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 #include "ptlrpc_internal.h"
  33
  34 static int ptl_send_buf(struct ptlrpc_request *request,
  35                         struct ptlrpc_connection *conn, int portal)
  36 {
  37         int rc;
  38         int rc2;
  39         ptl_process_id_t remote_id;
  40         ptl_handle_md_t md_h;
  41         ptl_ack_req_t ack_req;
  42
  43         LASSERT (portal != 0);
  44         LASSERT (conn != NULL);
  45         CDEBUG (D_INFO, "conn=%p ni %s nid "LPX64" on %s\n",
  46                 conn, conn->c_peer.peer_ni->pni_name,
  47                 conn->c_peer.peer_nid, conn->c_peer.peer_ni->pni_name);
  48
  49         request->rq_req_md.user_ptr = request;
  50
  51         switch (request->rq_type) {
  52         case PTL_RPC_MSG_REQUEST:
  53                 request->rq_reqmsg->type = request->rq_type;
  54                 request->rq_req_md.start = request->rq_reqmsg;
  55                 request->rq_req_md.length = request->rq_reqlen;
  56                 request->rq_req_md.eventq =
  57                         conn->c_peer.peer_ni->pni_request_out_eq_h;
  58                 LASSERT (!request->rq_want_ack);
  59                 break;
  60         case PTL_RPC_MSG_ERR:
  61         case PTL_RPC_MSG_REPLY:
  62                 request->rq_repmsg->type = request->rq_type;
  63                 request->rq_req_md.start = request->rq_repmsg;
  64                 request->rq_req_md.length = request->rq_replen;
  65                 request->rq_req_md.eventq =
  66                         conn->c_peer.peer_ni->pni_reply_out_eq_h;
  67                 break;
  68         default:
  69                 LBUG();
  70                 return -1; /* notreached */
  71         }
  72         if (request->rq_want_ack) {
  73                 request->rq_req_md.threshold = 2; /* SENT and ACK */
  74                 ack_req = PTL_ACK_REQ;
  75         } else {
  76                 request->rq_req_md.threshold = 1;
  77                 ack_req = PTL_NOACK_REQ;
  78         }
  79         request->rq_req_md.options = PTL_MD_OP_PUT;
  80         request->rq_req_md.user_ptr = request;
  81
  82         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
  83                 request->rq_req_md.options |= PTL_MD_ACK_DISABLE;
  84                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
  85         }
  86
  87         /* NB if the send fails, we back out of the send and return
  88          * failure; it's down to the caller to handle missing callbacks */
  89
  90         rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h, request->rq_req_md,
  91                        &md_h);
  92         if (rc != PTL_OK) {
  93                 CERROR("PtlMDBind failed: %d\n", rc);
  94                 LASSERT (rc == PTL_NOSPACE);
  95                 RETURN (-ENOMEM);
  96         }
  97         if (request->rq_type != PTL_RPC_MSG_REQUEST)
  98                 memcpy(&request->rq_reply_md_h, &md_h, sizeof(md_h));
  99
 100         remote_id.nid = conn->c_peer.peer_nid;
 101         remote_id.pid = 0;
 102
 103         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
 104                request->rq_req_md.length, portal, request->rq_xid);
 105
 106         rc = PtlPut(md_h, ack_req, remote_id, portal, 0, request->rq_xid, 0, 0);
 107         if (rc != PTL_OK) {
 108                 CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n",
 109                        remote_id.nid, portal, request->rq_xid, rc);
 110                 rc2 = PtlMDUnlink(md_h);
 111                 LASSERT (rc2 == PTL_OK);
 112                 RETURN ((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 113         }
 114
 115         return 0;
 116 }
 117
 118 static inline ptl_kiov_t *
 119 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
 120 {
 121         ptl_kiov_t *iov;
 122
 123         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov))
 124                 return (desc->bd_iov);
 125
 126         OBD_ALLOC (iov, desc->bd_page_count * sizeof (*iov));
 127         if (iov == NULL)
 128                 LBUG();
 129
 130         return (iov);
 131 }
 132
 133 static inline void
 134 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, ptl_kiov_t *iov)
 135 {
 136         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov))
 137                 return;
 138
 139         OBD_FREE (iov, desc->bd_page_count * sizeof (*iov));
 140 }
 141
 142 int ptlrpc_bulk_put(struct ptlrpc_bulk_desc *desc)
 143 {
 144         int rc;
 145         int rc2;
 146         struct ptlrpc_peer *peer;
 147         struct list_head *tmp, *next;
 148         ptl_process_id_t remote_id;
 149         ptl_kiov_t *iov;
 150         __u64 xid;
 151         ENTRY;
 152
 153         /* NB no locking required until desc is on the network */
 154         LASSERT (!desc->bd_network_rw);
 155         LASSERT (desc->bd_type == BULK_PUT_SOURCE);
