lustre/ptlrpc/rpc.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define EXPORT_SYMTAB
  24
  25 #include <linux/config.h>
  26 #include <linux/module.h>
  27 #include <linux/kernel.h>
  28
  29 #define DEBUG_SUBSYSTEM S_RPC
  30
  31 #include <linux/obd_support.h>
  32 #include <linux/lustre_net.h>
  33
  34 static ptl_handle_eq_t sent_pkt_eq, rcvd_rep_eq,
  35         bulk_source_eq, bulk_sink_eq;
  36
  37
  38 struct ptlrpc_request *ptlrpc_prep_req(struct ptlrpc_client *cl,
  39                                        int opcode, int namelen, char *name,
  40                                        int tgtlen, char *tgt)
  41 {
  42         struct ptlrpc_request *request;
  43         int rc;
  44         ENTRY;
  45
  46         OBD_ALLOC(request, sizeof(*request));
  47         if (!request) {
  48                 CERROR("request allocation out of memory\n");
  49                 return NULL;
  50         }
  51
  52         memset(request, 0, sizeof(*request));
  53         request->rq_xid = cl->cli_xid++;
  54
  55         rc = cl->cli_req_pack(name, namelen, tgt, tgtlen,
  56                           &request->rq_reqhdr, &request->rq_req,
  57                           &request->rq_reqlen, &request->rq_reqbuf);
  58         if (rc) {
  59                 CERROR("cannot pack request %d\n", rc);
  60                 return NULL;
  61         }
  62         request->rq_reqhdr->opc = opcode;
  63         request->rq_reqhdr->seqno = request->rq_xid;
  64
  65         EXIT;
  66         return request;
  67 }
  68
  69 void ptlrpc_free_req(struct ptlrpc_request *request)
  70 {
  71         OBD_FREE(request, sizeof(*request));
  72 }
  73
  74 int ptlrpc_queue_wait(struct ptlrpc_request *req,
  75                              struct ptlrpc_client *cl)
  76 {
  77         int rc;
  78         DECLARE_WAITQUEUE(wait, current);
  79
  80         init_waitqueue_head(&req->rq_wait_for_rep);
  81
  82         if (cl->cli_enqueue) {
  83                 /* Local delivery */
  84                 ENTRY;
  85                 rc = cl->cli_enqueue(req);
  86         } else {
  87                 /* Remote delivery via portals. */
  88                 req->rq_req_portal = cl->cli_request_portal;
  89                 req->rq_reply_portal = cl->cli_reply_portal;
  90                 rc = ptl_send_rpc(req, &cl->cli_server);
  91         }
  92         if (rc) {
  93                 CERROR("error %d, opcode %d\n", rc,
  94                        req->rq_reqhdr->opc);
  95                 return -rc;
  96         }
  97
  98         CDEBUG(0, "-- sleeping\n");
  99         add_wait_queue(&req->rq_wait_for_rep, &wait);
 100         while (req->rq_repbuf == NULL) {
 101                 set_current_state(TASK_INTERRUPTIBLE);
 102
 103                 /* if this process really wants to die, let it go */
 104                 if (sigismember(&(current->pending.signal), SIGKILL) ||
 105                     sigismember(&(current->pending.signal), SIGINT))
 106                         break;
 107
 108                 schedule();
 109         }
 110         remove_wait_queue(&req->rq_wait_for_rep, &wait);
 111         set_current_state(TASK_RUNNING);
 112         CDEBUG(0, "-- done\n");
 113
 114         if (req->rq_repbuf == NULL) {
 115                 /* We broke out because of a signal */
 116                 EXIT;
 117                 return -EINTR;
 118         }
 119
 120         rc = cl->cli_rep_unpack(req->rq_repbuf, req->rq_replen, &req->rq_rephdr, &req->rq_rep);
 121         if (rc) {
 122                 CERROR("unpack_rep failed: %d\n", rc);
 123                 return rc;
 124         }
 125         CERROR("got rep %lld\n", req->rq_rephdr->seqno);
 126         if ( req->rq_rephdr->status == 0 )
 127                 CDEBUG(0, "--> buf %p len %d status %d\n",
 128                        req->rq_repbuf, req->rq_replen,
 129                        req->rq_rephdr->status);
 130
 131         EXIT;
 132         return 0;
 133 }
 134 /*
 135  *  Free the packet when it has gone out
 136  */
 137 static int sent_packet_callback(ptl_event_t *ev, void *data)
 138 {
 139         ENTRY;
 140
 141         if (ev->type == PTL_EVENT_SENT) {
 142                 OBD_FREE(ev->mem_desc.start, ev->mem_desc.length);
 143         } else {
 144                 // XXX make sure we understand all events, including ACK's
 145                 CERROR("Unknown event %d\n", ev->type);
 146                 BUG();
 147         }
 148
 149         EXIT;
 150         return 1;
 151 }
 152
 153 /*
 154  * Wake up the thread waiting for the reply once it comes in.
