lustre/ptlrpc/niobuf.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 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 DEBUG_SUBSYSTEM S_RPC
  24 #ifndef __KERNEL__
  25 #include <liblustre.h>
  26 #include <portals/lib-types.h>
  27 #endif
  28 #include <linux/obd_support.h>
  29 #include <linux/lustre_net.h>
  30 #include <linux/lustre_lib.h>
  31 #include <linux/obd.h>
  32
  33 static int ptl_send_buf(struct ptlrpc_request *request,
  34                         struct ptlrpc_connection *conn, int portal)
  35 {
  36         int rc;
  37         int rc2;
  38         ptl_process_id_t remote_id;
  39         ptl_handle_md_t md_h;
  40         ptl_ack_req_t ack_req;
  41
  42         LASSERT (portal != 0);
  43         LASSERT (conn != NULL);
  44         CDEBUG (D_INFO, "conn=%p ni %s nid "LPX64" on %s\n",
  45                 conn, conn->c_peer.peer_ni->pni_name,
  46                 conn->c_peer.peer_nid, conn->c_peer.peer_ni->pni_name);
  47
  48         request->rq_req_md.user_ptr = request;
  49
  50         switch (request->rq_type) {
  51         case PTL_RPC_MSG_REQUEST:
  52                 request->rq_reqmsg->type = request->rq_type;
  53                 request->rq_req_md.start = request->rq_reqmsg;
  54                 request->rq_req_md.length = request->rq_reqlen;
  55                 request->rq_req_md.eventq =
  56                         conn->c_peer.peer_ni->pni_request_out_eq_h;
  57                 LASSERT (!request->rq_want_ack);
  58                 break;
  59         case PTL_RPC_MSG_ERR:
  60         case PTL_RPC_MSG_REPLY:
  61                 request->rq_repmsg->type = request->rq_type;
  62                 request->rq_req_md.start = request->rq_repmsg;
  63                 request->rq_req_md.length = request->rq_replen;
  64                 request->rq_req_md.eventq =
  65                         conn->c_peer.peer_ni->pni_reply_out_eq_h;
  66                 break;
  67         default:
  68                 LBUG();
  69                 return -1; /* notreached */
  70         }
  71         if (request->rq_want_ack) {
  72                 request->rq_req_md.threshold = 2; /* SENT and ACK */
  73                 ack_req = PTL_ACK_REQ;
  74         } else {
  75                 request->rq_req_md.threshold = 1;
  76                 ack_req = PTL_NOACK_REQ;
  77         }
  78         request->rq_req_md.options = PTL_MD_OP_PUT;
  79         request->rq_req_md.user_ptr = request;
  80
  81         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
  82                 request->rq_req_md.options |= PTL_MD_ACK_DISABLE;
  83                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
  84         }
  85
  86         /* NB if the send fails, we back out of the send and return
  87          * failure; it's down to the caller to handle missing callbacks */
  88
  89         rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h, request->rq_req_md,
  90                        &md_h);
  91         if (rc != PTL_OK) {
  92                 CERROR("PtlMDBind failed: %d\n", rc);
  93                 LASSERT (rc == PTL_NOSPACE);
  94                 RETURN (-ENOMEM);
  95         }
  96         if (request->rq_type != PTL_RPC_MSG_REQUEST)
  97                 memcpy(&request->rq_reply_md_h, &md_h, sizeof(md_h));
  98
  99         remote_id.nid = conn->c_peer.peer_nid;
 100         remote_id.pid = 0;
 101
 102         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
 103                request->rq_req_md.length, portal, request->rq_xid);
 104
 105         rc = PtlPut(md_h, ack_req, remote_id, portal, 0, request->rq_xid, 0, 0);
 106         if (rc != PTL_OK) {
 107                 CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n",
 108                        remote_id.nid, portal, request->rq_xid, rc);
 109                 rc2 = PtlMDUnlink(md_h);
 110                 LASSERT (rc2 == PTL_OK);
 111                 RETURN ((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 112         }
 113
 114         return 0;
 115 }
 116
 117 static inline ptl_kiov_t *
 118 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
 119 {
 120         ptl_kiov_t *iov;
 121
 122         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov))
 123                 return (desc->bd_iov);
 124
 125         OBD_ALLOC (iov, desc->bd_page_count * sizeof (*iov));