 156         desc->bd_complete = 0;
 157
 158         iov = ptlrpc_get_bulk_iov (desc);
 159         if (iov == NULL)
 160                 RETURN (-ENOMEM);
 161
 162         peer = &desc->bd_export->exp_connection->c_peer;
 163
 164         desc->bd_md.start = iov;
 165         desc->bd_md.niov = 0;
 166         desc->bd_md.length = 0;
 167         desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_source_eq_h;
 168         desc->bd_md.threshold = 2; /* SENT and ACK */
 169         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV;
 170         desc->bd_md.user_ptr = desc;
 171
 172         desc->bd_callback_count = 2;
 173
 174         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 175                 struct ptlrpc_bulk_page *bulk;
 176                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 177
 178                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 179
 180                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 181                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 182                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 183
 184                 LASSERT (iov[desc->bd_md.niov].kiov_offset +
 185                          iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE);
 186                 desc->bd_md.niov++;
 187                 desc->bd_md.length += bulk->bp_buflen;
 188         }
 189
 190         /* NB total length may be 0 for a read past EOF, so we send a 0
 191          * length bulk, since the client expects a bulk event. */
 192         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 193
 194         rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md,
 195                        &desc->bd_md_h);
 196
 197         ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/
 198
 199         if (rc != PTL_OK) {
 200                 CERROR("PtlMDBind failed: %d\n", rc);
 201                 LASSERT (rc == PTL_NOSPACE);
 202                 RETURN(-ENOMEM);
 203         }
 204
 205         /* Client's bulk and reply matchbits are the same */
 206         xid = desc->bd_req->rq_xid;
 207         remote_id.nid = peer->peer_nid;
 208         remote_id.pid = 0;
 209
 210         CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d on %s "
 211                "nid "LPX64" pid %d xid "LPX64"\n",
 212                desc->bd_md.niov, desc->bd_md.length,
 213                desc->bd_portal, peer->peer_ni->pni_name,
 214                remote_id.nid, remote_id.pid, xid);
 215
 216         desc->bd_network_rw = 1;
 217         rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
 218                     desc->bd_portal, 0, xid, 0, 0);
 219         if (rc != PTL_OK) {
 220                 desc->bd_network_rw = 0;
 221                 CERROR("PtlPut("LPU64", %d, "LPX64") failed: %d\n",
 222                        remote_id.nid, desc->bd_portal, xid, rc);
 223                 rc2 = PtlMDUnlink(desc->bd_md_h);
 224                 LASSERT (rc2 == PTL_OK);
 225                 RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 226         }
 227
 228         RETURN(0);
 229 }
 230
 231 int ptlrpc_bulk_get(struct ptlrpc_bulk_desc *desc)
 232 {
 233         int rc;
 234         int rc2;
 235         struct ptlrpc_peer *peer;
 236         struct list_head *tmp, *next;
 237         ptl_process_id_t remote_id;
 238         ptl_kiov_t *iov;
 239         __u64 xid;
 240         ENTRY;
 241
 242         /* NB no locking required until desc is on the network */
 243         LASSERT (!desc->bd_network_rw);
 244         LASSERT (desc->bd_type == BULK_GET_SINK);
 245         desc->bd_complete = 0;
 246
 247         iov = ptlrpc_get_bulk_iov (desc);
 248         if (iov == NULL)
 249                 RETURN(-ENOMEM);
 250
 251         peer = &desc->bd_export->exp_connection->c_peer;
 252
 253         desc->bd_md.start = iov;
 254         desc->bd_md.niov = 0;
 255         desc->bd_md.length = 0;
 256         desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_sink_eq_h;
 257         desc->bd_md.threshold = 2; /* SENT and REPLY */
 258         desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV;
 259         desc->bd_md.user_ptr = desc;
 260
 261         desc->bd_callback_count = 2;
 262
 263         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 264                 struct ptlrpc_bulk_page *bulk;
 265                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 266
 267                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 268
 269                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 270                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 271                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 272
 273                 LASSERT (iov[desc->bd_md.niov].kiov_offset +
 274                          iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE);