 155  */
 156 static int rcvd_reply_callback(ptl_event_t *ev, void *data)
 157 {
 158         struct ptlrpc_request *rpc = ev->mem_desc.user_ptr;
 159         ENTRY;
 160
 161         if (ev->type == PTL_EVENT_PUT) {
 162                 rpc->rq_repbuf = ev->mem_desc.start + ev->offset;
 163                 wake_up_interruptible(&rpc->rq_wait_for_rep);
 164         } else {
 165                 // XXX make sure we understand all events, including ACK's
 166                 CERROR("Unknown event %d\n", ev->type);
 167                 BUG();
 168         }
 169
 170         EXIT;
 171         return 1;
 172 }
 173
 174 static int server_request_callback(ptl_event_t *ev, void *data)
 175 {
 176         struct ptlrpc_service *service = data;
 177         int rc;
 178
 179         if (ev->rlength != ev->mlength)
 180                 CERROR("Warning: Possibly truncated rpc (%d/%d)\n",
 181                        ev->mlength, ev->rlength);
 182
 183         /* The ME is unlinked when there is less than 1024 bytes free
 184          * on its MD.  This ensures we are always able to handle the rpc,
 185          * although the 1024 value is a guess as to the size of a
 186          * large rpc (the known safe margin should be determined).
 187          *
 188          * NOTE: The portals API by default unlinks all MD's associated
 189          *       with an ME when it's unlinked.  For now, this behavior
 190          *       has been commented out of the portals library so the
 191          *       MD can be unlinked when its ref count drops to zero.
 192          *       A new MD and ME will then be created that use the same
 193          *       kmalloc()'ed memory and inserted at the ring tail.
 194          */
 195
 196         service->srv_ref_count[service->srv_md_active]++;
 197
 198         if (ev->offset >= (service->srv_buf_size - 1024)) {
 199                 CDEBUG(D_INODE, "Unlinking ME %d\n", service->srv_me_active);
 200
 201                 rc = PtlMEUnlink(service->srv_me_h[service->srv_me_active]);
 202                 service->srv_me_h[service->srv_me_active] = 0;
 203
 204                 if (rc != PTL_OK) {
 205                         CERROR("PtlMEUnlink failed - DROPPING soon: %d\n", rc);
 206                         return rc;
 207                 }
 208
 209                 service->srv_me_active = NEXT_INDEX(service->srv_me_active,
 210                         service->srv_ring_length);
 211
 212                 if (service->srv_me_h[service->srv_me_active] == 0)
 213                         CERROR("All %d ring ME's are unlinked!\n",
 214                                service->srv_ring_length);
 215         }
 216
 217         if (ev->type == PTL_EVENT_PUT) {
 218                 wake_up(service->srv_wait_queue);
 219         } else {
 220                 CERROR("Unexpected event type: %d\n", ev->type);
 221         }
 222
 223         return 0;
 224 }
 225
 226 static int bulk_source_callback(ptl_event_t *ev, void *data)
 227 {
 228         struct ptlrpc_request *rpc = ev->mem_desc.user_ptr;
 229
 230         ENTRY;
 231
 232         if (ev->type == PTL_EVENT_SENT) {
 233                 CDEBUG(D_NET, "got SENT event\n");
 234         } else if (ev->type == PTL_EVENT_ACK) {
 235                 CDEBUG(D_NET, "got ACK event\n");
 236                 wake_up_interruptible(&rpc->rq_wait_for_bulk);
