lnet/selftest/rpc.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/selftest/rpc.c
  37  *
  38  * Author: Isaac Huang <isaac@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include "selftest.h"
  44
  45
  46 #define SRPC_PEER_HASH_SIZE       101  /* # peer lists */
  47
  48 typedef enum {
  49         SRPC_STATE_NONE,
  50         SRPC_STATE_NI_INIT,
  51         SRPC_STATE_EQ_INIT,
  52         SRPC_STATE_WI_INIT,
  53         SRPC_STATE_RUNNING,
  54         SRPC_STATE_STOPPING,
  55 } srpc_state_t;
  56
  57 struct smoketest_rpc {
  58         spinlock_t        rpc_glock;     /* global lock */
  59         srpc_service_t   *rpc_services[SRPC_SERVICE_MAX_ID + 1];
  60         struct list_head *rpc_peers;     /* hash table of known peers */
  61         lnet_handle_eq_t  rpc_lnet_eq;   /* _the_ LNet event queue */
  62         srpc_state_t      rpc_state;
  63         srpc_counters_t   rpc_counters;
  64         __u64             rpc_matchbits; /* matchbits counter */
  65 } srpc_data;
  66
  67 static int srpc_peer_credits = 16;
  68 CFS_MODULE_PARM(srpc_peer_credits, "i", int, 0444,
  69                 "# in-flight RPCs per peer (16 by default)");
  70
  71 /* forward ref's */
  72 int srpc_handle_rpc (swi_workitem_t *wi);
  73
  74 void srpc_get_counters (srpc_counters_t *cnt)
  75 {
  76         spin_lock(&srpc_data.rpc_glock);
  77         *cnt = srpc_data.rpc_counters;
  78         spin_unlock(&srpc_data.rpc_glock);
  79 }
  80
  81 void srpc_set_counters (const srpc_counters_t *cnt)
  82 {
  83         spin_lock(&srpc_data.rpc_glock);
  84         srpc_data.rpc_counters = *cnt;
  85         spin_unlock(&srpc_data.rpc_glock);
  86 }
  87
  88 void
  89 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
  90 {
  91         LASSERT (i >= 0 && i < bk->bk_niov);
  92
  93 #ifdef __KERNEL__
  94         bk->bk_iovs[i].kiov_offset = 0;
  95         bk->bk_iovs[i].kiov_page   = pg;
  96         bk->bk_iovs[i].kiov_len    = CFS_PAGE_SIZE;
  97 #else
  98         LASSERT (bk->bk_pages != NULL);
  99
 100         bk->bk_pages[i] = pg;
 101         bk->bk_iovs[i].iov_len  = CFS_PAGE_SIZE;
 102         bk->bk_iovs[i].iov_base = cfs_page_address(pg);
 103 #endif
 104         return;
 105 }
 106
 107 void
 108 srpc_free_bulk (srpc_bulk_t *bk)
 109 {
 110         int         i;
 111         cfs_page_t *pg;
 112
 113         LASSERT (bk != NULL);
 114 #ifndef __KERNEL__
 115         LASSERT (bk->bk_pages != NULL);
 116 #endif
 117
 118         for (i = 0; i < bk->bk_niov; i++) {
 119 #ifdef __KERNEL__
 120                 pg = bk->bk_iovs[i].kiov_page;
 121 #else
 122                 pg = bk->bk_pages[i];
 123 #endif
 124                 if (pg == NULL) break;
 125
 126                 cfs_free_page(pg);
 127         }
 128
 129 #ifndef __KERNEL__
 130         LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
 131 #endif
 132         LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
 133         return;
 134 }
 135
 136 srpc_bulk_t *
 137 srpc_alloc_bulk (int npages, int sink)
 138 {
 139         srpc_bulk_t  *bk;
 140         cfs_page_t  **pages;
 141         int           i;
 142
 143         LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
 144
 145         LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 146         if (bk == NULL) {
 147                 CERROR ("Can't allocate descriptor for %d pages\n", npages);
 148                 return NULL;
 149         }
 150
 151         memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
 152         bk->bk_sink = sink;
 153         bk->bk_niov = npages;
 154         bk->bk_len  = npages * CFS_PAGE_SIZE;
 155 #ifndef __KERNEL__
 156         LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
 157         if (pages == NULL) {
 158                 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
 159                 CERROR ("Can't allocate page array for %d pages\n", npages);
 160                 return NULL;
 161         }
 162
 163         memset(pages, 0, sizeof(cfs_page_t *) * npages);
 164         bk->bk_pages = pages;
 165 #else
 166         UNUSED (pages);
 167 #endif
 168
 169         for (i = 0; i < npages; i++) {
 170                 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
 171
 172                 if (pg == NULL) {
 173                         CERROR ("Can't allocate page %d of %d\n", i, npages);
 174                         srpc_free_bulk(bk);
 175                         return NULL;
 176                 }
 177
 178                 srpc_add_bulk_page(bk, pg, i);
 179         }
 180
 181         return bk;
 182 }
 183
 184
 185 static inline struct list_head *
 186 srpc_nid2peerlist (lnet_nid_t nid)
 187 {
 188         unsigned int hash = ((unsigned int)nid) % SRPC_PEER_HASH_SIZE;
 189
 190         return &srpc_data.rpc_peers[hash];
 191 }
 192
 193 static inline srpc_peer_t *
 194 srpc_create_peer (lnet_nid_t nid)
 195 {
 196         srpc_peer_t *peer;
 197
 198         LASSERT (nid != LNET_NID_ANY);
 199
 200         LIBCFS_ALLOC(peer, sizeof(srpc_peer_t));
 201         if (peer == NULL) {
 202                 CERROR ("Failed to allocate peer structure for %s\n",
 203                         libcfs_nid2str(nid));
 204                 return NULL;
 205         }
 206
 207         memset(peer, 0, sizeof(srpc_peer_t));
 208         peer->stp_nid     = nid;
 209         peer->stp_credits = srpc_peer_credits;
 210
 211         spin_lock_init(&peer->stp_lock);
 212         CFS_INIT_LIST_HEAD(&peer->stp_rpcq);
 213         CFS_INIT_LIST_HEAD(&peer->stp_ctl_rpcq);
 214         return peer;
 215 }
 216
 217 srpc_peer_t *
 218 srpc_find_peer_locked (lnet_nid_t nid)
 219 {
 220         struct list_head *peer_list = srpc_nid2peerlist(nid);
 221         srpc_peer_t      *peer;
 222
 223         LASSERT (nid != LNET_NID_ANY);
 224
 225         list_for_each_entry (peer, peer_list, stp_list) {
 226                 if (peer->stp_nid == nid)
 227                         return peer;
 228         }
 229
 230         return NULL;
 231 }
 232
 233 static srpc_peer_t *
 234 srpc_nid2peer (lnet_nid_t nid)
 235 {
 236         srpc_peer_t *peer;
 237         srpc_peer_t *new_peer;
 238
 239         spin_lock(&srpc_data.rpc_glock);
 240         peer = srpc_find_peer_locked(nid);
 241         spin_unlock(&srpc_data.rpc_glock);
 242
 243         if (peer != NULL)
 244                 return peer;
 245
 246         new_peer = srpc_create_peer(nid);
 247
 248         spin_lock(&srpc_data.rpc_glock);
 249
 250         peer = srpc_find_peer_locked(nid);
 251         if (peer != NULL) {
 252                 spin_unlock(&srpc_data.rpc_glock);
 253                 if (new_peer != NULL)