 275                 desc->bd_md.niov++;
 276                 desc->bd_md.length += bulk->bp_buflen;
 277         }
 278
 279         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 280         LASSERT(desc->bd_md.niov != 0);
 281
 282         rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md, &desc->bd_md_h);
 283
 284         ptlrpc_put_bulk_iov(desc, iov); /*move down to reduce latency to send*/
 285
 286         if (rc != PTL_OK) {
 287                 CERROR("PtlMDBind failed: %d\n", rc);
 288                 LASSERT (rc == PTL_NOSPACE);
 289                 RETURN(-ENOMEM);
 290         }
 291
 292         /* Client's bulk and reply matchbits are the same */
 293         xid = desc->bd_req->rq_xid;
 294         remote_id.nid = desc->bd_export->exp_connection->c_peer.peer_nid;
 295         remote_id.pid = 0;
 296
 297         CDEBUG(D_NET, "Fetching %u pages %u bytes from portal %d on %s "
 298                "nid "LPX64" pid %d xid "LPX64"\n",
 299                desc->bd_md.niov, desc->bd_md.length, desc->bd_portal,
 300                peer->peer_ni->pni_name, remote_id.nid, remote_id.pid,
 301                xid);
 302
 303         desc->bd_network_rw = 1;
 304         rc = PtlGet(desc->bd_md_h, remote_id, desc->bd_portal, 0,
 305                     xid, 0);
 306         if (rc != PTL_OK) {
 307                 desc->bd_network_rw = 0;
 308                 CERROR("PtlGet("LPU64", %d, "LPX64") failed: %d\n",
 309                        remote_id.nid, desc->bd_portal, xid, rc);
 310                 rc2 = PtlMDUnlink(desc->bd_md_h);
 311                 LASSERT (rc2 == PTL_OK);
 312                 RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 313         }
 314
 315         RETURN(0);
 316 }
 317
 318 void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
 319 {
 320         /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 321          * serialises with completion callback) */
 322         unsigned long      flags;
 323         struct l_wait_info lwi;
 324         int                callback_count;
 325         int                rc;
 326
 327         LASSERT (!in_interrupt ());             /* might sleep */
 328
 329         /* NB. server-side bulk gets 2 events, so we have to keep trying to
 330          * unlink the MD until all callbacks have happened, or
 331          * PtlMDUnlink() returns OK or INVALID */
 332  again:
 333         spin_lock_irqsave (&desc->bd_lock, flags);
 334         if (!desc->bd_network_rw) {
 335                 /* completed or never even registered. NB holding bd_lock
 336                  * guarantees callback has completed if it ran. */
 337                 spin_unlock_irqrestore (&desc->bd_lock, flags);
 338                 return;
 339         }
 340
 341         /* sample callback count while we have the lock */
 342         callback_count = desc->bd_callback_count;
 343         spin_unlock_irqrestore (&desc->bd_lock, flags);
 344
 345         rc = PtlMDUnlink (desc->bd_md_h);
 346         switch (rc) {
 347         default:
 348                 CERROR("PtlMDUnlink returned %d\n", rc);
 349                 LBUG ();
 350         case PTL_OK:                    /* Won the race with the network */
 351                 LASSERT (!desc->bd_complete); /* Not all callbacks ran */
 352                 desc->bd_network_rw = 0;
 353                 return;
 354
 355         case PTL_MD_INUSE:              /* MD is being accessed right now */
 356                 for (;;) {
 357                         /* Network access will complete in finite time but the
 358                          * timeout lets us CERROR for visibility */
 359                         lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL);
 360                         rc = l_wait_event(desc->bd_waitq,
 361                                           desc->bd_callback_count !=
 362                                           callback_count, &lwi);
 363                         if (rc == -ETIMEDOUT) {
 364                                 CERROR("Unexpectedly long timeout: desc %p\n",
 365                                        desc);
 366                                 continue;
 367                         }
 368                         LASSERT (rc == 0);
 369                         break;
 370                 }
 371                 /* go back and try again... */
 372                 goto again;
 373
 374         case PTL_INV_MD:            /* Lost the race with completion */
 375                 LASSERT (desc->bd_complete);    /* Callbacks all ran */
 376                 LASSERT (!desc->bd_network_rw);
 377                 return;
 378         }
 379 }
 380
 381 int ptlrpc_register_bulk (struct ptlrpc_request *req)
 382 {
 383         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 384         struct ptlrpc_peer *peer;
 385         struct list_head *tmp, *next;