 126         if (iov == NULL)
 127                 LBUG();
 128
 129         return (iov);
 130 }
 131
 132 static inline void
 133 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, ptl_kiov_t *iov)
 134 {
 135         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov))
 136                 return;
 137
 138         OBD_FREE (iov, desc->bd_page_count * sizeof (*iov));
 139 }
 140
 141 int ptlrpc_bulk_put(struct ptlrpc_bulk_desc *desc)
 142 {
 143         int rc;
 144         int rc2;
 145         struct ptlrpc_peer *peer;
 146         struct list_head *tmp, *next;
 147         ptl_process_id_t remote_id;
 148         ptl_kiov_t *iov;
 149         __u64 xid;
 150         ENTRY;
 151
 152         /* NB no locking required until desc is on the network */
 153         LASSERT (!desc->bd_network_rw);
 154         LASSERT (desc->bd_type == BULK_PUT_SOURCE);
 155         desc->bd_complete = 0;
 156
 157         iov = ptlrpc_get_bulk_iov (desc);
 158         if (iov == NULL)
 159                 RETURN (-ENOMEM);
 160
 161         peer = &desc->bd_export->exp_connection->c_peer;
 162
 163         desc->bd_md.start = iov;
 164         desc->bd_md.niov = 0;
 165         desc->bd_md.length = 0;
 166         desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_source_eq_h;
 167         desc->bd_md.threshold = 2; /* SENT and ACK */
 168         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV;
 169         desc->bd_md.user_ptr = desc;
 170
 171         desc->bd_callback_count = 2;
 172
 173         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 174                 struct ptlrpc_bulk_page *bulk;
 175                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 176
 177                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 178
 179                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 180                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 181                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 182
 183                 LASSERT (iov[desc->bd_md.niov].kiov_offset +
 184                          iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE);
 185                 desc->bd_md.niov++;
 186                 desc->bd_md.length += bulk->bp_buflen;
 187         }
 188
 189         /* NB total length may be 0 for a read past EOF, so we send a 0
 190          * length bulk, since the client expects a bulk event. */
 191         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 192
 193         rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md,
 194                        &desc->bd_md_h);
 195
 196         ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/
 197
 198         if (rc != PTL_OK) {
 199                 CERROR("PtlMDBind failed: %d\n", rc);
 200                 LASSERT (rc == PTL_NOSPACE);
 201                 RETURN(-ENOMEM);
 202         }
 203
 204         /* Client's bulk and reply matchbits are the same */
 205         xid = desc->bd_req->rq_xid;
 206         remote_id.nid = peer->peer_nid;
 207         remote_id.pid = 0;
 208
 209         CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d on %s "
 210                "nid "LPX64" pid %d xid "LPX64"\n",
 211                desc->bd_md.niov, desc->bd_md.length,
 212                desc->bd_portal, peer->peer_ni->pni_name,
 213                remote_id.nid, remote_id.pid, xid);
 214
 215         desc->bd_network_rw = 1;
 216         rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
 217                     desc->bd_portal, 0, xid, 0, 0);
 218         if (rc != PTL_OK) {
 219                 desc->bd_network_rw = 0;
 220                 CERROR("PtlPut("LPU64", %d, "LPX64") failed: %d\n",
 221                        remote_id.nid, desc->bd_portal, xid, rc);
 222                 rc2 = PtlMDUnlink(desc->bd_md_h);
 223                 LASSERT (rc2 == PTL_OK);
 224                 RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 225         }
 226
 227         RETURN(0);
 228 }
 229
 230 int ptlrpc_bulk_get(struct ptlrpc_bulk_desc *desc)
 231 {
 232         int rc;
 233         int rc2;
 234         struct ptlrpc_peer *peer;
 235         struct list_head *tmp, *next;
 236         ptl_process_id_t remote_id;
 237         ptl_kiov_t *iov;
 238         __u64 xid;
 239         ENTRY;
 240
