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         cfs_list_for_each_entry_typed (peer, peer_list,
 226                                        srpc_peer_t, stp_list) {
 227                 if (peer->stp_nid == nid)
 228                         return peer;
 229         }
 230
 231         return NULL;
 232 }
 233
 234 static srpc_peer_t *
 235 srpc_nid2peer (lnet_nid_t nid)
 236 {
 237         srpc_peer_t *peer;
 238         srpc_peer_t *new_peer;
 239
 240         spin_lock(&srpc_data.rpc_glock);
 241         peer = srpc_find_peer_locked(nid);
 242         spin_unlock(&srpc_data.rpc_glock);
 243
 244         if (peer != NULL)
 245                 return peer;
 246
 247         new_peer = srpc_create_peer(nid);
 248
 249         spin_lock(&srpc_data.rpc_glock);
 250
 251         peer = srpc_find_peer_locked(nid);
 252         if (peer != NULL) {
 253                 spin_unlock(&srpc_data.rpc_glock);
 254                 if (new_peer != NULL)
 255                         LIBCFS_FREE(new_peer, sizeof(srpc_peer_t));
 256
 257                 return peer;
 258         }
 259
 260         if (new_peer == NULL) {
 261                 spin_unlock(&srpc_data.rpc_glock);
 262                 return NULL;
 263         }
 264
 265         list_add_tail(&new_peer->stp_list, srpc_nid2peerlist(nid));
 266         spin_unlock(&srpc_data.rpc_glock);
 267         return new_peer;
 268 }
 269
 270 static inline __u64
 271 srpc_next_id (void)
 272 {
 273         __u64 id;
 274
 275         spin_lock(&srpc_data.rpc_glock);
 276         id = srpc_data.rpc_matchbits++;
 277         spin_unlock(&srpc_data.rpc_glock);
 278         return id;
 279 }
 280
 281 void
 282 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
 283                       srpc_service_t *sv, srpc_buffer_t *buffer)
 284 {
 285         memset(rpc, 0, sizeof(*rpc));
 286         swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc);
 287
 288         rpc->srpc_ev.ev_fired = 1; /* no event expected now */
 289
 290         rpc->srpc_service  = sv;
 291         rpc->srpc_reqstbuf = buffer;
 292         rpc->srpc_peer     = buffer->buf_peer;
 293         rpc->srpc_self     = buffer->buf_self;
 294         LNetInvalidateHandle(&rpc->srpc_replymdh);
 295 }
 296
 297 int
 298 srpc_add_service (srpc_service_t *sv)
 299 {
 300         int                id = sv->sv_id;
 301         int                i;
 302         srpc_server_rpc_t *rpc;
 303
 304         LASSERT (sv->sv_concur > 0);
 305         LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
 306
 307         spin_lock(&srpc_data.rpc_glock);
 308
 309         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
 310
 311         if (srpc_data.rpc_services[id] != NULL) {
 312                 spin_unlock(&srpc_data.rpc_glock);
 313                 return -EBUSY;
 314         }
 315
 316         srpc_data.rpc_services[id] = sv;
 317         spin_unlock(&srpc_data.rpc_glock);
 318
 319         sv->sv_nprune       = 0;
 320         sv->sv_nposted_msg  = 0;
 321         sv->sv_shuttingdown = 0;
 322         spin_lock_init(&sv->sv_lock);
 323         CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
 324         CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
 325         CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
 326         CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
 327
 328         sv->sv_ev.ev_data = sv;
 329         sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
 330
 331         for (i = 0; i < sv->sv_concur; i++) {
 332                 LIBCFS_ALLOC(rpc, sizeof(*rpc));
 333                 if (rpc == NULL) goto enomem;
 334
 335                 list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 336         }
 337
 338         CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
 339                 id, sv->sv_name, sv->sv_concur);
 340         return 0;
 341
 342 enomem:
 343         while (!list_empty(&sv->sv_free_rpcq)) {
 344                 rpc = list_entry(sv->sv_free_rpcq.next,
 345                                  srpc_server_rpc_t, srpc_list);
 346                 list_del(&rpc->srpc_list);
 347                 LIBCFS_FREE(rpc, sizeof(*rpc));
 348         }
 349
 350         spin_lock(&srpc_data.rpc_glock);
 351         srpc_data.rpc_services[id] = NULL;
 352         spin_unlock(&srpc_data.rpc_glock);
 353         return -ENOMEM;
 354 }
 355
 356 int
 357 srpc_remove_service (srpc_service_t *sv)
 358 {
 359         int id = sv->sv_id;
 360
 361         spin_lock(&srpc_data.rpc_glock);
 362
 363         if (srpc_data.rpc_services[id] != sv) {
 364                 spin_unlock(&srpc_data.rpc_glock);
 365                 return -ENOENT;
 366         }
 367
 368         srpc_data.rpc_services[id] = NULL;
 369         spin_unlock(&srpc_data.rpc_glock);
 370         return 0;
 371 }
 372
 373 int
 374 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
 375                        int len, int options, lnet_process_id_t peer,
 376                        lnet_handle_md_t *mdh, srpc_event_t *ev)
 377 {
 378         int              rc;
 379         lnet_md_t        md;
 380         lnet_handle_me_t meh;
 381
 382         rc = LNetMEAttach(portal, peer, matchbits, 0,
 383                           LNET_UNLINK, LNET_INS_AFTER, &meh);
 384         if (rc != 0) {
 385                 CERROR ("LNetMEAttach failed: %d\n", rc);
 386                 LASSERT (rc == -ENOMEM);
 387                 return -ENOMEM;
 388         }
 389
 390         md.threshold = 1;
 391         md.user_ptr  = ev;
 392         md.start     = buf;
 393         md.length    = len;
 394         md.options   = options;
 395         md.eq_handle = srpc_data.rpc_lnet_eq;
 396
 397         rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
 398         if (rc != 0) {
 399                 CERROR ("LNetMDAttach failed: %d\n", rc);
 400                 LASSERT (rc == -ENOMEM);
 401
 402                 rc = LNetMEUnlink(meh);
 403                 LASSERT (rc == 0);
 404                 return -ENOMEM;
 405         }
 406
 407         CDEBUG (D_NET,
 408                 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
 409                 libcfs_id2str(peer), portal, matchbits);
 410         return 0;
 411 }
 412
 413 int
 414 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
 415                       int options, lnet_process_id_t peer, lnet_nid_t self,
 416                       lnet_handle_md_t *mdh, srpc_event_t *ev)
 417 {
 418         int       rc;
 419         lnet_md_t md;
 420
 421         md.user_ptr  = ev;
 422         md.start     = buf;
 423         md.length    = len;
 424         md.eq_handle = srpc_data.rpc_lnet_eq;
 425         md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
 426         md.options   = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
 427
 428         rc = LNetMDBind(md, LNET_UNLINK, mdh);
 429         if (rc != 0) {
 430                 CERROR ("LNetMDBind failed: %d\n", rc);
 431                 LASSERT (rc == -ENOMEM);
 432                 return -ENOMEM;
 433         }
 434
 435         /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
 436          * they're only meaningful for MDs attached to an ME (i.e. passive
 437          * buffers... */
 438         if ((options & LNET_MD_OP_PUT) != 0) {
 439                 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
 440                              portal, matchbits, 0, 0);
 441         } else {
 442                 LASSERT ((options & LNET_MD_OP_GET) != 0);
 443
 444                 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
 445         }
 446
 447         if (rc != 0) {
 448                 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
 449                         ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
 450                         libcfs_id2str(peer), portal, matchbits, rc);
 451
 452                 /* The forthcoming unlink event will complete this operation
 453                  * with failure, so fall through and return success here.