 254                         LIBCFS_FREE(new_peer, sizeof(srpc_peer_t));
 255
 256                 return peer;
 257         }
 258
 259         if (new_peer == NULL) {
 260                 spin_unlock(&srpc_data.rpc_glock);
 261                 return NULL;
 262         }
 263
 264         list_add_tail(&new_peer->stp_list, srpc_nid2peerlist(nid));
 265         spin_unlock(&srpc_data.rpc_glock);
 266         return new_peer;
 267 }
 268
 269 static inline __u64
 270 srpc_next_id (void)
 271 {
 272         __u64 id;
 273
 274         spin_lock(&srpc_data.rpc_glock);
 275         id = srpc_data.rpc_matchbits++;
 276         spin_unlock(&srpc_data.rpc_glock);
 277         return id;
 278 }
 279
 280 void
 281 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
 282                       srpc_service_t *sv, srpc_buffer_t *buffer)
 283 {
 284         memset(rpc, 0, sizeof(*rpc));
 285         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc);
 286
 287         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 288
 289         rpc->srpc_service  = sv;
 290         rpc->srpc_reqstbuf = buffer;
 291         rpc->srpc_peer     = buffer->buf_peer;
 292         rpc->srpc_self     = buffer->buf_self;
 293         rpc->srpc_replymdh = LNET_INVALID_HANDLE;
 294 }
 295
 296 int
 297 srpc_add_service (srpc_service_t *sv)
 298 {
 299         int                id = sv->sv_id;
 300         int                i;
 301         srpc_server_rpc_t *rpc;
 302
 303         LASSERT (sv->sv_concur > 0);
 304         LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
 305
 306         spin_lock(&srpc_data.rpc_glock);
 307
 308         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
 309
 310         if (srpc_data.rpc_services[id] != NULL) {
 311                 spin_unlock(&srpc_data.rpc_glock);
 312                 return -EBUSY;
 313         }
 314
 315         srpc_data.rpc_services[id] = sv;
 316         spin_unlock(&srpc_data.rpc_glock);
 317
 318         sv->sv_nprune       = 0;
 319         sv->sv_nposted_msg  = 0;
 320         sv->sv_shuttingdown = 0;
 321         spin_lock_init(&sv->sv_lock);
 322         CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
 323         CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
 324         CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
 325         CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
 326
 327         sv->sv_ev.ev_data = sv;
 328         sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
 329
 330         for (i = 0; i < sv->sv_concur; i++) {
 331                 LIBCFS_ALLOC(rpc, sizeof(*rpc));
 332                 if (rpc == NULL) goto enomem;
 333
 334                 list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 335         }
 336
 337         CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
 338                 id, sv->sv_name, sv->sv_concur);
 339         return 0;
 340
 341 enomem:
 342         while (!list_empty(&sv->sv_free_rpcq)) {
 343                 rpc = list_entry(sv->sv_free_rpcq.next,
 344                                  srpc_server_rpc_t, srpc_list);
 345                 list_del(&rpc->srpc_list);
 346                 LIBCFS_FREE(rpc, sizeof(*rpc));
 347         }
 348
 349         spin_lock(&srpc_data.rpc_glock);
 350         srpc_data.rpc_services[id] = NULL;
 351         spin_unlock(&srpc_data.rpc_glock);
 352         return -ENOMEM;
 353 }
 354
 355 int
 356 srpc_remove_service (srpc_service_t *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 int
 373 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
 374                        int len, int options, lnet_process_id_t peer,
 375                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 376 {
 377         int              rc;
 378         lnet_md_t        md;
 379         lnet_handle_me_t meh;
 380
 381         rc = LNetMEAttach(portal, peer, matchbits, 0,
 382                           LNET_UNLINK, LNET_INS_AFTER, &meh);
 383         if (rc != 0) {
 384                 CERROR ("LNetMEAttach failed: %d\n", rc);
 385                 LASSERT (rc == -ENOMEM);
 386                 return -ENOMEM;
 387         }
 388
 389         md.threshold = 1;
 390         md.user_ptr  = ev;
 391         md.start     = buf;
 392         md.length    = len;
 393         md.options   = options;
 394         md.eq_handle = srpc_data.rpc_lnet_eq;
 395
 396         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 397         if (rc != 0) {
 398                 CERROR ("LNetMDAttach failed: %d\n", rc);
 399                 LASSERT (rc == -ENOMEM);
 400
 401                 rc = LNetMEUnlink(meh);
 402                 LASSERT (rc == 0);
 403                 return -ENOMEM;
 404         }
 405
 406         CDEBUG (D_NET,
 407                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 408                 libcfs_id2str(peer), portal, matchbits);
 409         return 0;
 410 }
 411
 412 int
 413 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 414                       int options, lnet_process_id_t peer, lnet_nid_t self,
 415                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 416 {
 417         int       rc;
 418         lnet_md_t md;
 419
 420         md.user_ptr  = ev;
 421         md.start     = buf;
 422         md.length    = len;
 423         md.eq_handle = srpc_data.rpc_lnet_eq;
 424         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 425         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 426
 427         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 428         if (rc != 0) {
 429                 CERROR ("LNetMDBind failed: %d\n", rc);
 430                 LASSERT (rc == -ENOMEM);
 431                 return -ENOMEM;
 432         }
 433
 434         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 435          * they're only meaningful for MDs attached to an ME (i.e. passive
 436          * buffers... */
 437         if ((options & LNET_MD_OP_PUT) != 0) {
 438                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 439                              portal, matchbits, 0, 0);
 440         } else {
 441                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 442
 443                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 444         }
 445
 446         if (rc != 0) {
 447                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 448                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 449                         libcfs_id2str(peer), portal, matchbits, rc);
 450
 451                 /* The forthcoming unlink event will complete this operation
 452                  * with failure, so fall through and return success here.