 241         /* NB no locking required until desc is on the network */
 242         LASSERT (!desc->bd_network_rw);
 243         LASSERT (desc->bd_type == BULK_GET_SINK);
 244         desc->bd_complete = 0;
 245
 246         iov = ptlrpc_get_bulk_iov (desc);
 247         if (iov == NULL)
 248                 RETURN(-ENOMEM);
 249
 250         peer = &desc->bd_export->exp_connection->c_peer;
 251
 252         desc->bd_md.start = iov;
 253         desc->bd_md.niov = 0;
 254         desc->bd_md.length = 0;
 255         desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_sink_eq_h;
 256         desc->bd_md.threshold = 2; /* SENT and REPLY */
 257         desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV;
 258         desc->bd_md.user_ptr = desc;
 259
 260         desc->bd_callback_count = 2;
 261
 262         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 263                 struct ptlrpc_bulk_page *bulk;
 264                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 265
 266                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 267
 268                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 269                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 270                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 271
 272                 LASSERT (iov[desc->bd_md.niov].kiov_offset +
 273                          iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE);
 274                 desc->bd_md.niov++;
 275                 desc->bd_md.length += bulk->bp_buflen;
 276         }
 277
 278         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 279         LASSERT(desc->bd_md.niov != 0);
 280
 281         rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md, &desc->bd_md_h);
 282
 283         ptlrpc_put_bulk_iov(desc, iov); /*move down to reduce latency to send*/
 284
 285         if (rc != PTL_OK) {
 286                 CERROR("PtlMDBind failed: %d\n", rc);
 287                 LASSERT (rc == PTL_NOSPACE);
 288                 RETURN(-ENOMEM);
 289         }
 290
 291         /* Client's bulk and reply matchbits are the same */
 292         xid = desc->bd_req->rq_xid;
 293         remote_id.nid = desc->bd_export->exp_connection->c_peer.peer_nid;
 294         remote_id.pid = 0;
 295
 296         CDEBUG(D_NET, "Fetching %u pages %u bytes from portal %d on %s "
 297                "nid "LPX64" pid %d xid "LPX64"\n",
 298                desc->bd_md.niov, desc->bd_md.length, desc->bd_portal,
 299                peer->peer_ni->pni_name, remote_id.nid, remote_id.pid,
 300                xid);
 301
 302         desc->bd_network_rw = 1;
 303         rc = PtlGet(desc->bd_md_h, remote_id, desc->bd_portal, 0,
 304                     xid, 0);
 305         if (rc != PTL_OK) {
 306                 desc->bd_network_rw = 0;
 307                 CERROR("PtlGet("LPU64", %d, "LPX64") failed: %d\n",
 308                        remote_id.nid, desc->bd_portal, xid, rc);
 309                 rc2 = PtlMDUnlink(desc->bd_md_h);
 310                 LASSERT (rc2 == PTL_OK);
 311                 RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 312         }
 313
 314         RETURN(0);
 315 }
 316
 317 void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
 318 {
 319         /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 320          * serialises with completion callback) */
 321         unsigned long      flags;
 322         struct l_wait_info lwi;
 323         int                callback_count;
 324         int                rc;
 325
 326         LASSERT (!in_interrupt ());             /* might sleep */
 327
 328         /* NB. server-side bulk gets 2 events, so we have to keep trying to
 329          * unlink the MD until all callbacks have happened, or
 330          * PtlMDUnlink() returns OK or INVALID */
 331  again:
 332         spin_lock_irqsave (&desc->bd_lock, flags);
 333         if (!desc->bd_network_rw) {
 334                 /* completed or never even registered. NB holding bd_lock
 335                  * guarantees callback has completed if it ran. */
 336                 spin_unlock_irqrestore (&desc->bd_lock, flags);
 337                 return;
 338         }
 339
 340         /* sample callback count while we have the lock */
 341         callback_count = desc->bd_callback_count;