 237         } else {
 238                 CERROR("Unexpected event type!\n");
 239                 BUG();
 240         }
 241
 242         EXIT;
 243         return 1;
 244 }
 245
 246 static int bulk_sink_callback(ptl_event_t *ev, void *data)
 247 {
 248         struct ptlrpc_request *rpc = ev->mem_desc.user_ptr;
 249
 250         ENTRY;
 251
 252         if (ev->type == PTL_EVENT_PUT) {
 253                 if (rpc->rq_bulkbuf != ev->mem_desc.start + ev->offset)
 254                         CERROR("bulkbuf != mem_desc -- why?\n");
 255                 wake_up_interruptible(&rpc->rq_wait_for_bulk);
 256         } else {
 257                 CERROR("Unexpected event type!\n");
 258                 BUG();
 259         }
 260
 261         EXIT;
 262         return 1;
 263 }
 264
 265 int ptl_send_buf(struct ptlrpc_request *request, struct lustre_peer *peer,
 266                  int portal, int is_request)
 267 {
 268         int rc;
 269         ptl_process_id_t remote_id;
 270         ptl_handle_md_t md_h;
 271
 272         /* FIXME: This is bad. */
 273         if (request->rq_bulklen) {
 274                 request->rq_req_md.start = request->rq_bulkbuf;
 275                 request->rq_req_md.length = request->rq_bulklen;
 276                 request->rq_req_md.eventq = bulk_source_eq;
 277         } else if (is_request) {
 278                 request->rq_req_md.start = request->rq_reqbuf;
 279                 request->rq_req_md.length = request->rq_reqlen;
 280                 request->rq_req_md.eventq = sent_pkt_eq;
 281         } else {
 282                 request->rq_req_md.start = request->rq_repbuf;
 283                 request->rq_req_md.length = request->rq_replen;
 284                 request->rq_req_md.eventq = sent_pkt_eq;
 285         }
 286         request->rq_req_md.threshold = 1;
 287         request->rq_req_md.options = PTL_MD_OP_PUT;
 288         request->rq_req_md.user_ptr = request;
 289
 290         rc = PtlMDBind(peer->peer_ni, request->rq_req_md, &md_h);
 291         if (rc != 0) {
 292                 BUG();
 293                 CERROR("PtlMDBind failed: %d\n", rc);
 294                 return rc;
 295         }
 296
 297         remote_id.addr_kind = PTL_ADDR_NID;
 298         remote_id.nid = peer->peer_nid;
 299         remote_id.pid = 0;
 300
 301         if (request->rq_bulklen) {
 302                 rc = PtlPut(md_h, PTL_ACK_REQ, remote_id, portal, 0,
 303                             request->rq_xid, 0, 0);
 304         } else {
 305                 rc = PtlPut(md_h, PTL_NOACK_REQ, remote_id, portal, 0,
 306                             request->rq_xid, 0, 0);
 307         }
 308         if (rc != PTL_OK) {
 309                 BUG();
 310                 CERROR("PtlPut(%d, %d, %d) failed: %d\n", remote_id.nid,
 311                        portal, request->rq_xid, rc);
 312                 /* FIXME: tear down md */
 313         }
 314
 315         return rc;
 316 }
 317
 318 int ptl_send_rpc(struct ptlrpc_request *request, struct lustre_peer *peer)
 319 {
 320         ptl_handle_me_t me_h, bulk_me_h;
 321         ptl_process_id_t local_id;
 322         int rc;
 323         char *repbuf;
 324
 325         ENTRY;
 326
 327         if (request->rq_replen == 0) {
 328                 CERROR("request->rq_replen is 0!\n");
 329                 EXIT;
 330                 return -EINVAL;
 331         }