 454                  */
 455                 rc = LNetMDUnlink(*mdh);
 456                 LASSERT (rc == 0);
 457         } else {
 458                 CDEBUG (D_NET,
 459                         "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
 460                         libcfs_id2str(peer), portal, matchbits);
 461         }
 462         return 0;
 463 }
 464
 465 int
 466 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
 467                         int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
 468 {
 469         int rc;
 470         int portal;
 471
 472         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 473                 portal = SRPC_REQUEST_PORTAL;
 474         else
 475                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 476
 477         rc = srpc_post_active_rdma(portal, service, buf, len,
 478                                    LNET_MD_OP_PUT, peer,
 479                                    LNET_NID_ANY, mdh, ev);
 480         return rc;
 481 }
 482
 483 int
 484 srpc_post_passive_rqtbuf(int service, void *buf, int len,
 485                          lnet_handle_md_t *mdh, srpc_event_t *ev)
 486 {
 487         int               rc;
 488         int               portal;
 489         lnet_process_id_t any = {0};
 490
 491         any.nid = LNET_NID_ANY;
 492         any.pid = LNET_PID_ANY;
 493
 494         if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 495                 portal = SRPC_REQUEST_PORTAL;
 496         else
 497                 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
 498
 499         rc = srpc_post_passive_rdma(portal, service, buf, len,
 500                                     LNET_MD_OP_PUT, any, mdh, ev);
 501         return rc;
 502 }
 503
 504 int
 505 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 506 {
 507         srpc_msg_t *msg = &buf->buf_msg;
 508         int         rc;
 509
 510         LASSERT (!sv->sv_shuttingdown);
 511
 512         LNetInvalidateHandle(&buf->buf_mdh);
 513         list_add(&buf->buf_list, &sv->sv_posted_msgq);
 514         sv->sv_nposted_msg++;
 515         spin_unlock(&sv->sv_lock);
 516
 517         rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
 518                                       &buf->buf_mdh, &sv->sv_ev);
 519
 520         /* At this point, a RPC (new or delayed) may have arrived in
 521          * msg and its event handler has been called. So we must add
 522          * buf to sv_posted_msgq _before_ dropping sv_lock */
 523
 524         spin_lock(&sv->sv_lock);
 525
 526         if (rc == 0) {
 527                 if (sv->sv_shuttingdown) {
 528                         spin_unlock(&sv->sv_lock);
 529
 530                         /* srpc_shutdown_service might have tried to unlink me
 531                          * when my buf_mdh was still invalid */
 532                         LNetMDUnlink(buf->buf_mdh);
 533
 534                         spin_lock(&sv->sv_lock);
 535                 }
 536                 return 0;
 537         }
 538
 539         sv->sv_nposted_msg--;
 540         if (sv->sv_shuttingdown) return rc;
 541
 542         list_del(&buf->buf_list);
 543
 544         spin_unlock(&sv->sv_lock);
 545         LIBCFS_FREE(buf, sizeof(*buf));
 546         spin_lock(&sv->sv_lock);
 547         return rc;
 548 }
 549
 550 int
 551 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
 552 {
 553         int                rc;
 554         int                posted;
 555         srpc_buffer_t     *buf;
 556
 557         LASSERTF (nbuffer > 0,
 558                   "nbuffer must be positive: %d\n", nbuffer);
 559
 560         for (posted = 0; posted < nbuffer; posted++) {
 561                 LIBCFS_ALLOC(buf, sizeof(*buf));
 562                 if (buf == NULL) break;
 563
 564                 spin_lock(&sv->sv_lock);
 565                 rc = srpc_service_post_buffer(sv, buf);
 566                 spin_unlock(&sv->sv_lock);
 567
 568                 if (rc != 0) break;
 569         }
 570
 571         return posted;
 572 }
 573
 574 void
 575 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
 576 {
 577         LASSERTF (nbuffer > 0,
 578                   "nbuffer must be positive: %d\n", nbuffer);
 579
 580         spin_lock(&sv->sv_lock);
 581
 582         LASSERT (sv->sv_nprune >= 0);
 583         LASSERT (!sv->sv_shuttingdown);
 584
 585         sv->sv_nprune += nbuffer;
 586
 587         spin_unlock(&sv->sv_lock);
 588         return;
 589 }
 590
 591 /* returns 1 if sv has finished, otherwise 0 */
 592 int
 593 srpc_finish_service (srpc_service_t *sv)
 594 {
 595         srpc_server_rpc_t *rpc;
 596         srpc_buffer_t     *buf;
 597
 598         spin_lock(&sv->sv_lock);
 599
 600         LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
 601
 602         if (sv->sv_nposted_msg != 0 || !list_empty(&sv->sv_active_rpcq)) {
 603                 CDEBUG (D_NET,
 604                         "waiting for %d posted buffers to unlink and "
 605                         "in-flight RPCs to die.\n",
 606                         sv->sv_nposted_msg);
 607
 608                 if (!list_empty(&sv->sv_active_rpcq)) {
 609                         rpc = list_entry(sv->sv_active_rpcq.next,
 610                                          srpc_server_rpc_t, srpc_list);
 611                         CDEBUG (D_NETERROR,
 612                                 "Active RPC on shutdown: sv %s, peer %s, "
 613                                 "wi %s scheduled %d running %d, "
 614                                 "ev fired %d type %d status %d lnet %d\n",
 615                                 sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 616                                 swi_state2str(rpc->srpc_wi.wi_state),
 617                                 rpc->srpc_wi.wi_scheduled,
 618                                 rpc->srpc_wi.wi_running,
 619                                 rpc->srpc_ev.ev_fired,
 620                                 rpc->srpc_ev.ev_type,
 621                                 rpc->srpc_ev.ev_status,
 622                                 rpc->srpc_ev.ev_lnet);
 623                 }
 624
 625                 spin_unlock(&sv->sv_lock);
 626                 return 0;
 627         }
 628
 629         spin_unlock(&sv->sv_lock); /* no lock needed from now on */
 630
 631         for (;;) {
 632                 struct list_head *q;
 633
 634                 if (!list_empty(&sv->sv_posted_msgq))
 635                         q = &sv->sv_posted_msgq;
 636                 else if (!list_empty(&sv->sv_blocked_msgq))
 637                         q = &sv->sv_blocked_msgq;
 638                 else
 639                         break;
 640
 641                 buf = list_entry(q->next, srpc_buffer_t, buf_list);
 642                 list_del(&buf->buf_list);
 643
 644                 LIBCFS_FREE(buf, sizeof(*buf));
 645         }
 646
 647         while (!list_empty(&sv->sv_free_rpcq)) {
 648                 rpc = list_entry(sv->sv_free_rpcq.next,
 649                                  srpc_server_rpc_t, srpc_list);
 650                 list_del(&rpc->srpc_list);
 651                 LIBCFS_FREE(rpc, sizeof(*rpc));
 652         }
 653
 654         return 1;
 655 }
 656
 657 /* called with sv->sv_lock held */
 658 void
 659 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
 660 {
 661         if (sv->sv_shuttingdown) goto free;