 453                  */
 454                 rc = LNetMDUnlink(*mdh);
 455                 LASSERT (rc == 0);
 456         } else {
 457                 CDEBUG (D_NET,
 458                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 459                         libcfs_id2str(peer), portal, matchbits);
 460         }
 461         return 0;
 462 }
 463
 464 int
 465 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 466                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 467 {
 468         int rc;
 469         int portal;
 470
 471         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 472                 portal = SRPC_REQUEST_PORTAL;
 473         else
 474                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 475
 476         rc = srpc_post_active_rdma(portal, service, buf, len,
 477                                    LNET_MD_OP_PUT, peer,
 478                                    LNET_NID_ANY, mdh, ev);
 479         return rc;
 480 }
 481
 482 int
 483 srpc_post_passive_rqtbuf(int service, void *buf, int len,
 484                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 485 {
 486         int               rc;
 487         int               portal;
 488         lnet_process_id_t any = {.nid = LNET_NID_ANY,
 489                                  .pid = LNET_PID_ANY};
 490
 491         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 492                 portal = SRPC_REQUEST_PORTAL;
 493         else
 494                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 495
 496         rc = srpc_post_passive_rdma(portal, service, buf, len,
 497                                     LNET_MD_OP_PUT, any, mdh, ev);
 498         return rc;
 499 }
 500
 501 int
 502 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 503 {
 504         srpc_msg_t *msg = &buf->buf_msg;
 505         int         rc;
 506
 507         LASSERT (!sv->sv_shuttingdown);
 508
 509         buf->buf_mdh = LNET_INVALID_HANDLE;
 510         list_add(&buf->buf_list, &sv->sv_posted_msgq);
 511         sv->sv_nposted_msg++;
 512         spin_unlock(&sv->sv_lock);
 513
 514         rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
 515                                       &buf->buf_mdh, &sv->sv_ev);
 516
 517         /* At this point, a RPC (new or delayed) may have arrived in
 518          * msg and its event handler has been called. So we must add
 519          * buf to sv_posted_msgq _before_ dropping sv_lock */
 520
 521         spin_lock(&sv->sv_lock);
 522
 523         if (rc == 0) {
 524                 if (sv->sv_shuttingdown) {
 525                         spin_unlock(&sv->sv_lock);
 526
 527                         /* srpc_shutdown_service might have tried to unlink me
 528                          * when my buf_mdh was still invalid */
 529                         LNetMDUnlink(buf->buf_mdh);
 530
 531                         spin_lock(&sv->sv_lock);
 532                 }
 533                 return 0;
 534         }
 535
 536         sv->sv_nposted_msg--;
 537         if (sv->sv_shuttingdown) return rc;
 538
 539         list_del(&buf->buf_list);
 540
 541         spin_unlock(&sv->sv_lock);
 542         LIBCFS_FREE(buf, sizeof(*buf));
 543         spin_lock(&sv->sv_lock);
 544         return rc;
 545 }
 546
 547 int
 548 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
 549 {
 550         int                rc;
 551         int                posted;
 552         srpc_buffer_t     *buf;
 553
 554         LASSERTF (nbuffer > 0,
 555                   "nbuffer must be positive: %d\n", nbuffer);
 556
 557         for (posted = 0; posted < nbuffer; posted++) {
 558                 LIBCFS_ALLOC(buf, sizeof(*buf));
 559                 if (buf == NULL) break;
 560
 561                 spin_lock(&sv->sv_lock);
 562                 rc = srpc_service_post_buffer(sv, buf);
 563                 spin_unlock(&sv->sv_lock);
 564
 565                 if (rc != 0) break;
 566         }
 567
 568         return posted;
 569 }
 570
 571 void
 572 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
 573 {
 574         LASSERTF (nbuffer > 0,
 575                   "nbuffer must be positive: %d\n", nbuffer);
 576
 577         spin_lock(&sv->sv_lock);
 578
 579         LASSERT (sv->sv_nprune >= 0);
 580         LASSERT (!sv->sv_shuttingdown);
 581
 582         sv->sv_nprune += nbuffer;
 583
 584         spin_unlock(&sv->sv_lock);
 585         return;
 586 }
 587
 588 /* returns 1 if sv has finished, otherwise 0 */
 589 int
 590 srpc_finish_service (srpc_service_t *sv)
 591 {
 592         srpc_server_rpc_t *rpc;
 593         srpc_buffer_t     *buf;
 594
 595         spin_lock(&sv->sv_lock);
 596
 597         LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
 598
 599         if (sv->sv_nposted_msg != 0 || !list_empty(&sv->sv_active_rpcq)) {
 600                 CDEBUG (D_NET,
 601                         "waiting for %d posted buffers to unlink and "
 602                         "in-flight RPCs to die.\n",
 603                         sv->sv_nposted_msg);
 604
 605                 if (!list_empty(&sv->sv_active_rpcq)) {
 606                         rpc = list_entry(sv->sv_active_rpcq.next,
 607                                          srpc_server_rpc_t, srpc_list);
 608                         CDEBUG (D_NETERROR,
 609                                 "Active RPC on shutdown: sv %s, peer %s, "
 610                                 "wi %s scheduled %d running %d, "
 611                                 "ev fired %d type %d status %d lnet %d\n",
 612                                 sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 613                                 swi_state2str(rpc->srpc_wi.wi_state),
 614                                 rpc->srpc_wi.wi_scheduled,
 615                                 rpc->srpc_wi.wi_running,
 616                                 rpc->srpc_ev.ev_fired,
 617                                 rpc->srpc_ev.ev_type,
 618                                 rpc->srpc_ev.ev_status,
 619                                 rpc->srpc_ev.ev_lnet);
 620                 }
 621
 622                 spin_unlock(&sv->sv_lock);
 623                 return 0;
 624         }
 625
 626         spin_unlock(&sv->sv_lock); /* no lock needed from now on */
 627
 628         for (;;) {
 629                 struct list_head *q;
 630
 631                 if (!list_empty(&sv->sv_posted_msgq))
 632                         q = &sv->sv_posted_msgq;
 633                 else if (!list_empty(&sv->sv_blocked_msgq))
 634                         q = &sv->sv_blocked_msgq;
 635                 else
 636                         break;
 637
 638                 buf = list_entry(q->next, srpc_buffer_t, buf_list);
 639                 list_del(&buf->buf_list);
 640
 641                 LIBCFS_FREE(buf, sizeof(*buf));
 642         }
 643
 644         while (!list_empty(&sv->sv_free_rpcq)) {
 645                 rpc = list_entry(sv->sv_free_rpcq.next,
 646                                  srpc_server_rpc_t, srpc_list);
 647                 list_del(&rpc->srpc_list);
 648                 LIBCFS_FREE(rpc, sizeof(*rpc));
 649         }
 650
 651         return 1;
 652 }
 653
 654 /* called with sv->sv_lock held */
 655 void
 656 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 657 {
