi->swi_state = SWI_STATE_DONE;
1184
1185         spin_unlock(&rpc->crpc_lock);
1186
1187         (*rpc->crpc_done)(rpc);
1188 }
1189
1190 /* sends an outgoing RPC */
1191 void
1192 srpc_send_rpc(struct swi_workitem *wi)
1193 {
1194         int rc = 0;
1195         struct srpc_client_rpc *rpc;
1196         struct srpc_msg *reply;
1197         int do_bulk;
1198
1199         LASSERT(wi != NULL);
1200
1201         rpc = container_of(wi, struct srpc_client_rpc, crpc_wi);
1202
1203         LASSERT(rpc != NULL);
1204         LASSERT(wi == &rpc->crpc_wi);
1205
1206         reply = &rpc->crpc_replymsg;
1207         do_bulk = rpc->crpc_bulk.bk_niov > 0;
1208
1209         spin_lock(&rpc->crpc_lock);
1210         if (wi->swi_state == SWI_STATE_DONE) {
1211                 spin_unlock(&rpc->crpc_lock);
1212                 return;
1213         }
1214
1215         if (rpc->crpc_aborted) {
1216                 spin_unlock(&rpc->crpc_lock);
1217                 goto abort;
1218         }
1219
1220         spin_unlock(&rpc->crpc_lock);
1221
1222         switch (wi->swi_state) {
1223         default:
1224                 LBUG();
1225         case SWI_STATE_NEWBORN:
1226                 LASSERT(!srpc_event_pending(rpc));
1227
1228                 rc = srpc_prepare_reply(rpc);
1229                 if (rc != 0) {
1230                         srpc_client_rpc_done(rpc, rc);
1231                         return;
1232                 }
1233
1234                 rc = srpc_prepare_bulk(rpc);
1235                 if (rc != 0)
1236                         break;
1237
1238                 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1239                 rc = srpc_send_request(rpc);
1240                 break;
1241
1242         case SWI_STATE_REQUEST_SUBMITTED:
1243                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1244                  * order; however, they're processed in a strict order:
1245                  * rqt, rpy, and bulk.
1246                  */
1247                 if (!rpc->crpc_reqstev.ev_fired)
1248                         break;
1249
1250                 rc = rpc->crpc_reqstev.ev_status;
1251                 if (rc != 0)
1252                         break;
1253
1254                 wi->swi_state = SWI_STATE_REQUEST_SENT;
1255                 fallthrough;
1256         case SWI_STATE_REQUEST_SENT: {
1257                 enum srpc_msg_type type;
1258
1259                 type = srpc_service2reply(rpc->crpc_service);
1260
1261                 if (!rpc->crpc_replyev.ev_fired)
1262                         break;
1263
1264                 rc = rpc->crpc_replyev.ev_status;
1265                 if (rc != 0)
1266                         break;
1267
1268                 srpc_unpack_msg_hdr(reply);
1269                 if (reply->msg_type != type ||
1270                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1271                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1272                         CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1273                               libcfs_id2str(rpc->crpc_dest),
1274                               reply->msg_type, type,
1275                               reply->msg_magic, SRPC_MSG_MAGIC);
1276                         rc = -EBADMSG;
1277                         break;
1278                 }
1279
1280                 if (do_bulk && reply->msg_body.reply.status != 0) {
1281                         CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1282                               reply->msg_body.reply.status,
1283                               libcfs_id2str(rpc->crpc_dest));
1284                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1285                 }
1286
1287                 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1288         }
1289         fallthrough;
1290         case SWI_STATE_REPLY_RECEIVED:
1291                 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1292                         break;
1293
1294                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1295
1296                 /* Bulk buffer was unlinked due to remote error. Clear error
1297                  * since reply buffer still contains valid data.
1298                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1299                  * remote error.
1300                  */
1301                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1302                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1303                         rc = 0;
1304
1305                 wi->swi_state = SWI_STATE_DONE;
1306                 srpc_client_rpc_done(rpc, rc);
1307                 return;
1308         }
1309
1310         if (rc != 0) {
1311                 spin_lock(&rpc->crpc_lock);
1312                 srpc_abort_rpc(rpc, rc);
1313                 spin_unlock(&rpc->crpc_lock);
1314         }
1315
1316 abort:
1317         if (rpc->crpc_aborted) {
1318                 LNetMDUnlink(rpc->crpc_reqstmdh);
1319                 LNetMDUnlink(rpc->crpc_replymdh);
1320                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1321
1322                 if (!srpc_event_pending(rpc)) {
1323                         srpc_client_rpc_done(rpc, -EINTR);
1324                         return;
1325                 }
1326         }
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 %px, crpc %px, reqstev %px, replyev %px, bulkev %px\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_first_entry(&scd->scd_rpc_free,
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, true);
1697                 fallthrough;
1698
1699         case SRPC_STATE_NI_INIT:
1700                 LNetNIFini();
1701         }
1702 }