 662
 663         if (sv->sv_nprune == 0) {
 664                 if (srpc_service_post_buffer(sv, buf) != 0)
 665                         CWARN ("Failed to post %s buffer\n", sv->sv_name);
 666                 return;
 667         }
 668
 669         sv->sv_nprune--;
 670 free:
 671         spin_unlock(&sv->sv_lock);
 672         LIBCFS_FREE(buf, sizeof(*buf));
 673         spin_lock(&sv->sv_lock);
 674 }
 675
 676 void
 677 srpc_shutdown_service (srpc_service_t *sv)
 678 {
 679         srpc_server_rpc_t *rpc;
 680         srpc_buffer_t     *buf;
 681
 682         spin_lock(&sv->sv_lock);
 683
 684         CDEBUG (D_NET, "Shutting down service: id %d, name %s\n",
 685                 sv->sv_id, sv->sv_name);
 686
 687         sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
 688
 689         /* schedule in-flight RPCs to notice the shutdown */
 690         cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
 691                                        srpc_server_rpc_t, srpc_list) {
 692                 swi_schedule_workitem(&rpc->srpc_wi);
 693         }
 694
 695         spin_unlock(&sv->sv_lock);
 696
 697         /* OK to traverse sv_posted_msgq without lock, since no one
 698          * touches sv_posted_msgq now */
 699         cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
 700                                        srpc_buffer_t, buf_list)
 701                 LNetMDUnlink(buf->buf_mdh);
 702
 703         return;
 704 }
 705
 706 int
 707 srpc_send_request (srpc_client_rpc_t *rpc)
 708 {
 709         srpc_event_t *ev = &rpc->crpc_reqstev;
 710         int           rc;
 711
 712         ev->ev_fired = 0;
 713         ev->ev_data  = rpc;
 714         ev->ev_type  = SRPC_REQUEST_SENT;
 715
 716         rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
 717                                      &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
 718                                      &rpc->crpc_reqstmdh, ev);
 719         if (rc != 0) {
 720                 LASSERT (rc == -ENOMEM);
 721                 ev->ev_fired = 1;  /* no more event expected */
 722         }
 723         return rc;
 724 }
 725
 726 int
 727 srpc_prepare_reply (srpc_client_rpc_t *rpc)
 728 {
 729         srpc_event_t *ev = &rpc->crpc_replyev;
 730         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
 731         int           rc;
 732
 733         ev->ev_fired = 0;
 734         ev->ev_data  = rpc;
 735         ev->ev_type  = SRPC_REPLY_RCVD;
 736
 737         *id = srpc_next_id();
 738
 739         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 740                                     &rpc->crpc_replymsg, sizeof(srpc_msg_t),
 741                                     LNET_MD_OP_PUT, rpc->crpc_dest,
 742                                     &rpc->crpc_replymdh, ev);
 743         if (rc != 0) {
 744                 LASSERT (rc == -ENOMEM);
 745                 ev->ev_fired = 1;  /* no more event expected */
 746         }
 747         return rc;
 748 }
 749
 750 int
 751 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
 752 {
 753         srpc_bulk_t  *bk = &rpc->crpc_bulk;
 754         srpc_event_t *ev = &rpc->crpc_bulkev;
 755         __u64        *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
 756         int           rc;
 757         int           opt;
 758
 759         LASSERT (bk->bk_niov <= LNET_MAX_IOV);
 760
 761         if (bk->bk_niov == 0) return 0; /* nothing to do */
 762
 763         opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
 764 #ifdef __KERNEL__
 765         opt |= LNET_MD_KIOV;
 766 #else
 767         opt |= LNET_MD_IOVEC;
 768 #endif
 769
 770         ev->ev_fired = 0;
 771         ev->ev_data  = rpc;
 772         ev->ev_type  = SRPC_BULK_REQ_RCVD;
 773
 774         *id = srpc_next_id();
 775
 776         rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
 777                                     &bk->bk_iovs[0], bk->bk_niov, opt,
 778                                     rpc->crpc_dest, &bk->bk_mdh, ev);
 779         if (rc != 0) {
 780                 LASSERT (rc == -ENOMEM);
 781                 ev->ev_fired = 1;  /* no more event expected */
 782         }
 783         return rc;
 784 }
 785
 786 int
 787 srpc_do_bulk (srpc_server_rpc_t *rpc)
 788 {
 789         srpc_event_t  *ev = &rpc->srpc_ev;
 790         srpc_bulk_t   *bk = rpc->srpc_bulk;
 791         __u64          id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
 792         int            rc;
 793         int            opt;
 794
 795         LASSERT (bk != NULL);
 796
 797         opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
 798 #ifdef __KERNEL__
 799         opt |= LNET_MD_KIOV;
 800 #else
 801         opt |= LNET_MD_IOVEC;
 802 #endif
 803
 804         ev->ev_fired = 0;
 805         ev->ev_data  = rpc;
 806         ev->ev_type  = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
 807
 808         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
 809                                    &bk->bk_iovs[0], bk->bk_niov, opt,
 810                                    rpc->srpc_peer, rpc->srpc_self,
 811                                    &bk->bk_mdh, ev);
 812         if (rc != 0)
 813                 ev->ev_fired = 1;  /* no more event expected */
 814         return rc;
 815 }
 816
 817 /* called with srpc_service_t::sv_lock held */
 818 inline void
 819 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
 820 {
 821         srpc_service_t *sv = rpc->srpc_service;
 822
 823         if (sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID)
 824                 swi_schedule_workitem(&rpc->srpc_wi);
 825         else    /* framework RPCs are handled one by one */
 826                 swi_schedule_serial_workitem(&rpc->srpc_wi);
 827
 828         return;
 829 }
 830
 831 /* only called from srpc_handle_rpc */
 832 void
 833 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
 834 {
 835         srpc_service_t *sv = rpc->srpc_service;
 836         srpc_buffer_t  *buffer;
 837
 838         LASSERT (status != 0 || rpc->srpc_wi.wi_state == SWI_STATE_DONE);
 839
 840         rpc->srpc_status = status;
 841
 842         CDEBUG (status == 0 ? D_NET : D_NETERROR,
 843                 "Server RPC done: service %s, peer %s, status %s:%d\n",
 844                 sv->sv_name, libcfs_id2str(rpc->srpc_peer),
 845                 swi_state2str(rpc->srpc_wi.wi_state), status);
 846
 847         if (status != 0) {
 848                 spin_lock(&srpc_data.rpc_glock);
 849                 srpc_data.rpc_counters.rpcs_dropped++;
 850                 spin_unlock(&srpc_data.rpc_glock);
 851         }
 852
 853         if (rpc->srpc_done != NULL)
 854                 (*rpc->srpc_done) (rpc);
 855         LASSERT (rpc->srpc_bulk == NULL);
 856
 857         spin_lock(&sv->sv_lock);
 858
 859         if (rpc->srpc_reqstbuf != NULL) {
 860                 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
 861                  * sv won't go away for sv_active_rpcq must not be empty */
 862                 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
 863                 rpc->srpc_reqstbuf = NULL;
 864         }
 865
 866         list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
 867
 868         /*
 869          * No one can schedule me now since:
 870          * - I'm not on sv_active_rpcq.