 658         if (sv->sv_shuttingdown) goto free;
 659
 660         if (sv->sv_nprune == 0) {
 661                 if (srpc_service_post_buffer(sv, buf) != 0)
 662                         CWARN ("Failed to post %s buffer\n", sv->sv_name);
 663                 return;
 664         }
 665
 666         sv->sv_nprune--;
 667 free:
 668         spin_unlock(&sv->sv_lock);
 669         LIBCFS_FREE(buf, sizeof(*buf));
 670         spin_lock(&sv->sv_lock);
 671 }
 672
 673 void
 674 srpc_shutdown_service (srpc_service_t *sv)
 675 {
 676         srpc_server_rpc_t *rpc;
 677         srpc_buffer_t     *buf;
 678
 679         spin_lock(&sv->sv_lock);
 680
 681         CDEBUG (D_NET, "Shutting down service: id %d, name %s\n",
 682                 sv->sv_id, sv->sv_name);
 683
 684         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 685
 686         /* schedule in-flight RPCs to notice the shutdown */
 687         list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
 688                 swi_schedule_workitem(&rpc->srpc_wi);
 689         }
 690
 691         spin_unlock(&sv->sv_lock);
 692
 693         /* OK to traverse sv_posted_msgq without lock, since no one
 694          * touches sv_posted_msgq now */
 695         list_for_each_entry (buf, &sv->sv_posted_msgq, buf_list)
 696                 LNetMDUnlink(buf->buf_mdh);
 697
 698         return;
 699 }
 700
 701 int
 702 srpc_send_request (srpc_client_rpc_t *rpc)
 703 {
 704         srpc_event_t *ev = &rpc->crpc_reqstev;
 705         int           rc;
 706
 707         ev->ev_fired = 0;
 708         ev->ev_data  = rpc;
 709         ev->ev_type  = SRPC_REQUEST_SENT;
 710
 711         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 712                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 713                                      &rpc->crpc_reqstmdh, ev);
 714         if (rc != 0) {
 715                 LASSERT (rc == -ENOMEM);
 716                 ev->ev_fired = 1;  /* no more event expected */
 717         }
 718         return rc;
 719 }
 720
 721 int
 722 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 723 {
 724         srpc_event_t *ev = &rpc->crpc_replyev;
 725         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 726         int           rc;
 727
 728         ev->ev_fired = 0;
 729         ev->ev_data  = rpc;
 730         ev->ev_type  = SRPC_REPLY_RCVD;
 731
 732         *id = srpc_next_id();
 733
 734         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 735                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 736                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 737                                     &rpc->crpc_replymdh, ev);
 738         if (rc != 0) {
 739                 LASSERT (rc == -ENOMEM);
 740                 ev->ev_fired = 1;  /* no more event expected */
 741         }
 742         return rc;
 743 }
 744
 745 int
 746 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 747 {
 748         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 749         srpc_event_t *ev = &rpc->crpc_bulkev;
 750         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 751         int           rc;
 752         int           opt;
 753
 754         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 755
 756         if (bk->bk_niov == 0) return 0; /* nothing to do */
 757
 758         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 759 #ifdef __KERNEL__
 760         opt |= LNET_MD_KIOV;
 761 #else
 762         opt |= LNET_MD_IOVEC;
 763 #endif
 764
 765         ev->ev_fired = 0;
 766         ev->ev_data  = rpc;
 767         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 768
 769         *id = srpc_next_id();
 770
 771         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 772                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 773                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 774         if (rc != 0) {
 775                 LASSERT (rc == -ENOMEM);
 776                 ev->ev_fired = 1;  /* no more event expected */
 777         }
 778         return rc;
 779 }
 780
 781 int
 782 srpc_do_bulk (srpc_server_rpc_t *rpc)
 783 {
 784         srpc_event_t  *ev = &rpc->srpc_ev;
 785         srpc_bulk_t   *bk = rpc->srpc_bulk;
 786         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 787         int            rc;
 788         int            opt;
 789
 790         LASSERT (bk != NULL);
 791
 792         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 793 #ifdef __KERNEL__
 794         opt |= LNET_MD_KIOV;
 795 #else
 796         opt |= LNET_MD_IOVEC;
 797 #endif
 798
 799         ev->ev_fired = 0;
 800         ev->ev_data  = rpc;
 801         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 802
 803         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 804                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 805                                    rpc->srpc_peer, rpc->srpc_self,
 806                                    &bk->bk_mdh, ev);
 807         if (rc != 0)
 808                 ev->ev_fired = 1;  /* no more event expected */
 809         return rc;
 810 }
 811
 812 /* called with srpc_service_t::sv_lock held */
 813 inline void
 814 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
 815 {
 816         srpc_service_t *sv = rpc->srpc_service;
 817
 818         if (sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 819                 swi_schedule_workitem(&rpc->srpc_wi);
 820         else    /* framework RPCs are handled one by one */
 821                 swi_schedule_serial_workitem(&rpc->srpc_wi);
 822
 823         return;
 824 }
 825
 826 /* only called from srpc_handle_rpc */
 827 void
 828 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
 829 {
 830         srpc_service_t *sv = rpc->srpc_service;
 831         srpc_buffer_t  *buffer;
 832
 833         LASSERT (status != 0 || rpc->srpc_wi.wi_state == SWI_STATE_DONE);
 834
 835         rpc->srpc_status = status;
 836
 837         CDEBUG (status == 0 ? D_NET : D_NETERROR,
 838                 "Server RPC done: service %s, peer %s, status %s:%d\n",
 839                 sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 840                 swi_state2str(rpc->srpc_wi.wi_state), status);
 841
 842         if (status != 0) {
 843                 spin_lock(&srpc_data.rpc_glock);
 844                 srpc_data.rpc_counters.rpcs_dropped++;
 845                 spin_unlock(&srpc_data.rpc_glock);
 846         }
 847
 848         if (rpc->srpc_done != NULL)
 849                 (*rpc->srpc_done) (rpc);
 850         LASSERT (rpc->srpc_bulk == NULL);
 851
 852         spin_lock(&sv->sv_lock);
 853
 854         if (rpc->srpc_reqstbuf != NULL) {
 855                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 856                  * sv won't go away for sv_active_rpcq must not be empty */
 857                 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
 858                 rpc->srpc_reqstbuf = NULL;
 859         }
 860
 861         list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
 862
 863         /*
 864          * No one can schedule me now since:
 865          * - I'm not on sv_active_rpcq.
 866          * - all LNet events have been fired.