 342         spin_unlock_irqrestore (&desc->bd_lock, flags);
 343
 344         rc = PtlMDUnlink (desc->bd_md_h);
 345         switch (rc) {
 346         default:
 347                 CERROR("PtlMDUnlink returned %d\n", rc);
 348                 LBUG ();
 349         case PTL_OK:                    /* Won the race with the network */
 350                 LASSERT (!desc->bd_complete); /* Not all callbacks ran */
 351                 desc->bd_network_rw = 0;
 352                 return;
 353
 354         case PTL_MD_INUSE:              /* MD is being accessed right now */
 355                 for (;;) {
 356                         /* Network access will complete in finite time but the
 357                          * timeout lets us CERROR for visibility */
 358                         lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL);
 359                         rc = l_wait_event(desc->bd_waitq,
 360                                           desc->bd_callback_count !=
 361                                           callback_count, &lwi);
 362                         if (rc == -ETIMEDOUT) {
 363                                 CERROR("Unexpectedly long timeout: desc %p\n",
 364                                        desc);
 365                                 continue;
 366                         }
 367                         LASSERT (rc == 0);
 368                         break;
 369                 }
 370                 /* go back and try again... */
 371                 goto again;
 372
 373         case PTL_INV_MD:            /* Lost the race with completion */
 374                 LASSERT (desc->bd_complete);    /* Callbacks all ran */
 375                 LASSERT (!desc->bd_network_rw);
 376                 return;
 377         }
 378 }
 379
 380 int ptlrpc_register_bulk (struct ptlrpc_request *req)
 381 {
 382         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 383         struct ptlrpc_peer *peer;
 384         struct list_head *tmp, *next;
 385         int rc;
 386         int rc2;
 387         ptl_kiov_t *iov;
 388         ptl_process_id_t source_id;
 389         ENTRY;
 390
 391         /* NB no locking required until desc is on the network */
 392         LASSERT (!desc->bd_network_rw);
 393         LASSERT (desc->bd_page_count <= PTL_MD_MAX_IOV);
 394         LASSERT (desc->bd_req != NULL);
 395         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 396                  desc->bd_type == BULK_GET_SOURCE);
 397
 398         desc->bd_complete = 0;
 399
 400         iov = ptlrpc_get_bulk_iov (desc);
 401         if (iov == NULL)
 402                 return (-ENOMEM);
 403
 404         peer = &desc->bd_import->imp_connection->c_peer;
 405
 406         desc->bd_md.start = iov;
 407         desc->bd_md.niov = 0;
 408         desc->bd_md.length = 0;
 409         desc->bd_md.threshold = 1;
 410         desc->bd_md.user_ptr = desc;
 411
 412         if (desc->bd_type == BULK_GET_SOURCE) {
 413                 desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV;
 414                 desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_source_eq_h;
 415         } else {
 416                 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV;
 417                 desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_sink_eq_h;
 418         }
 419
 420         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 421                 struct ptlrpc_bulk_page *bulk;
 422                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 423
 424                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 425
 426                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 427                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 428                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 429
 430                 LASSERT (bulk->bp_pageoffset + bulk->bp_buflen <= PAGE_SIZE);
 431                 desc->bd_md.niov++;
 432                 desc->bd_md.length += bulk->bp_buflen;
 433         }
 434
 435         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 436         LASSERT(desc->bd_md.niov != 0);
 437
 438         /* XXX Registering the same xid on retried bulk makes my head
 439          * explode trying to understand how the original request's bulk