 871          * - all LNet events have been fired.
 872          * Cancel pending schedules and prevent future schedule attempts:
 873          */
 874         LASSERT (rpc->srpc_ev.ev_fired);
 875         swi_kill_workitem(&rpc->srpc_wi);
 876
 877         if (!sv->sv_shuttingdown && !list_empty(&sv->sv_blocked_msgq)) {
 878                 buffer = list_entry(sv->sv_blocked_msgq.next,
 879                                     srpc_buffer_t, buf_list);
 880                 list_del(&buffer->buf_list);
 881
 882                 srpc_init_server_rpc(rpc, sv, buffer);
 883                 list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
 884                 srpc_schedule_server_rpc(rpc);
 885         } else {
 886                 list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
 887         }
 888
 889         spin_unlock(&sv->sv_lock);
 890         return;
 891 }
 892
 893 /* handles an incoming RPC */
 894 int
 895 srpc_handle_rpc (swi_workitem_t *wi)
 896 {
 897         srpc_server_rpc_t *rpc = wi->wi_data;
 898         srpc_service_t    *sv = rpc->srpc_service;
 899         srpc_event_t      *ev = &rpc->srpc_ev;
 900         int                rc = 0;
 901
 902         LASSERT (wi == &rpc->srpc_wi);
 903
 904         spin_lock(&sv->sv_lock);
 905
 906         if (sv->sv_shuttingdown) {
 907                 spin_unlock(&sv->sv_lock);
 908
 909                 if (rpc->srpc_bulk != NULL)
 910                         LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 911                 LNetMDUnlink(rpc->srpc_replymdh);
 912
 913                 if (ev->ev_fired) { /* no more event, OK to finish */
 914                         srpc_server_rpc_done(rpc, -ESHUTDOWN);
 915                         return 1;
 916                 }
 917                 return 0;
 918         }
 919
 920         spin_unlock(&sv->sv_lock);
 921
 922         switch (wi->wi_state) {
 923         default:
 924                 LBUG ();
 925         case SWI_STATE_NEWBORN: {
 926                 srpc_msg_t           *msg;
 927                 srpc_generic_reply_t *reply;
 928
 929                 msg = &rpc->srpc_reqstbuf->buf_msg;
 930                 reply = &rpc->srpc_replymsg.msg_body.reply;
 931
 932                 if (msg->msg_magic == 0) {
 933                         /* moaned already in srpc_lnet_ev_handler */
 934                         rc = EBADMSG;
 935                 } else if (msg->msg_version != SRPC_MSG_VERSION &&
 936                            msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
 937                         CWARN ("Version mismatch: %u, %u expected, from %s\n",
 938                                msg->msg_version, SRPC_MSG_VERSION,
 939                                libcfs_id2str(rpc->srpc_peer));
 940                         reply->status = EPROTO;
 941                 } else {
 942                         reply->status = 0;
 943                         rc = (*sv->sv_handler) (rpc);
 944                         LASSERT (reply->status == 0 || !rpc->srpc_bulk);
 945                 }
 946
 947                 if (rc != 0) {
 948                         srpc_server_rpc_done(rpc, rc);
 949                         return 1;
 950                 }
 951
 952                 wi->wi_state = SWI_STATE_BULK_STARTED;
 953
 954                 if (rpc->srpc_bulk != NULL) {
 955                         rc = srpc_do_bulk(rpc);
 956                         if (rc == 0)
 957                                 return 0; /* wait for bulk */
 958
 959                         LASSERT (ev->ev_fired);
 960                         ev->ev_status = rc;
 961                 }
 962         }
 963         case SWI_STATE_BULK_STARTED:
 964                 /* we cannot LASSERT ev_fired right here because it
 965                  * may be set only upon an event with unlinked==1 */
 966
 967                 if (rpc->srpc_bulk != NULL) {
 968                         rc = ev->ev_status;
 969
 970                         if (sv->sv_bulk_ready != NULL)
 971                                 rc = (*sv->sv_bulk_ready) (rpc, rc);
 972
 973                         if (rc != 0) {
 974                                 if (ev->ev_fired) {
 975                                         srpc_server_rpc_done(rpc, rc);
 976                                         return 1;
 977                                 }
 978
 979                                 rpc->srpc_status = rc;
 980                                 wi->wi_state     = SWI_STATE_BULK_ERRORED;
 981                                 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
 982                                 return 0; /* wait for UNLINK event  */
 983                         }
 984                 }
 985
 986                 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
 987
 988                 wi->wi_state = SWI_STATE_REPLY_SUBMITTED;
 989                 rc = srpc_send_reply(rpc);
 990                 if (rc == 0)
 991                         return 0; /* wait for reply */
 992                 srpc_server_rpc_done(rpc, rc);
 993                 return 1;
 994
 995         case SWI_STATE_REPLY_SUBMITTED:
 996                 LASSERT (ev->ev_fired);
 997
 998                 wi->wi_state = SWI_STATE_DONE;
 999                 srpc_server_rpc_done(rpc, ev->ev_status);
1000                 return 1;
1001
1002         case SWI_STATE_BULK_ERRORED:
1003                 LASSERT (rpc->srpc_bulk != NULL && ev->ev_fired);
1004                 LASSERT (rpc->srpc_status != 0);
1005
1006                 srpc_server_rpc_done(rpc, rpc->srpc_status);
1007                 return 1;
1008         }
1009
1010         return 0;
1011 }
1012
1013 void
1014 srpc_client_rpc_expired (void *data)
1015 {
1016         srpc_client_rpc_t *rpc = data;
1017
1018         CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1019                rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1020                rpc->crpc_timeout);
1021
1022         spin_lock(&rpc->crpc_lock);
1023
1024         rpc->crpc_timeout = 0;
1025         srpc_abort_rpc(rpc, -ETIMEDOUT);
1026
1027         spin_unlock(&rpc->crpc_lock);
1028
1029         spin_lock(&srpc_data.rpc_glock);
1030         srpc_data.rpc_counters.rpcs_expired++;
1031         spin_unlock(&srpc_data.rpc_glock);
1032         return;
1033 }
1034
1035 inline void
1036 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
1037 {
1038         stt_timer_t *timer = &rpc->crpc_timer;
1039
1040         if (rpc->crpc_timeout == 0) return;
1041
1042         CFS_INIT_LIST_HEAD(&timer->stt_list);
1043         timer->stt_data    = rpc;
1044         timer->stt_func    = srpc_client_rpc_expired;
1045         timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1046                                           cfs_time_current_sec());
1047         stt_add_timer(timer);
1048         return;
1049 }
1050
1051 /*
1052  * Called with rpc->crpc_lock held.