 332
 333         /* request->rq_repbuf is set only when the reply comes in, in
 334          * client_packet_callback() */
 335         OBD_ALLOC(repbuf, request->rq_replen);
 336         if (!repbuf) {
 337                 EXIT;
 338                 return -ENOMEM;
 339         }
 340
 341         local_id.addr_kind = PTL_ADDR_GID;
 342         local_id.gid = PTL_ID_ANY;
 343         local_id.rid = PTL_ID_ANY;
 344
 345         CERROR("sending req %d\n", request->rq_xid);
 346         rc = PtlMEAttach(peer->peer_ni, request->rq_reply_portal, local_id,
 347                          request->rq_xid, 0, PTL_UNLINK, &me_h);
 348         if (rc != PTL_OK) {
 349                 CERROR("PtlMEAttach failed: %d\n", rc);
 350                 BUG();
 351                 EXIT;
 352                 goto cleanup;
 353         }
 354
 355         request->rq_reply_md.start = repbuf;
 356         request->rq_reply_md.length = request->rq_replen;
 357         request->rq_reply_md.threshold = 1;
 358         request->rq_reply_md.options = PTL_MD_OP_PUT;
 359         request->rq_reply_md.user_ptr = request;
 360         request->rq_reply_md.eventq = rcvd_rep_eq;
 361
 362         rc = PtlMDAttach(me_h, request->rq_reply_md, PTL_UNLINK,
 363                          &request->rq_reply_md_h);
 364         if (rc != PTL_OK) {
 365                 CERROR("PtlMDAttach failed: %d\n", rc);
 366                 BUG();
 367                 EXIT;
 368                 goto cleanup2;
 369         }
 370
 371         if (request->rq_bulklen != 0) {
 372                 rc = PtlMEAttach(peer->peer_ni, request->rq_bulk_portal,
 373                                  local_id, request->rq_xid, 0, PTL_UNLINK,
 374                                  &bulk_me_h);
 375                 if (rc != PTL_OK) {
 376                         CERROR("PtlMEAttach failed: %d\n", rc);
 377                         BUG();
 378                         EXIT;
 379                         goto cleanup3;
 380                 }
 381
 382                 request->rq_bulk_md.start = request->rq_bulkbuf;
 383                 request->rq_bulk_md.length = request->rq_bulklen;
 384                 request->rq_bulk_md.threshold = 1;
 385                 request->rq_bulk_md.options = PTL_MD_OP_PUT;
 386                 request->rq_bulk_md.user_ptr = request;
 387                 request->rq_bulk_md.eventq = bulk_sink_eq;
 388
 389                 rc = PtlMDAttach(bulk_me_h, request->rq_bulk_md, PTL_UNLINK,
 390                                  &request->rq_bulk_md_h);
 391                 if (rc != PTL_OK) {
 392                         CERROR("PtlMDAttach failed: %d\n", rc);
 393                         BUG();
 394                         EXIT;
 395                         goto cleanup4;
 396                 }
 397         }
 398
 399         return ptl_send_buf(request, peer, request->rq_req_portal, 1);
 400
 401  cleanup4:
 402         PtlMEUnlink(bulk_me_h);
 403  cleanup3:
 404         PtlMDUnlink(request->rq_reply_md_h);
 405  cleanup2:
 406         PtlMEUnlink(me_h);
 407  cleanup:
 408         OBD_FREE(repbuf, request->rq_replen);
 409
 410         return rc;
 411 }
 412
 413 /* ptl_received_rpc() should be called by the sleeping process once
 414  * it finishes processing an event.  This ensures the ref count is
 415  * decremented and that the rpc ring buffer cycles properly.