 867          * Cancel pending schedules and prevent future schedule attempts:
 868          */
 869         LASSERT (rpc->srpc_ev.ev_fired);
 870         swi_kill_workitem(&rpc->srpc_wi);
 871
 872         if (!sv->sv_shuttingdown && !list_empty(&sv->sv_blocked_msgq)) {
 873                 buffer = list_entry(sv->sv_blocked_msgq.next,
 874                                     srpc_buffer_t, buf_list);
 875                 list_del(&buffer->buf_list);
 876
 877                 srpc_init_server_rpc(rpc, sv, buffer);
 878                 list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
 879                 srpc_schedule_server_rpc(rpc);
 880         } else {
 881                 list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 882         }
 883
 884         spin_unlock(&sv->sv_lock);
 885         return;
 886 }
 887
 888 /* handles an incoming RPC */
 889 int
 890 srpc_handle_rpc (swi_workitem_t *wi)
 891 {
 892         srpc_server_rpc_t *rpc = wi->wi_data;
 893         srpc_service_t    *sv = rpc->srpc_service;
 894         srpc_event_t      *ev = &rpc->srpc_ev;
 895         int                rc = 0;
 896
 897         LASSERT (wi == &rpc->srpc_wi);
 898
 899         spin_lock(&sv->sv_lock);
 900
 901         if (sv->sv_shuttingdown) {
 902                 spin_unlock(&sv->sv_lock);
 903
 904                 if (rpc->srpc_bulk != NULL)
 905                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 906                 LNetMDUnlink(rpc->srpc_replymdh);
 907
 908                 if (ev->ev_fired) { /* no more event, OK to finish */
 909                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 910                         return 1;
 911                 }
 912                 return 0;
 913         }
 914
 915         spin_unlock(&sv->sv_lock);
 916
 917         switch (wi->wi_state) {
 918         default:
 919                 LBUG ();
 920         case SWI_STATE_NEWBORN: {
 921                 srpc_msg_t           *msg;
 922                 srpc_generic_reply_t *reply;
 923
 924                 msg = &rpc->srpc_reqstbuf->buf_msg;
 925                 reply = &rpc->srpc_replymsg.msg_body.reply;
 926
 927                 if (msg->msg_magic == 0) {
 928                         /* moaned already in srpc_lnet_ev_handler */
 929                         rc = EBADMSG;
 930                 } else if (msg->msg_version != SRPC_MSG_VERSION &&
 931                            msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
 932                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
 933                                msg->msg_version, SRPC_MSG_VERSION,
 934                                libcfs_id2str(rpc->srpc_peer));
 935                         reply->status = EPROTO;
 936                 } else {
 937                         reply->status = 0;
 938                         rc = (*sv->sv_handler) (rpc);
 939                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
 940                 }
 941
 942                 if (rc != 0) {
 943                         srpc_server_rpc_done(rpc, rc);
 944                         return 1;
 945                 }
 946
 947                 wi->wi_state = SWI_STATE_BULK_STARTED;
 948
 949                 if (rpc->srpc_bulk != NULL) {
 950                         rc = srpc_do_bulk(rpc);
 951                         if (rc == 0)
 952                                 return 0; /* wait for bulk */
 953
 954                         LASSERT (ev->ev_fired);
 955                         ev->ev_status = rc;
 956                 }
 957         }
 958         case SWI_STATE_BULK_STARTED:
 959                 /* we cannot LASSERT ev_fired right here because it
 960                  * may be set only upon an event with unlinked==1 */
 961
 962                 if (rpc->srpc_bulk != NULL) {
 963                         rc = ev->ev_status;
 964
 965                         if (sv->sv_bulk_ready != NULL)
 966                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
 967
 968                         if (rc != 0) {
 969                                 if (ev->ev_fired) {
 970                                         srpc_server_rpc_done(rpc, rc);
 971                                         return 1;
 972                                 }
 973
 974                                 rpc->srpc_status = rc;
 975                                 wi->wi_state     = SWI_STATE_BULK_ERRORED;
 976                                 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 977                                 return 0; /* wait for UNLINK event  */
 978                         }
 979                 }
 980
 981                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
 982
 983                 wi->wi_state = SWI_STATE_REPLY_SUBMITTED;
 984                 rc = srpc_send_reply(rpc);
 985                 if (rc == 0)
 986                         return 0; /* wait for reply */
 987                 srpc_server_rpc_done(rpc, rc);
 988                 return 1;
 989
 990         case SWI_STATE_REPLY_SUBMITTED:
 991                 LASSERT (ev->ev_fired);
 992
 993                 wi->wi_state = SWI_STATE_DONE;
 994                 srpc_server_rpc_done(rpc, ev->ev_status);
 995                 return 1;
 996
 997         case SWI_STATE_BULK_ERRORED:
 998                 LASSERT (rpc->srpc_bulk != NULL && ev->ev_fired);
 999                 LASSERT (rpc->srpc_status != 0);
1000
1001                 srpc_server_rpc_done(rpc, rpc->srpc_status);
1002                 return 1;
1003         }
1004
1005         return 0;
1006 }
1007
1008 void
1009 srpc_client_rpc_expired (void *data)
1010 {
1011         srpc_client_rpc_t *rpc = data;
1012
1013         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1014                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1015                rpc->crpc_timeout);
1016
1017         spin_lock(&rpc->crpc_lock);
1018
1019         rpc->crpc_timeout = 0;
1020         srpc_abort_rpc(rpc, -ETIMEDOUT);
1021
1022         spin_unlock(&rpc->crpc_lock);
1023
1024         spin_lock(&srpc_data.rpc_glock);
1025         srpc_data.rpc_counters.rpcs_expired++;
1026         spin_unlock(&srpc_data.rpc_glock);
1027         return;
1028 }
1029
1030 inline void
1031 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
1032 {
1033         stt_timer_t *timer = &rpc->crpc_timer;
1034
1035         if (rpc->crpc_timeout == 0) return;
1036
1037         CFS_INIT_LIST_HEAD(&timer->stt_list);
1038         timer->stt_data    = rpc;
1039         timer->stt_func    = srpc_client_rpc_expired;
1040         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1041                                           cfs_time_current_sec());
1042         stt_add_timer(timer);
1043         return;
1044 }
1045
1046 /*
1047  * Called with rpc->crpc_lock held.
1048  *
1049  * Upon exit the RPC expiry timer is not queued and the handler is not
1050  * running on any CPU. */
1051 void
1052 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1053 {
1054         /* timer not planted or already exploded */
1055         if (rpc->crpc_timeout == 0) return;
1056
1057         /* timer sucessfully defused */
1058         if (stt_del_timer(&rpc->crpc_timer)) return;
1059
1060 #ifdef __KERNEL__
1061         /* timer detonated, wait for it to explode */
1062         while (rpc->crpc_timeout != 0) {
1063                 spin_unlock(&rpc->crpc_lock);
1064
1065                 cfs_schedule();
1066
1067                 spin_lock(&rpc->crpc_lock);
1068         }
1069 #else
1070         LBUG(); /* impossible in single-threaded runtime */
1071 #endif
1072         return;
1073 }
1074
1075 void
1076 srpc_check_sends (srpc_peer_t *peer, int credits)
1077 {
1078         struct list_head  *q;
1079         srpc_client_rpc_t *rpc;
1080
1081         LASSERT (credits >= 0);
1082         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1083
1084         spin_lock(&peer->stp_lock);
1085         peer->stp_credits += credits;
1086
1087         while (peer->stp_credits) {
1088                 if (!list_empty(&peer->stp_ctl_rpcq))
1089                         q = &peer->stp_ctl_rpcq;
1090                 else if (!list_empty(&peer->stp_rpcq))
1091                         q = &peer->stp_rpcq;
1092                 else
1093                         break;
1094
1095                 peer->stp_credits--;
1096
1097                 rpc = list_entry(q->next, srpc_client_rpc_t, crpc_privl);
1098                 list_del_init(&rpc->crpc_privl);
1099                 srpc_client_rpc_decref(rpc);  /* --ref for peer->*rpcq */
1100
1101                 swi_schedule_workitem(&rpc->crpc_wi);
1102         }
1103
1104         spin_unlock(&peer->stp_lock);
1105         return;
1106 }
1107
1108 void
1109 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1110 {
1111         swi_workitem_t *wi = &rpc->crpc_wi;
1112         srpc_peer_t    *peer = rpc->crpc_peer;
1113
1114         LASSERT (status != 0 || wi->wi_state == SWI_STATE_DONE);
1115
1116         spin_lock(&rpc->crpc_lock);
1117
1118         rpc->crpc_closed = 1;
1119         if (rpc->crpc_status == 0)
1120                 rpc->crpc_status = status;
1121
1122         srpc_del_client_rpc_timer(rpc);
1123
1124         CDEBUG ((status == 0) ? D_NET : D_NETERROR,
1125                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1126                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1127                 swi_state2str(wi->wi_state), rpc->crpc_aborted, status);
1128
1129         /*
1130          * No one can schedule me now since:
1131          * - RPC timer has been defused.
1132          * - all LNet events have been fired.