1053  *
1054  * Upon exit the RPC expiry timer is not queued and the handler is not
1055  * running on any CPU. */
1056 void
1057 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1058 {
1059         /* timer not planted or already exploded */
1060         if (rpc->crpc_timeout == 0) return;
1061
1062         /* timer sucessfully defused */
1063         if (stt_del_timer(&rpc->crpc_timer)) return;
1064
1065 #ifdef __KERNEL__
1066         /* timer detonated, wait for it to explode */
1067         while (rpc->crpc_timeout != 0) {
1068                 spin_unlock(&rpc->crpc_lock);
1069
1070                 cfs_schedule();
1071
1072                 spin_lock(&rpc->crpc_lock);
1073         }
1074 #else
1075         LBUG(); /* impossible in single-threaded runtime */
1076 #endif
1077         return;
1078 }
1079
1080 void
1081 srpc_check_sends (srpc_peer_t *peer, int credits)
1082 {
1083         struct list_head  *q;
1084         srpc_client_rpc_t *rpc;
1085
1086         LASSERT (credits >= 0);
1087         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1088
1089         spin_lock(&peer->stp_lock);
1090         peer->stp_credits += credits;
1091
1092         while (peer->stp_credits) {
1093                 if (!list_empty(&peer->stp_ctl_rpcq))
1094                         q = &peer->stp_ctl_rpcq;
1095                 else if (!list_empty(&peer->stp_rpcq))
1096                         q = &peer->stp_rpcq;
1097                 else
1098                         break;
1099
1100                 peer->stp_credits--;
1101
1102                 rpc = list_entry(q->next, srpc_client_rpc_t, crpc_privl);
1103                 list_del_init(&rpc->crpc_privl);
1104                 srpc_client_rpc_decref(rpc);  /* --ref for peer->*rpcq */
1105
1106                 swi_schedule_workitem(&rpc->crpc_wi);
1107         }
1108
1109         spin_unlock(&peer->stp_lock);
1110         return;
1111 }
1112
1113 void
1114 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1115 {
1116         swi_workitem_t *wi = &rpc->crpc_wi;
1117         srpc_peer_t    *peer = rpc->crpc_peer;
1118
1119         LASSERT (status != 0 || wi->wi_state == SWI_STATE_DONE);
1120
1121         spin_lock(&rpc->crpc_lock);
1122
1123         rpc->crpc_closed = 1;
1124         if (rpc->crpc_status == 0)
1125                 rpc->crpc_status = status;
1126
1127         srpc_del_client_rpc_timer(rpc);
1128
1129         CDEBUG ((status == 0) ? D_NET : D_NETERROR,
1130                 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1131                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1132                 swi_state2str(wi->wi_state), rpc->crpc_aborted, status);
1133
1134         /*
1135          * No one can schedule me now since:
1136          * - RPC timer has been defused.
1137          * - all LNet events have been fired.
1138          * - crpc_closed has been set, preventing srpc_abort_rpc from
1139          *   scheduling me.
1140          * Cancel pending schedules and prevent future schedule attempts:
1141          */
1142         LASSERT (!srpc_event_pending(rpc));
1143         swi_kill_workitem(wi);
1144
1145         spin_unlock(&rpc->crpc_lock);
1146
1147         (*rpc->crpc_done) (rpc);
1148
1149         if (peer != NULL)
1150                 srpc_check_sends(peer, 1);
1151         return;
1152 }
1153
1154 /* sends an outgoing RPC */
1155 int
1156 srpc_send_rpc (swi_workitem_t *wi)
1157 {
1158         int                rc = 0;
1159         srpc_client_rpc_t *rpc = wi->wi_data;
1160         srpc_msg_t        *reply = &rpc->crpc_replymsg;
1161         int                do_bulk = rpc->crpc_bulk.bk_niov > 0;
1162
1163         LASSERT (rpc != NULL);
1164         LASSERT (wi == &rpc->crpc_wi);
1165
1166         spin_lock(&rpc->crpc_lock);
1167
1168         if (rpc->crpc_aborted) {
1169                 spin_unlock(&rpc->crpc_lock);
1170                 goto abort;
1171         }
1172
1173         spin_unlock(&rpc->crpc_lock);
1174
1175         switch (wi->wi_state) {
1176         default:
1177                 LBUG ();
1178         case SWI_STATE_NEWBORN:
1179                 LASSERT (!srpc_event_pending(rpc));
1180
1181                 rc = srpc_prepare_reply(rpc);
1182                 if (rc != 0) {
1183                         srpc_client_rpc_done(rpc, rc);
1184                         return 1;
1185                 }
1186
1187                 rc = srpc_prepare_bulk(rpc);
1188                 if (rc != 0) break;
1189
1190                 wi->wi_state = SWI_STATE_REQUEST_SUBMITTED;
1191                 rc = srpc_send_request(rpc);
1192                 break;
1193
1194         case SWI_STATE_REQUEST_SUBMITTED:
1195                 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1196                  * order; however, they're processed in a strict order:
1197                  * rqt, rpy, and bulk. */
1198                 if (!rpc->crpc_reqstev.ev_fired) break;
1199
1200                 rc = rpc->crpc_reqstev.ev_status;
1201                 if (rc != 0) break;
1202
1203                 wi->wi_state = SWI_STATE_REQUEST_SENT;
1204                 /* perhaps more events, fall thru */
1205         case SWI_STATE_REQUEST_SENT: {
1206                 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1207
1208                 if (!rpc->crpc_replyev.ev_fired) break;
1209
1210                 rc = rpc->crpc_replyev.ev_status;
1211                 if (rc != 0) break;
1212
1213                 if ((reply->msg_type != type &&
1214                      reply->msg_type != __swab32(type)) ||
1215                     (reply->msg_magic != SRPC_MSG_MAGIC &&
1216                      reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1217                         CWARN ("Bad message from %s: type %u (%d expected),"
1218                                " magic %u (%d expected).\n",
1219                                libcfs_id2str(rpc->crpc_dest),
1220                                reply->msg_type, type,
1221                                reply->msg_magic, SRPC_MSG_MAGIC);
1222                         rc = -EBADMSG;
1223                         break;
1224                 }
1225
1226                 if (do_bulk && reply->msg_body.reply.status != 0) {
1227                         CWARN ("Remote error %d at %s, unlink bulk buffer in "
1228                                "case peer didn't initiate bulk transfer\n",
1229                                reply->msg_body.reply.status,
1230                                libcfs_id2str(rpc->crpc_dest));
1231                         LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1232                 }
1233
1234                 wi->wi_state = SWI_STATE_REPLY_RECEIVED;
1235         }
1236         case SWI_STATE_REPLY_RECEIVED:
1237                 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1238
1239                 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1240
1241                 /* Bulk buffer was unlinked due to remote error. Clear error
1242                  * since reply buffer still contains valid data.