 440          * might interfere with the retried request -eeb */
 441         LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
 442         desc->bd_registered = 1;
 443         desc->bd_last_xid = desc->bd_last_xid;
 444
 445         source_id.nid = desc->bd_import->imp_connection->c_peer.peer_nid;
 446         source_id.pid = PTL_PID_ANY;
 447
 448         rc = PtlMEAttach(peer->peer_ni->pni_ni_h,
 449                          desc->bd_portal, source_id, req->rq_xid, 0,
 450                          PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
 451
 452         if (rc != PTL_OK) {
 453                 CERROR("PtlMEAttach failed: %d\n", rc);
 454                 LASSERT (rc == PTL_NOSPACE);
 455                 GOTO(out, rc = -ENOMEM);
 456         }
 457
 458         /* About to let the network at it... */
 459         desc->bd_network_rw = 1;
 460         rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
 461                          &desc->bd_md_h);
 462         if (rc != PTL_OK) {
 463                 CERROR("PtlMDAttach failed: %d\n", rc);
 464                 LASSERT (rc == PTL_NOSPACE);
 465                 desc->bd_network_rw = 0;
 466                 rc2 = PtlMEUnlink (desc->bd_me_h);
 467                 LASSERT (rc2 == PTL_OK);
 468                 GOTO(out, rc = -ENOMEM);
 469         }
 470         rc = 0;
 471
 472         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
 473                "portal %u on %s\n",
 474                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 475                desc->bd_md.niov, desc->bd_md.length,
 476                req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
 477
 478  out:
 479         ptlrpc_put_bulk_iov (desc, iov);
 480         RETURN(rc);
 481 }
 482
 483 void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
 484 {
 485         /* Disconnect a bulk desc from the network. Idempotent. Not
 486          * thread-safe (i.e. only interlocks with completion callback). */
 487         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 488         wait_queue_head_t       *wq;
 489         unsigned long            flags;
 490         struct l_wait_info       lwi;
 491         int                      rc;
 492
 493         LASSERT (!in_interrupt ());             /* might sleep */
 494
 495         spin_lock_irqsave (&desc->bd_lock, flags);
 496         if (!desc->bd_network_rw) {     /* completed or never even registered */
 497                 spin_unlock_irqrestore (&desc->bd_lock, flags);
 498                 return;
 499         }
 500         spin_unlock_irqrestore (&desc->bd_lock, flags);
 501
 502         LASSERT (desc->bd_req == req);     /* NB bd_req NULL until registered */
 503
 504         /* NB...
 505          * 1. If the MD unlink is successful, the ME gets unlinked too.
 506          * 2. Since client-side bulk only gets a single event and a
 507          * .. threshold of 1.  If the MD was inuse at the first link
 508          * .. attempt, the callback is due any minute, and the MD/ME will
 509          * .. unlink themselves.
 510          */
 511         rc = PtlMDUnlink (desc->bd_md_h);
 512         switch (rc) {
 513         default:
 514                 CERROR("PtlMDUnlink returned %d\n", rc);
 515                 LBUG ();
 516         case PTL_OK:                          /* Won the race with completion */
 517                 LASSERT (!desc->bd_complete);   /* Callback hasn't happened */
 518                 desc->bd_network_rw = 0;
 519                 return;
 520         case PTL_MD_INUSE:                  /* MD is being accessed right now */
 521                 for (;;) {
 522                         /* Network access will complete in finite time but the
 523                          * timeout lets us CERROR for visibility */
 524                         if (desc->bd_req->rq_set != NULL)
 525                                 wq = &req->rq_set->set_waitq;
 526                         else
 527                                 wq = &req->rq_wait_for_rep;
 528                         lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL);
 529                         rc = l_wait_event(*wq, ptlrpc_bulk_complete(desc), &lwi);
 530                         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 531                         if (rc == 0)
 532                                 break;