1133          * - crpc_closed has been set, preventing srpc_abort_rpc from
1134          *   scheduling me.
1135          * Cancel pending schedules and prevent future schedule attempts:
1136          */
1137         LASSERT (!srpc_event_pending(rpc));
1138         swi_kill_workitem(wi);
1139
1140         spin_unlock(&rpc->crpc_lock);
1141
1142         (*rpc->crpc_done) (rpc);
1143
1144         if (peer != NULL)
1145                 srpc_check_sends(peer, 1);
1146         return;
1147 }
1148
1149 /* sends an outgoing RPC */
1150 int
1151 srpc_send_rpc (swi_workitem_t *wi)
1152 {
1153         int                rc = 0;
1154         srpc_client_rpc_t *rpc = wi->wi_data;
1155         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1156         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1157
1158         LASSERT (rpc != NULL);
1159         LASSERT (wi == &rpc->crpc_wi);
1160
1161         spin_lock(&rpc->crpc_lock);
1162
1163         if (rpc->crpc_aborted) {
1164                 spin_unlock(&rpc->crpc_lock);
1165                 goto abort;
1166         }
1167
1168         spin_unlock(&rpc->crpc_lock);
1169
1170         switch (wi->wi_state) {
1171         default:
1172                 LBUG ();
1173         case SWI_STATE_NEWBORN:
1174                 LASSERT (!srpc_event_pending(rpc));
1175
1176                 rc = srpc_prepare_reply(rpc);
1177                 if (rc != 0) {
1178                         srpc_client_rpc_done(rpc, rc);
1179                         return 1;
1180                 }
1181
1182                 rc = srpc_prepare_bulk(rpc);
1183                 if (rc != 0) break;
1184
1185                 wi->wi_state = SWI_STATE_REQUEST_SUBMITTED;
1186                 rc = srpc_send_request(rpc);
1187                 break;
1188
1189         case SWI_STATE_REQUEST_SUBMITTED:
1190                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1191                  * order; however, they're processed in a strict order:
1192                  * rqt, rpy, and bulk. */
1193                 if (!rpc->crpc_reqstev.ev_fired) break;
1194
1195                 rc = rpc->crpc_reqstev.ev_status;
1196                 if (rc != 0) break;
1197
1198                 wi->wi_state = SWI_STATE_REQUEST_SENT;
1199                 /* perhaps more events, fall thru */
1200         case SWI_STATE_REQUEST_SENT: {
1201                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1202
1203                 if (!rpc->crpc_replyev.ev_fired) break;
1204
1205                 rc = rpc->crpc_replyev.ev_status;
1206                 if (rc != 0) break;
1207
1208                 if ((reply->msg_type != type &&
1209                      reply->msg_type != __swab32(type)) ||
1210                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1211                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1212                         CWARN ("Bad message from %s: type %u (%d expected),"
1213                                " magic %u (%d expected).\n",
1214                                libcfs_id2str(rpc->crpc_dest),
1215                                reply->msg_type, type,
1216                                reply->msg_magic, SRPC_MSG_MAGIC);
1217                         rc = -EBADMSG;
1218                         break;
1219                 }
1220
1221                 if (do_bulk && reply->msg_body.reply.status != 0) {
1222                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1223                                "case peer didn't initiate bulk transfer\n",
1224                                reply->msg_body.reply.status,
1225                                libcfs_id2str(rpc->crpc_dest));
1226                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1227                 }
1228
1229                 wi->wi_state = SWI_STATE_REPLY_RECEIVED;
1230         }
1231         case SWI_STATE_REPLY_RECEIVED:
1232                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1233
1234                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1235
1236                 /* Bulk buffer was unlinked due to remote error. Clear error
1237                  * since reply buffer still contains valid data.
1238                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1239                  * remote error. */
1240                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1241                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1242                         rc = 0;
1243
1244                 wi->wi_state = SWI_STATE_DONE;
1245                 srpc_client_rpc_done(rpc, rc);
1246                 return 1;
1247         }
1248
1249         if (rc != 0) {
1250                 spin_lock(&rpc->crpc_lock);
1251                 srpc_abort_rpc(rpc, rc);
1252                 spin_unlock(&rpc->crpc_lock);
1253         }
1254
1255 abort:
1256         if (rpc->crpc_aborted) {
1257                 LNetMDUnlink(rpc->crpc_reqstmdh);
1258                 LNetMDUnlink(rpc->crpc_replymdh);
1259                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1260
1261                 if (!srpc_event_pending(rpc)) {
1262                         srpc_client_rpc_done(rpc, -EINTR);
1263                         return 1;
1264                 }
1265         }
1266         return 0;
1267 }
1268
1269 srpc_client_rpc_t *
1270 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1271                         int nbulkiov, int bulklen,
1272                         void (*rpc_done)(srpc_client_rpc_t *),
1273                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1274 {
1275         srpc_client_rpc_t *rpc;
1276
1277         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1278                                    crpc_bulk.bk_iovs[nbulkiov]));
1279         if (rpc == NULL)
1280                 return NULL;
1281
1282         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1283                              bulklen, rpc_done, rpc_fini, priv);
1284         return rpc;
1285 }
1286
1287 /* called with rpc->crpc_lock held */
1288 static inline void
1289 srpc_queue_rpc (srpc_peer_t *peer, srpc_client_rpc_t *rpc)
1290 {
1291         int service = rpc->crpc_service;
1292
1293         LASSERT (peer->stp_nid == rpc->crpc_dest.nid);
1294         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1295
1296         rpc->crpc_peer = peer;
1297
1298         spin_lock(&peer->stp_lock);
1299
1300         /* Framework RPCs that alter session state shall take precedence
1301          * over test RPCs and framework query RPCs */
1302         if (service <= SRPC_FRAMEWORK_SERVICE_MAX_ID &&
1303             service != SRPC_SERVICE_DEBUG &&
1304             service != SRPC_SERVICE_QUERY_STAT)
1305                 list_add_tail(&rpc->crpc_privl, &peer->stp_ctl_rpcq);
1306         else
1307                 list_add_tail(&rpc->crpc_privl, &peer->stp_rpcq);
1308
1309         srpc_client_rpc_addref(rpc); /* ++ref for peer->*rpcq */
1310         spin_unlock(&peer->stp_lock);
1311         return;
1312 }
1313
1314 /* called with rpc->crpc_lock held */
1315 void
1316 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1317 {
1318         srpc_peer_t *peer = rpc->crpc_peer;
1319
1320         LASSERT (why != 0);
1321
1322         if (rpc->crpc_aborted || /* already aborted */
1323             rpc->crpc_closed)    /* callback imminent */
1324                 return;
1325
1326         CDEBUG (D_NET,
1327                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1328                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1329                 swi_state2str(rpc->crpc_wi.wi_state), why);
1330
1331         rpc->crpc_aborted = 1;
1332         rpc->crpc_status  = why;
1333
1334         if (peer != NULL) {
1335                 spin_lock(&peer->stp_lock);
1336
1337                 if (!list_empty(&rpc->crpc_privl)) { /* still queued */
1338                         list_del_init(&rpc->crpc_privl);