1243                  * NB rpc->crpc_done shouldn't look into bulk data in case of
1244                  * remote error. */
1245                 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1246                     rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1247                         rc = 0;
1248
1249                 wi->wi_state = SWI_STATE_DONE;
1250                 srpc_client_rpc_done(rpc, rc);
1251                 return 1;
1252         }
1253
1254         if (rc != 0) {
1255                 spin_lock(&rpc->crpc_lock);
1256                 srpc_abort_rpc(rpc, rc);
1257                 spin_unlock(&rpc->crpc_lock);
1258         }
1259
1260 abort:
1261         if (rpc->crpc_aborted) {
1262                 LNetMDUnlink(rpc->crpc_reqstmdh);
1263                 LNetMDUnlink(rpc->crpc_replymdh);
1264                 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1265
1266                 if (!srpc_event_pending(rpc)) {
1267                         srpc_client_rpc_done(rpc, -EINTR);
1268                         return 1;
1269                 }
1270         }
1271         return 0;
1272 }
1273
1274 srpc_client_rpc_t *
1275 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1276                         int nbulkiov, int bulklen,
1277                         void (*rpc_done)(srpc_client_rpc_t *),
1278                         void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1279 {
1280         srpc_client_rpc_t *rpc;
1281
1282         LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1283                                    crpc_bulk.bk_iovs[nbulkiov]));
1284         if (rpc == NULL)
1285                 return NULL;
1286
1287         srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1288                              bulklen, rpc_done, rpc_fini, priv);
1289         return rpc;
1290 }
1291
1292 /* called with rpc->crpc_lock held */
1293 static inline void
1294 srpc_queue_rpc (srpc_peer_t *peer, srpc_client_rpc_t *rpc)
1295 {
1296         int service = rpc->crpc_service;
1297
1298         LASSERT (peer->stp_nid == rpc->crpc_dest.nid);
1299         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1300
1301         rpc->crpc_peer = peer;
1302
1303         spin_lock(&peer->stp_lock);
1304
1305         /* Framework RPCs that alter session state shall take precedence
1306          * over test RPCs and framework query RPCs */
1307         if (service <= SRPC_FRAMEWORK_SERVICE_MAX_ID &&
1308             service != SRPC_SERVICE_DEBUG &&
1309             service != SRPC_SERVICE_QUERY_STAT)
1310                 list_add_tail(&rpc->crpc_privl, &peer->stp_ctl_rpcq);
1311         else
1312                 list_add_tail(&rpc->crpc_privl, &peer->stp_rpcq);
1313
1314         srpc_client_rpc_addref(rpc); /* ++ref for peer->*rpcq */
1315         spin_unlock(&peer->stp_lock);
1316         return;
1317 }
1318
1319 /* called with rpc->crpc_lock held */
1320 void
1321 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1322 {
1323         srpc_peer_t *peer = rpc->crpc_peer;
1324
1325         LASSERT (why != 0);
1326
1327         if (rpc->crpc_aborted || /* already aborted */
1328             rpc->crpc_closed)    /* callback imminent */
1329                 return;
1330
1331         CDEBUG (D_NET,
1332                 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1333                 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1334                 swi_state2str(rpc->crpc_wi.wi_state), why);
1335
1336         rpc->crpc_aborted = 1;
1337         rpc->crpc_status  = why;
1338
1339         if (peer != NULL) {
1340                 spin_lock(&peer->stp_lock);
1341
1342                 if (!list_empty(&rpc->crpc_privl)) { /* still queued */
1343                         list_del_init(&rpc->crpc_privl);
1344                         srpc_client_rpc_decref(rpc); /* --ref for peer->*rpcq */
1345                         rpc->crpc_peer = NULL;       /* no credit taken */
1346                 }
1347
1348                 spin_unlock(&peer->stp_lock);
1349         }
1350
1351         swi_schedule_workitem(&rpc->crpc_wi);
1352         return;
1353 }
1354
1355 /* called with rpc->crpc_lock held */
1356 void
1357 srpc_post_rpc (srpc_client_rpc_t *rpc)
1358 {
1359         srpc_peer_t *peer;
1360
1361         LASSERT (!rpc->crpc_aborted);
1362         LASSERT (rpc->crpc_peer == NULL);
1363         LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1364         LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1365
1366         CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1367                 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1368                 rpc->crpc_timeout);
1369
1370         srpc_add_client_rpc_timer(rpc);
1371
1372         peer = srpc_nid2peer(rpc->crpc_dest.nid);
1373         if (peer == NULL) {
1374                 srpc_abort_rpc(rpc, -ENOMEM);
1375                 return;
1376         }
1377
1378         srpc_queue_rpc(peer, rpc);
1379
1380         spin_unlock(&rpc->crpc_lock);
1381         srpc_check_sends(peer, 0);
1382         spin_lock(&rpc->crpc_lock);
1383         return;
1384 }
1385
1386
1387 int
1388 srpc_send_reply (srpc_server_rpc_t *rpc)
1389 {
1390         srpc_event_t   *ev = &rpc->srpc_ev;
1391         srpc_msg_t     *msg = &rpc->srpc_replymsg;
1392         srpc_buffer_t  *buffer = rpc->srpc_reqstbuf;
1393         srpc_service_t *sv = rpc->srpc_service;
1394         __u64           rpyid;
1395         int             rc;
1396
1397         LASSERT (buffer != NULL);
1398         rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1399
1400         spin_lock(&sv->sv_lock);
1401
1402         if (!sv->sv_shuttingdown &&
1403             sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1404                 /* Repost buffer before replying since test client
1405                  * might send me another RPC once it gets the reply */
1406                 if (srpc_service_post_buffer(sv, buffer) != 0)
1407                         CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1408                 rpc->srpc_reqstbuf = NULL;
1409         }
1410
1411         spin_unlock(&sv->sv_lock);
1412
1413         ev->ev_fired = 0;
1414         ev->ev_data  = rpc;
1415         ev->ev_type  = SRPC_REPLY_SENT;
1416
1417         msg->msg_magic   = SRPC_MSG_MAGIC;
1418         msg->msg_version = SRPC_MSG_VERSION;
1419         msg->msg_type    = srpc_service2reply(sv->sv_id);
1420
1421         rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1422                                    sizeof(*msg), LNET_MD_OP_PUT,