 416  */
 417 int ptl_received_rpc(struct ptlrpc_service *service) {
 418         int rc, index;
 419
 420         index = service->srv_md_active;
 421         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 422                service->srv_ref_count[index]);
 423         service->srv_ref_count[index]--;
 424
 425         if ((service->srv_ref_count[index] <= 0) &&
 426             (service->srv_me_h[index] == 0)) {
 427
 428                 /* Replace the unlinked ME and MD */
 429                 rc = PtlMEInsert(service->srv_me_h[service->srv_me_tail],
 430                                  service->srv_id, 0, ~0, PTL_RETAIN,
 431                                  PTL_INS_AFTER, &(service->srv_me_h[index]));
 432                 CDEBUG(D_INFO, "Inserting new ME and MD in ring, rc %d\n", rc);
 433                 service->srv_me_tail = index;
 434                 service->srv_ref_count[index] = 0;
 435
 436                 if (rc != PTL_OK) {
 437                         CERROR("PtlMEInsert failed: %d\n", rc);
 438                         return rc;
 439                 }
 440
 441                 service->srv_md[index].start        = service->srv_buf[index];
 442                 service->srv_md[index].length       = service->srv_buf_size;
 443                 service->srv_md[index].threshold    = PTL_MD_THRESH_INF;
 444                 service->srv_md[index].options      = PTL_MD_OP_PUT;
 445                 service->srv_md[index].user_ptr     = service;
 446                 service->srv_md[index].eventq       = service->srv_eq_h;
 447
 448                 rc = PtlMDAttach(service->srv_me_h[index],
 449                                  service->srv_md[index],
 450                                  PTL_RETAIN, &(service->srv_md_h[index]));
 451
 452                 CDEBUG(D_INFO, "Attach MD in ring, rc %d\n", rc);
 453                 if (rc != PTL_OK) {
 454                         /* XXX cleanup */
 455                         BUG();
 456                         CERROR("PtlMDAttach failed: %d\n", rc);
 457                         return rc;
 458                 }
 459
 460                 service->srv_md_active =
 461                         NEXT_INDEX(index, service->srv_ring_length);
 462         }
 463
 464         return 0;
 465 }
 466
 467 int rpc_register_service(struct ptlrpc_service *service, char *uuid)
 468 {
 469         struct lustre_peer peer;
 470         int rc, i;
 471
 472         rc = kportal_uuid_to_peer(uuid, &peer);
 473         if (rc != 0) {
 474                 CERROR("Invalid uuid \"%s\"\n", uuid);
 475                 return -EINVAL;
 476         }
 477
 478         service->srv_ring_length = RPC_RING_LENGTH;
 479         service->srv_me_active = 0;
 480         service->srv_md_active = 0;
 481
 482         service->srv_id.addr_kind = PTL_ADDR_GID;
 483         service->srv_id.gid = PTL_ID_ANY;
 484         service->srv_id.rid = PTL_ID_ANY;
 485
 486         rc = PtlEQAlloc(peer.peer_ni, 128, server_request_callback,
 487                         service, &(service->srv_eq_h));
 488
 489         if (rc != PTL_OK) {
 490                 CERROR("PtlEQAlloc failed: %d\n", rc);
 491                 return rc;
 492         }
 493
 494         /* Attach the leading ME on which we build the ring */
 495         rc = PtlMEAttach(peer.peer_ni, service->srv_portal,
 496                          service->srv_id, 0, ~0, PTL_RETAIN,
 497                          &(service->srv_me_h[0]));
 498
 499         if (rc != PTL_OK) {
 500                 CERROR("PtlMEAttach failed: %d\n", rc);
 501                 return rc;
 502         }
 503
 504         for (i = 0; i < service->srv_ring_length; i++) {
 505                 OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
 506
 507                 if (service->srv_buf[i] == NULL) {
 508                         CERROR("no memory\n");
 509                         return -ENOMEM;
 510                 }
 511
 512                 /* Insert additional ME's to the ring */