1339                         srpc_client_rpc_decref(rpc); /* --ref for peer->*rpcq */
1340                         rpc->crpc_peer = NULL;       /* no credit taken */
1341                 }
1342
1343                 spin_unlock(&peer->stp_lock);
1344         }
1345
1346         swi_schedule_workitem(&rpc->crpc_wi);
1347         return;
1348 }
1349
1350 /* called with rpc->crpc_lock held */
1351 void
1352 srpc_post_rpc (srpc_client_rpc_t *rpc)
1353 {
1354         srpc_peer_t *peer;
1355
1356         LASSERT (!rpc->crpc_aborted);
1357         LASSERT (rpc->crpc_peer == NULL);
1358         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1359         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1360
1361         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1362                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1363                 rpc->crpc_timeout);
1364
1365         srpc_add_client_rpc_timer(rpc);
1366
1367         peer = srpc_nid2peer(rpc->crpc_dest.nid);
1368         if (peer == NULL) {
1369                 srpc_abort_rpc(rpc, -ENOMEM);
1370                 return;
1371         }
1372
1373         srpc_queue_rpc(peer, rpc);
1374
1375         spin_unlock(&rpc->crpc_lock);
1376         srpc_check_sends(peer, 0);
1377         spin_lock(&rpc->crpc_lock);
1378         return;
1379 }
1380
1381
1382 int
1383 srpc_send_reply (srpc_server_rpc_t *rpc)
1384 {
1385         srpc_event_t   *ev = &rpc->srpc_ev;
1386         srpc_msg_t     *msg = &rpc->srpc_replymsg;
1387         srpc_buffer_t  *buffer = rpc->srpc_reqstbuf;
1388         srpc_service_t *sv = rpc->srpc_service;
1389         __u64           rpyid;
1390         int             rc;
1391
1392         LASSERT (buffer != NULL);
1393         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1394
1395         spin_lock(&sv->sv_lock);
1396
1397         if (!sv->sv_shuttingdown &&
1398             sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1399                 /* Repost buffer before replying since test client
1400                  * might send me another RPC once it gets the reply */
1401                 if (srpc_service_post_buffer(sv, buffer) != 0)
1402                         CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1403                 rpc->srpc_reqstbuf = NULL;
1404         }
1405
1406         spin_unlock(&sv->sv_lock);
1407
1408         ev->ev_fired = 0;
1409         ev->ev_data  = rpc;
1410         ev->ev_type  = SRPC_REPLY_SENT;
1411
1412         msg->msg_magic   = SRPC_MSG_MAGIC;
1413         msg->msg_version = SRPC_MSG_VERSION;
1414         msg->msg_type    = srpc_service2reply(sv->sv_id);
1415
1416         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1417                                    sizeof(*msg), LNET_MD_OP_PUT,
1418                                    rpc->srpc_peer, rpc->srpc_self,
1419                                    &rpc->srpc_replymdh, ev);
1420         if (rc != 0)
1421                 ev->ev_fired = 1;  /* no more event expected */
1422         return rc;
1423 }
1424
1425 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1426 void
1427 srpc_lnet_ev_handler (lnet_event_t *ev)
1428 {
1429         srpc_event_t      *rpcev = ev->md.user_ptr;
1430         srpc_client_rpc_t *crpc;
1431         srpc_server_rpc_t *srpc;
1432         srpc_buffer_t     *buffer;
1433         srpc_service_t    *sv;
1434         srpc_msg_t        *msg;
1435         srpc_msg_type_t    type;
1436         int                fired_flag = 1;
1437
1438         LASSERT (!in_interrupt());
1439
1440         if (ev->status != 0) {
1441                 spin_lock(&srpc_data.rpc_glock);
1442                 srpc_data.rpc_counters.errors++;
1443                 spin_unlock(&srpc_data.rpc_glock);
1444         }
1445
1446         rpcev->ev_lnet = ev->type;
1447
1448         switch (rpcev->ev_type) {
1449         default:
1450                 LBUG ();
1451         case SRPC_REQUEST_SENT:
1452                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1453                         spin_lock(&srpc_data.rpc_glock);
1454                         srpc_data.rpc_counters.rpcs_sent++;
1455                         spin_unlock(&srpc_data.rpc_glock);
1456                 }
1457         case SRPC_REPLY_RCVD:
1458         case SRPC_BULK_REQ_RCVD:
1459                 crpc = rpcev->ev_data;
1460
1461                 LASSERT (rpcev == &crpc->crpc_reqstev ||
1462                          rpcev == &crpc->crpc_replyev ||
1463                          rpcev == &crpc->crpc_bulkev);
1464
1465                 spin_lock(&crpc->crpc_lock);
1466
1467                 LASSERT (rpcev->ev_fired == 0);
1468                 rpcev->ev_fired  = 1;
1469                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1470                                                 -EINTR : ev->status;
1471                 swi_schedule_workitem(&crpc->crpc_wi);
1472
1473                 spin_unlock(&crpc->crpc_lock);
1474                 break;
1475
1476         case SRPC_REQUEST_RCVD:
1477                 sv = rpcev->ev_data;
1478
1479                 LASSERT (rpcev == &sv->sv_ev);
1480
1481                 spin_lock(&sv->sv_lock);
1482
1483                 LASSERT (ev->unlinked);
1484                 LASSERT (ev->type == LNET_EVENT_PUT ||
1485                          ev->type == LNET_EVENT_UNLINK);
1486                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1487                          sv->sv_shuttingdown);
1488
1489                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1490                 buffer->buf_peer = ev->initiator;
1491                 buffer->buf_self = ev->target.nid;
1492
1493                 sv->sv_nposted_msg--;
1494                 LASSERT (sv->sv_nposted_msg >= 0);
1495
1496                 if (sv->sv_shuttingdown) {
1497                         /* Leave buffer on sv->sv_posted_msgq since
1498                          * srpc_finish_service needs to traverse it. */
1499                         spin_unlock(&sv->sv_lock);
1500                         break;
1501                 }
1502
1503                 list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1504                 msg = &buffer->buf_msg;
1505                 type = srpc_service2request(sv->sv_id);
1506
1507                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1508                     (msg->msg_type != type &&
1509                      msg->msg_type != __swab32(type)) ||
1510                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1511                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1512                         CERROR ("Dropping RPC (%s) from %s: "
1513                                 "status %d mlength %d type %u magic %u.\n",
1514                                 sv->sv_name, libcfs_id2str(ev->initiator),
1515                                 ev->status, ev->mlength,
1516                                 msg->msg_type, msg->msg_magic);
1517
1518                         /* NB can't call srpc_service_recycle_buffer here since
1519                          * it may call LNetM[DE]Attach. The invalid magic tells
1520                          * srpc_handle_rpc to drop this RPC */
1521                         msg->msg_magic = 0;
1522                 }
1523
1524                 if (!list_empty(&sv->sv_free_rpcq)) {
1525                         srpc = list_entry(sv->sv_free_rpcq.next,
1526                                           srpc_server_rpc_t, srpc_list);
1527                         list_del(&srpc->srpc_list);
1528
1529                         srpc_init_server_rpc(srpc, sv, buffer);
1530                         list_add_tail(&srpc->srpc_list, &sv->sv_active_rpcq);
1531                         srpc_schedule_server_rpc(srpc);
1532                 } else {
1533                         list_add_tail(&buffer->buf_list, &sv->sv_blocked_msgq);
1534                 }
1535