1423                                    rpc->srpc_peer, rpc->srpc_self,
1424                                    &rpc->srpc_replymdh, ev);
1425         if (rc != 0)
1426                 ev->ev_fired = 1;  /* no more event expected */
1427         return rc;
1428 }
1429
1430 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1431 void
1432 srpc_lnet_ev_handler (lnet_event_t *ev)
1433 {
1434         srpc_event_t      *rpcev = ev->md.user_ptr;
1435         srpc_client_rpc_t *crpc;
1436         srpc_server_rpc_t *srpc;
1437         srpc_buffer_t     *buffer;
1438         srpc_service_t    *sv;
1439         srpc_msg_t        *msg;
1440         srpc_msg_type_t    type;
1441         int                fired_flag = 1;
1442
1443         LASSERT (!in_interrupt());
1444
1445         if (ev->status != 0) {
1446                 spin_lock(&srpc_data.rpc_glock);
1447                 srpc_data.rpc_counters.errors++;
1448                 spin_unlock(&srpc_data.rpc_glock);
1449         }
1450
1451         rpcev->ev_lnet = ev->type;
1452
1453         switch (rpcev->ev_type) {
1454         default:
1455                 LBUG ();
1456         case SRPC_REQUEST_SENT:
1457                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1458                         spin_lock(&srpc_data.rpc_glock);
1459                         srpc_data.rpc_counters.rpcs_sent++;
1460                         spin_unlock(&srpc_data.rpc_glock);
1461                 }
1462         case SRPC_REPLY_RCVD:
1463         case SRPC_BULK_REQ_RCVD:
1464                 crpc = rpcev->ev_data;
1465
1466                 LASSERT (rpcev == &crpc->crpc_reqstev ||
1467                          rpcev == &crpc->crpc_replyev ||
1468                          rpcev == &crpc->crpc_bulkev);
1469
1470                 spin_lock(&crpc->crpc_lock);
1471
1472                 LASSERT (rpcev->ev_fired == 0);
1473                 rpcev->ev_fired  = 1;
1474                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1475                                                 -EINTR : ev->status;
1476                 swi_schedule_workitem(&crpc->crpc_wi);
1477
1478                 spin_unlock(&crpc->crpc_lock);
1479                 break;
1480
1481         case SRPC_REQUEST_RCVD:
1482                 sv = rpcev->ev_data;
1483
1484                 LASSERT (rpcev == &sv->sv_ev);
1485
1486                 spin_lock(&sv->sv_lock);
1487
1488                 LASSERT (ev->unlinked);
1489                 LASSERT (ev->type == LNET_EVENT_PUT ||
1490                          ev->type == LNET_EVENT_UNLINK);
1491                 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1492                          sv->sv_shuttingdown);
1493
1494                 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1495                 buffer->buf_peer = ev->initiator;
1496                 buffer->buf_self = ev->target.nid;
1497
1498                 sv->sv_nposted_msg--;
1499                 LASSERT (sv->sv_nposted_msg >= 0);
1500
1501                 if (sv->sv_shuttingdown) {
1502                         /* Leave buffer on sv->sv_posted_msgq since
1503                          * srpc_finish_service needs to traverse it. */
1504                         spin_unlock(&sv->sv_lock);
1505                         break;
1506                 }
1507
1508                 list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1509                 msg = &buffer->buf_msg;
1510                 type = srpc_service2request(sv->sv_id);
1511
1512                 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1513                     (msg->msg_type != type &&
1514                      msg->msg_type != __swab32(type)) ||
1515                     (msg->msg_magic != SRPC_MSG_MAGIC &&
1516                      msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1517                         CERROR ("Dropping RPC (%s) from %s: "
1518                                 "status %d mlength %d type %u magic %u.\n",
1519                                 sv->sv_name, libcfs_id2str(ev->initiator),
1520                                 ev->status, ev->mlength,
1521                                 msg->msg_type, msg->msg_magic);
1522
1523                         /* NB can't call srpc_service_recycle_buffer here since
1524                          * it may call LNetM[DE]Attach. The invalid magic tells
1525                          * srpc_handle_rpc to drop this RPC */
1526                         msg->msg_magic = 0;
1527                 }
1528
1529                 if (!list_empty(&sv->sv_free_rpcq)) {
1530                         srpc = list_entry(sv->sv_free_rpcq.next,
1531                                           srpc_server_rpc_t, srpc_list);
1532                         list_del(&srpc->srpc_list);
1533
1534                         srpc_init_server_rpc(srpc, sv, buffer);
1535                         list_add_tail(&srpc->srpc_list, &sv->sv_active_rpcq);
1536                         srpc_schedule_server_rpc(srpc);
1537                 } else {
1538                         list_add_tail(&buffer->buf_list, &sv->sv_blocked_msgq);
1539                 }
1540
1541                 spin_unlock(&sv->sv_lock);
1542
1543                 spin_lock(&srpc_data.rpc_glock);
1544                 srpc_data.rpc_counters.rpcs_rcvd++;
1545                 spin_unlock(&srpc_data.rpc_glock);
1546                 break;
1547
1548         case SRPC_BULK_GET_RPLD:
1549                 LASSERT (ev->type == LNET_EVENT_SEND ||
1550                          ev->type == LNET_EVENT_REPLY ||
1551                          ev->type == LNET_EVENT_UNLINK);
1552
1553                 if (ev->type == LNET_EVENT_SEND && !ev->unlinked) {
1554                         if (ev->status == 0)
1555                                 break; /* wait for the final LNET_EVENT_REPLY */
1556                         else
1557                                 fired_flag = 0; /* LNET_EVENT_REPLY may arrive
1558                                                    (optimized GET case) */
1559                 }
1560         case SRPC_BULK_PUT_SENT:
1561                 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1562                         spin_lock(&srpc_data.rpc_glock);
1563
1564                         if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1565                                 srpc_data.rpc_counters.bulk_get += ev->mlength;
1566                         else
1567                                 srpc_data.rpc_counters.bulk_put += ev->mlength;
1568
1569                         spin_unlock(&srpc_data.rpc_glock);