 533                         CERROR ("Unexpectedly long timeout: desc %p\n", desc);
 534                         LBUG();
 535                 }
 536                 /* Fall through */
 537         case PTL_INV_MD:                     /* Lost the race with completion */
 538                 LASSERT (desc->bd_complete);/* Callback has run to completion */
 539                 LASSERT (!desc->bd_network_rw);
 540                 return;
 541         }
 542 }
 543
 544 int ptlrpc_reply(struct ptlrpc_request *req)
 545 {
 546         unsigned long flags;
 547         int rc;
 548
 549         /* We must already have a reply buffer (only ptlrpc_error() may be
 550          * called without one).  We must also have a request buffer which
 551          * is either the actual (swabbed) incoming request, or a saved copy
 552          * if this is a req saved in target_queue_final_reply(). */
 553         LASSERT (req->rq_repmsg != NULL);
 554         LASSERT (req->rq_reqmsg != NULL);
 555
 556         /* FIXME: we need to increment the count of handled events */
 557         if (req->rq_type != PTL_RPC_MSG_ERR)
 558                 req->rq_type = PTL_RPC_MSG_REPLY;
 559
 560         req->rq_repmsg->status = req->rq_status;
 561         req->rq_repmsg->opc = req->rq_reqmsg->opc;
 562
 563         init_waitqueue_head(&req->rq_wait_for_rep);
 564         rc = ptl_send_buf(req, req->rq_connection, req->rq_svc->srv_rep_portal);
 565         if (rc != 0) {
 566                 /* Do what the callback handler would have done */
 567                 OBD_FREE (req->rq_repmsg, req->rq_replen);
 568
 569                 spin_lock_irqsave (&req->rq_lock, flags);
 570                 req->rq_want_ack = 0;
 571                 spin_unlock_irqrestore (&req->rq_lock, flags);
 572         }
 573         return rc;
 574 }
 575
 576 int ptlrpc_error(struct ptlrpc_request *req)
 577 {
 578         int rc;
 579         ENTRY;
 580
 581         if (!req->rq_repmsg) {
 582                 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen,
 583                                      &req->rq_repmsg);
 584                 if (rc)
 585                         RETURN(rc);
 586         }
 587
 588
 589         req->rq_type = PTL_RPC_MSG_ERR;
 590
 591         rc = ptlrpc_reply(req);
 592         RETURN(rc);
 593 }
 594
 595 int ptl_send_rpc(struct ptlrpc_request *request)
 596 {
 597         int rc;
 598         int rc2;
 599         unsigned long flags;
 600         ptl_process_id_t source_id;
 601         ptl_handle_me_t  reply_me_h;
 602         ENTRY;
 603
 604         LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 605
 606         /* If this is a re-transmit, we're required to have disengaged
 607          * cleanly from the previous attempt */
 608         LASSERT (!request->rq_receiving_reply);
 609
 610         if (request->rq_bulk != NULL) {
 611                 rc = ptlrpc_register_bulk (request);
 612                 if (rc != 0)
 613                         RETURN(rc);
 614         }
 615
 616         request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
 617
 618         source_id.nid = request->rq_connection->c_peer.peer_nid;
 619         source_id.pid = PTL_PID_ANY;
 620
 621         LASSERT (request->rq_replen != 0);
 622         OBD_ALLOC(request->rq_repmsg, request->rq_replen);
 623         if (request->rq_repmsg == NULL) {
 624                 LBUG();
 625                 RETURN(-ENOMEM);
 626         }
 627
 628         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni->pni_ni_h,
 629                          request->rq_reply_portal, /* XXX FIXME bug 249 */
 630                          source_id, request->rq_xid, 0, PTL_UNLINK,
 631                          PTL_INS_AFTER, &reply_me_h);
 632         if (rc != PTL_OK) {
 633                 CERROR("PtlMEAttach failed: %d\n", rc);
 634                 LASSERT (rc == PTL_NOSPACE);
 635                 LBUG();
 636                 GOTO(cleanup, rc = -ENOMEM);
 637         }
 638
 639         request->rq_reply_md.start = request->rq_repmsg;
 640         request->rq_reply_md.length = request->rq_replen;
 641         request->rq_reply_md.threshold = 1;
 642         request->rq_reply_md.options = PTL_MD_OP_PUT;
 643         request->rq_reply_md.user_ptr = request;
 644         request->rq_reply_md.eventq =