1536                 spin_unlock(&sv->sv_lock);
1537
1538                 spin_lock(&srpc_data.rpc_glock);
1539                 srpc_data.rpc_counters.rpcs_rcvd++;
1540                 spin_unlock(&srpc_data.rpc_glock);
1541                 break;
1542
1543         case SRPC_BULK_GET_RPLD:
1544                 LASSERT (ev->type == LNET_EVENT_SEND ||
1545                          ev->type == LNET_EVENT_REPLY ||
1546                          ev->type == LNET_EVENT_UNLINK);
1547
1548                 if (ev->type == LNET_EVENT_SEND && !ev->unlinked) {
1549                         if (ev->status == 0)
1550                                 break; /* wait for the final LNET_EVENT_REPLY */
1551                         else
1552                                 fired_flag = 0; /* LNET_EVENT_REPLY may arrive
1553                                                    (optimized GET case) */
1554                 }
1555         case SRPC_BULK_PUT_SENT:
1556                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1557                         spin_lock(&srpc_data.rpc_glock);
1558
1559                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1560                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1561                         else
1562                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1563
1564                         spin_unlock(&srpc_data.rpc_glock);
1565                 }
1566         case SRPC_REPLY_SENT:
1567                 srpc = rpcev->ev_data;
1568                 sv = srpc->srpc_service;
1569
1570                 LASSERT (rpcev == &srpc->srpc_ev);
1571
1572                 spin_lock(&sv->sv_lock);
1573                 if (fired_flag)
1574                         rpcev->ev_fired  = 1;
1575
1576                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1577                                                 -EINTR : ev->status;
1578
1579                 srpc_schedule_server_rpc(srpc);
1580                 spin_unlock(&sv->sv_lock);
1581                 break;
1582         }
1583
1584         return;
1585 }
1586
1587 #ifndef __KERNEL__
1588
1589 int
1590 srpc_check_event (int timeout)
1591 {
1592         lnet_event_t ev;
1593         int          rc;
1594         int          i;
1595
1596         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1597                         timeout * 1000, &ev, &i);
1598         if (rc == 0) return 0;
1599
1600         LASSERT (rc == -EOVERFLOW || rc == 1);
1601
1602         /* We can't affort to miss any events... */
1603         if (rc == -EOVERFLOW) {
1604                 CERROR ("Dropped an event!!!\n");
1605                 abort();
1606         }
1607
1608         srpc_lnet_ev_handler(&ev);
1609         return 1;
1610 }
1611
1612 #endif
1613
1614 int
1615 srpc_startup (void)
1616 {
1617         int i;
1618         int rc;
1619
1620 #ifndef __KERNEL__
1621         char *s;
1622
1623         s = getenv("SRPC_PEER_CREDITS");
1624         srpc_peer_credits = (s != NULL) ? atoi(s) : srpc_peer_credits;
1625 #endif
1626
1627         if (srpc_peer_credits <= 0) {
1628                 CERROR("Peer credits must be positive: %d\n", srpc_peer_credits);
1629                 return -EINVAL;
1630         }
1631
1632         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1633         spin_lock_init(&srpc_data.rpc_glock);
1634
1635         /* 1 second pause to avoid timestamp reuse */
1636         cfs_pause(cfs_time_seconds(1));
1637         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1638
1639         srpc_data.rpc_state = SRPC_STATE_NONE;
1640
1641         LIBCFS_ALLOC(srpc_data.rpc_peers,
1642                      sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1643         if (srpc_data.rpc_peers == NULL) {
1644                 CERROR ("Failed to alloc peer hash.\n");
1645                 return -ENOMEM;
1646         }
1647
1648         for (i = 0; i < SRPC_PEER_HASH_SIZE; i++)
1649                 CFS_INIT_LIST_HEAD(&srpc_data.rpc_peers[i]);
1650
1651 #ifdef __KERNEL__
1652         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1653 #else
1654         if (the_lnet.ln_server_mode_flag)
1655                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1656         else
1657                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1658 #endif
1659         if (rc < 0) {
1660                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1661                 LIBCFS_FREE(srpc_data.rpc_peers,
1662                             sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1663                 return rc;
1664         }
1665
1666         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1667
1668         srpc_data.rpc_lnet_eq = LNET_EQ_NONE;
1669 #ifdef __KERNEL__
1670         rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1671 #else
1672         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1673 #endif
1674         if (rc != 0) {
1675                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1676                 goto bail;
1677         }
1678
1679         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1680         LASSERT (rc == 0);
1681
1682         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1683
1684         rc = swi_startup();
1685         if (rc != 0)
1686                 goto bail;
1687
1688         srpc_data.rpc_state = SRPC_STATE_WI_INIT;
1689
1690         rc = stt_startup();
1691
1692 bail:
1693         if (rc != 0)
1694                 srpc_shutdown();
1695         else
1696                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1697
1698         return rc;
1699 }
1700
1701 void
1702 srpc_shutdown (void)
1703 {
1704         int i;
1705         int rc;
1706         int state;
1707
1708         state = srpc_data.rpc_state;
1709         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1710
1711         switch (state) {
1712         default:
1713                 LBUG ();
1714         case SRPC_STATE_RUNNING:
1715                 spin_lock(&srpc_data.rpc_glock);
1716
1717                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1718                         srpc_service_t *sv = srpc_data.rpc_services[i];
1719
1720                         LASSERTF (sv == NULL,
1721                                   "service not empty: id %d, name %s\n",
1722                                   i, sv->sv_name);
1723                 }
1724
1725                 spin_unlock(&srpc_data.rpc_glock);
1726
1727                 stt_shutdown();
1728
1729         case SRPC_STATE_WI_INIT:
1730                 swi_shutdown();
1731
1732         case SRPC_STATE_EQ_INIT:
1733                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1734                 LASSERT (rc == 0);
1735                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1736                 LASSERT (rc == 0); /* the EQ should have no user by now */
1737
1738         case SRPC_STATE_NI_INIT:
1739                 LNetNIFini();
1740                 break;
1741         }
1742
1743         /* srpc_peer_t's are kept in hash until shutdown */
1744         for (i = 0; i < SRPC_PEER_HASH_SIZE; i++) {
1745                 srpc_peer_t *peer;
1746
1747                 while (!list_empty(&srpc_data.rpc_peers[i])) {
1748                         peer = list_entry(srpc_data.rpc_peers[i].next,
1749                                           srpc_peer_t, stp_list);
1750                         list_del(&peer->stp_list);
1751
1752                         LASSERT (list_empty(&peer->stp_rpcq));
1753                         LASSERT (list_empty(&peer->stp_ctl_rpcq));
1754                         LASSERT (peer->stp_credits == srpc_peer_credits);
1755
1756                         LIBCFS_FREE(peer, sizeof(srpc_peer_t));
1757                 }
1758         }
1759
1760         LIBCFS_FREE(srpc_data.rpc_peers,
1761                     sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1762         return;
1763 }