1570                 }
1571         case SRPC_REPLY_SENT:
1572                 srpc = rpcev->ev_data;
1573                 sv = srpc->srpc_service;
1574
1575                 LASSERT (rpcev == &srpc->srpc_ev);
1576
1577                 spin_lock(&sv->sv_lock);
1578                 if (fired_flag)
1579                         rpcev->ev_fired  = 1;
1580
1581                 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1582                                                 -EINTR : ev->status;
1583
1584                 srpc_schedule_server_rpc(srpc);
1585                 spin_unlock(&sv->sv_lock);
1586                 break;
1587         }
1588
1589         return;
1590 }
1591
1592 #ifndef __KERNEL__
1593
1594 int
1595 srpc_check_event (int timeout)
1596 {
1597         lnet_event_t ev;
1598         int          rc;
1599         int          i;
1600
1601         rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1602                         timeout * 1000, &ev, &i);
1603         if (rc == 0) return 0;
1604
1605         LASSERT (rc == -EOVERFLOW || rc == 1);
1606
1607         /* We can't affort to miss any events... */
1608         if (rc == -EOVERFLOW) {
1609                 CERROR ("Dropped an event!!!\n");
1610                 abort();
1611         }
1612
1613         srpc_lnet_ev_handler(&ev);
1614         return 1;
1615 }
1616
1617 #endif
1618
1619 int
1620 srpc_startup (void)
1621 {
1622         int i;
1623         int rc;
1624
1625 #ifndef __KERNEL__
1626         char *s;
1627
1628         s = getenv("SRPC_PEER_CREDITS");
1629         srpc_peer_credits = (s != NULL) ? atoi(s) : srpc_peer_credits;
1630 #endif
1631
1632         if (srpc_peer_credits <= 0) {
1633                 CERROR("Peer credits must be positive: %d\n", srpc_peer_credits);
1634                 return -EINVAL;
1635         }
1636
1637         memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1638         spin_lock_init(&srpc_data.rpc_glock);
1639
1640         /* 1 second pause to avoid timestamp reuse */
1641         cfs_pause(cfs_time_seconds(1));
1642         srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1643
1644         srpc_data.rpc_state = SRPC_STATE_NONE;
1645
1646         LIBCFS_ALLOC(srpc_data.rpc_peers,
1647                      sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1648         if (srpc_data.rpc_peers == NULL) {
1649                 CERROR ("Failed to alloc peer hash.\n");
1650                 return -ENOMEM;
1651         }
1652
1653         for (i = 0; i < SRPC_PEER_HASH_SIZE; i++)
1654                 CFS_INIT_LIST_HEAD(&srpc_data.rpc_peers[i]);
1655
1656 #ifdef __KERNEL__
1657         rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1658 #else
1659         if (the_lnet.ln_server_mode_flag)
1660                 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1661         else
1662                 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1663 #endif
1664         if (rc < 0) {
1665                 CERROR ("LNetNIInit() has failed: %d\n", rc);
1666                 LIBCFS_FREE(srpc_data.rpc_peers,
1667                             sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1668                 return rc;
1669         }
1670
1671         srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1672
1673         LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1674 #ifdef __KERNEL__
1675         rc = LNetEQAlloc(16, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1676 #else
1677         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1678 #endif
1679         if (rc != 0) {
1680                 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1681                 goto bail;
1682         }
1683
1684         rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1685         LASSERT (rc == 0);
1686
1687         srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1688
1689         rc = swi_startup();
1690         if (rc != 0)
1691                 goto bail;
1692
1693         srpc_data.rpc_state = SRPC_STATE_WI_INIT;
1694
1695         rc = stt_startup();
1696
1697 bail:
1698         if (rc != 0)
1699                 srpc_shutdown();
1700         else
1701                 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1702
1703         return rc;
1704 }
1705
1706 void
1707 srpc_shutdown (void)
1708 {
1709         int i;
1710         int rc;
1711         int state;
1712
1713         state = srpc_data.rpc_state;
1714         srpc_data.rpc_state = SRPC_STATE_STOPPING;
1715
1716         switch (state) {
1717         default:
1718                 LBUG ();
1719         case SRPC_STATE_RUNNING:
1720                 spin_lock(&srpc_data.rpc_glock);
1721
1722                 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1723                         srpc_service_t *sv = srpc_data.rpc_services[i];
1724
1725                         LASSERTF (sv == NULL,
1726                                   "service not empty: id %d, name %s\n",
1727                                   i, sv->sv_name);
1728                 }
1729
1730                 spin_unlock(&srpc_data.rpc_glock);
1731
1732                 stt_shutdown();
1733
1734         case SRPC_STATE_WI_INIT:
1735                 swi_shutdown();
1736
1737         case SRPC_STATE_EQ_INIT:
1738                 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1739                 LASSERT (rc == 0);
1740                 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1741                 LASSERT (rc == 0); /* the EQ should have no user by now */
1742
1743         case SRPC_STATE_NI_INIT:
1744                 LNetNIFini();
1745                 break;
1746         }
1747
1748         /* srpc_peer_t's are kept in hash until shutdown */
1749         for (i = 0; i < SRPC_PEER_HASH_SIZE; i++) {
1750                 srpc_peer_t *peer;
1751
1752                 while (!list_empty(&srpc_data.rpc_peers[i])) {
1753                         peer = list_entry(srpc_data.rpc_peers[i].next,
1754                                           srpc_peer_t, stp_list);
1755                         list_del(&peer->stp_list);
1756
1757                         LASSERT (list_empty(&peer->stp_rpcq));
1758                         LASSERT (list_empty(&peer->stp_ctl_rpcq));
1759                         LASSERT (peer->stp_credits == srpc_peer_credits);
1760
1761                         LIBCFS_FREE(peer, sizeof(srpc_peer_t));
1762                 }
1763         }
1764
1765         LIBCFS_FREE(srpc_data.rpc_peers,
1766                     sizeof(struct list_head) * SRPC_PEER_HASH_SIZE);
1767         return;
1768 }