 645                 request->rq_connection->c_peer.peer_ni->pni_reply_in_eq_h;
 646
 647         rc = PtlMDAttach(reply_me_h, request->rq_reply_md,
 648                          PTL_UNLINK, &request->rq_reply_md_h);
 649         if (rc != PTL_OK) {
 650                 CERROR("PtlMDAttach failed: %d\n", rc);
 651                 LASSERT (rc == PTL_NOSPACE);
 652                 LBUG();
 653                 GOTO(cleanup2, rc -ENOMEM);
 654         }
 655
 656         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 657                ", portal %u on %s\n",
 658                request->rq_replen, request->rq_xid,
 659                request->rq_reply_portal,
 660                request->rq_connection->c_peer.peer_ni->pni_name);
 661
 662         ptlrpc_request_addref(request);        /* 1 ref for the SENT callback */
 663
 664         spin_lock_irqsave (&request->rq_lock, flags);
 665         request->rq_receiving_reply = 1;
 666         /* Clear any flags that may be present from previous sends. */
 667         request->rq_replied = 0;
 668         request->rq_err = 0;
 669         request->rq_timedout = 0;
 670         request->rq_resend = 0;
 671         request->rq_restart = 0;
 672         spin_unlock_irqrestore (&request->rq_lock, flags);
 673
 674         request->rq_sent = LTIME_S(CURRENT_TIME);
 675         rc = ptl_send_buf(request, request->rq_connection,
 676                           request->rq_request_portal);
 677         if (rc == 0)
 678                 RETURN(rc);
 679
 680         spin_lock_irqsave (&request->rq_lock, flags);
 681         request->rq_receiving_reply = 0;
 682         spin_unlock_irqrestore (&request->rq_lock, flags);
 683         ptlrpc_req_finished (request);          /* drop callback ref */
 684  cleanup2:
 685         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 686          * nobody apart from the PUT's target has the right nid+XID to
 687          * access the reply buffer. */
 688         rc2 = PtlMEUnlink(reply_me_h);
 689         LASSERT (rc2 == PTL_OK);
 690  cleanup:
 691         OBD_FREE(request->rq_repmsg, request->rq_replen);
 692         request->rq_repmsg = NULL;
 693         return rc;
 694 }
 695
 696 void ptlrpc_link_svc_me(struct ptlrpc_request_buffer_desc *rqbd)
 697 {
 698         struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
 699         struct ptlrpc_service *service = srv_ni->sni_service;
 700         static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
 701         int rc;
 702         ptl_md_t dummy;
 703         ptl_handle_md_t md_h;
 704
 705         LASSERT(atomic_read(&rqbd->rqbd_refcount) == 0);
 706
 707         CDEBUG(D_NET, "PtlMEAttach: portal %d on %s h %lx."LPX64"\n",
 708                service->srv_req_portal, srv_ni->sni_ni->pni_name,
 709                srv_ni->sni_ni->pni_ni_h.nal_idx,
 710                srv_ni->sni_ni->pni_ni_h.cookie);
 711
 712         /* Attach the leading ME on which we build the ring */
 713         rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
 714                          match_id, 0, ~0,
 715                          PTL_UNLINK, PTL_INS_AFTER, &rqbd->rqbd_me_h);
 716         if (rc != PTL_OK) {
 717                 CERROR("PtlMEAttach failed: %d\n", rc);
 718                 /* BUG 1191 */
 719                 LBUG();
 720         }
 721
 722         dummy.start      = rqbd->rqbd_buffer;
 723         dummy.length     = service->srv_buf_size;
 724         dummy.max_size   = service->srv_max_req_size;
 725         dummy.threshold  = PTL_MD_THRESH_INF;
 726         dummy.options    = PTL_MD_OP_PUT | PTL_MD_MAX_SIZE | PTL_MD_AUTO_UNLINK;
 727         dummy.user_ptr   = rqbd;
 728         dummy.eventq     = srv_ni->sni_eq_h;
 729
 730         atomic_inc(&srv_ni->sni_nrqbds_receiving);
 731         atomic_set(&rqbd->rqbd_refcount, 1);   /* 1 ref for portals */
 732
 733         rc = PtlMDAttach(rqbd->rqbd_me_h, dummy, PTL_UNLINK, &md_h);
 734         if (rc != PTL_OK) {
 735                 CERROR("PtlMDAttach failed: %d\n", rc);
 736                 LASSERT (rc == PTL_NOSPACE);
 737                 LBUG();
 738                 /* BUG 1191 */
 739                 PtlMEUnlink (rqbd->rqbd_me_h);
 740                 atomic_set(&rqbd->rqbd_refcount, 0);
 741                 atomic_dec(&srv_ni->sni_nrqbds_receiving);
 742         }
 743 }