 513                 if (i > 0) {
 514                         rc = PtlMEInsert(service->srv_me_h[i-1],
 515                                          service->srv_id, 0, ~0, PTL_RETAIN,
 516                                          PTL_INS_AFTER,&(service->srv_me_h[i]));
 517                         service->srv_me_tail = i;
 518
 519                         if (rc != PTL_OK) {
 520                                 CERROR("PtlMEInsert failed: %d\n", rc);
 521                                 return rc;
 522                         }
 523                 }
 524
 525                 service->srv_ref_count[i] = 0;
 526                 service->srv_md[i].start        = service->srv_buf[i];
 527                 service->srv_md[i].length        = service->srv_buf_size;
 528                 service->srv_md[i].threshold        = PTL_MD_THRESH_INF;
 529                 service->srv_md[i].options        = PTL_MD_OP_PUT;
 530                 service->srv_md[i].user_ptr        = service;
 531                 service->srv_md[i].eventq        = service->srv_eq_h;
 532
 533                 rc = PtlMDAttach(service->srv_me_h[i], service->srv_md[i],
 534                                  PTL_RETAIN, &(service->srv_md_h[i]));
 535
 536                 if (rc != PTL_OK) {
 537                         /* cleanup */
 538                         CERROR("PtlMDAttach failed: %d\n", rc);
 539                         return rc;
 540                 }
 541         }
 542
 543         return 0;
 544 }
 545
 546 int rpc_unregister_service(struct ptlrpc_service *service)
 547 {
 548         int rc, i;
 549
 550         for (i = 0; i < service->srv_ring_length; i++) {
 551                 rc = PtlMDUnlink(service->srv_md_h[i]);
 552                 if (rc)
 553                         CERROR("PtlMDUnlink failed: %d\n", rc);
 554
 555                 rc = PtlMEUnlink(service->srv_me_h[i]);
 556                 if (rc)
 557                         CERROR("PtlMEUnlink failed: %d\n", rc);
 558
 559                 OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 560         }
 561
 562         rc = PtlEQFree(service->srv_eq_h);
 563         if (rc)
 564                 CERROR("PtlEQFree failed: %d\n", rc);
 565
 566         return 0;
 567 }
 568
 569 static int req_init_portals(void)
 570 {
 571         int rc;
 572         const ptl_handle_ni_t *nip;
 573         ptl_handle_ni_t ni;
 574
 575         nip = inter_module_get_request(LUSTRE_NAL "_ni", LUSTRE_NAL);
 576         if (nip == NULL) {
 577                 CERROR("get_ni failed: is the NAL module loaded?\n");
 578                 return -EIO;
 579         }
 580         ni = *nip;
 581
 582         rc = PtlEQAlloc(ni, 128, sent_packet_callback, NULL, &sent_pkt_eq);
 583         if (rc != PTL_OK)
 584                 CERROR("PtlEQAlloc failed: %d\n", rc);
 585
 586         rc = PtlEQAlloc(ni, 128, rcvd_reply_callback, NULL, &rcvd_rep_eq);
 587         if (rc != PTL_OK)
 588                 CERROR("PtlEQAlloc failed: %d\n", rc);
 589
 590         rc = PtlEQAlloc(ni, 128, bulk_source_callback, NULL, &bulk_source_eq);
 591         if (rc != PTL_OK)
 592                 CERROR("PtlEQAlloc failed: %d\n", rc);
 593
 594         rc = PtlEQAlloc(ni, 128, bulk_sink_callback, NULL, &bulk_sink_eq);
 595         if (rc != PTL_OK)
 596                 CERROR("PtlEQAlloc failed: %d\n", rc);
 597
 598         return rc;
 599 }
 600
 601 static int __init ptlrpc_init(void)
 602 {
 603         return req_init_portals();
 604 }
 605
 606 static void __exit ptlrpc_exit(void)
 607 {
 608         PtlEQFree(sent_pkt_eq);
 609         PtlEQFree(rcvd_rep_eq);
 610         PtlEQFree(bulk_source_eq);
 611         PtlEQFree(bulk_sink_eq);
 612
 613         inter_module_put(LUSTRE_NAL "_ni");
 614
 615         return;
 616 }
 617
 618 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
 619 MODULE_DESCRIPTION("Lustre Request Processor v1.0");
 620 MODULE_LICENSE("GPL");
 621
 622 module_init(ptlrpc_init);
 623 module_exit(ptlrpc_exit);