 386         int rc;
 387         int rc2;
 388         ptl_kiov_t *iov;
 389         ptl_process_id_t source_id;
 390         ENTRY;
 391
 392         /* NB no locking required until desc is on the network */
 393         LASSERT (!desc->bd_network_rw);
 394         LASSERT (desc->bd_page_count <= PTL_MD_MAX_IOV);
 395         LASSERT (desc->bd_req != NULL);
 396         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 397                  desc->bd_type == BULK_GET_SOURCE);
 398
 399         desc->bd_complete = 0;
 400
 401         iov = ptlrpc_get_bulk_iov (desc);
 402         if (iov == NULL)
 403                 return (-ENOMEM);
 404
 405         peer = &desc->bd_import->imp_connection->c_peer;
 406
 407         desc->bd_md.start = iov;
 408         desc->bd_md.niov = 0;
 409         desc->bd_md.length = 0;
 410         desc->bd_md.threshold = 1;
 411         desc->bd_md.user_ptr = desc;
 412
 413         if (desc->bd_type == BULK_GET_SOURCE) {
 414                 desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV;
 415                 desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_source_eq_h;
 416         } else {
 417                 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV;
 418                 desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_sink_eq_h;
 419         }
 420
 421         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 422                 struct ptlrpc_bulk_page *bulk;
 423                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 424
 425                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 426
 427                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 428                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 429                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 430
 431                 LASSERT (bulk->bp_pageoffset + bulk->bp_buflen <= PAGE_SIZE);
 432                 desc->bd_md.niov++;
 433                 desc->bd_md.length += bulk->bp_buflen;
 434         }
 435
 436         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 437         LASSERT(desc->bd_md.niov != 0);
 438
 439         /* XXX Registering the same xid on retried bulk makes my head
 440          * explode trying to understand how the original request's bulk
 441          * might interfere with the retried request -eeb */
 442         LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
 443         desc->bd_registered = 1;
 444         desc->bd_last_xid = desc->bd_last_xid;
 445
 446         source_id.nid = desc->bd_import->imp_connection->c_peer.peer_nid;
 447         source_id.pid = PTL_PID_ANY;
 448
 449         rc = PtlMEAttach(peer->peer_ni->pni_ni_h,
 450                          desc->bd_portal, source_id, req->rq_xid, 0,
 451                          PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
 452
 453         if (rc != PTL_OK) {
 454                 CERROR("PtlMEAttach failed: %d\n", rc);
 455                 LASSERT (rc == PTL_NOSPACE);
 456                 GOTO(out, rc = -ENOMEM);
 457         }
 458
 459         /* About to let the network at it... */
 460         desc->bd_network_rw = 1;
 461         rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
 462                          &desc->bd_md_h);
 463         if (rc != PTL_OK) {
 464                 CERROR("PtlMDAttach failed: %d\n", rc);
 465                 LASSERT (rc == PTL_NOSPACE);
 466                 desc->bd_network_rw = 0;
 467                 rc2 = PtlMEUnlink (desc->bd_me_h);
 468                 LASSERT (rc2 == PTL_OK);
 469                 GOTO(out, rc = -ENOMEM);
 470         }
 471         rc = 0;
 472
 473         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
 474                "portal %u on %s\n",
 475                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 476                desc->bd_md.niov, desc->bd_md.length,
 477                req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
 478
 479  out:
 480         ptlrpc_put_bulk_iov (desc, iov);
 481         RETURN(rc);
 482 }
 483
 484 void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
 485 {
 486         /* Disconnect a bulk desc from the network. Idempotent. Not
 487          * thread-safe (i.e. only interlocks with completion callback). */
 488         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 489         wait_queue_head_t       *wq;
 490         unsigned long            flags;
 491         struct l_wait_info       lwi;
 492         int                      rc;
 493
 494         LASSERT (!in_interrupt ());             /* might sleep */
 495
 496         spin_lock_irqsave (&desc->bd_lock, flags);
 497         if (!desc->bd_network_rw) {     /* completed or never even registered */
 498                 spin_unlock_irqrestore (&desc->bd_lock, flags);
 499                 return;
 500         }
 501         spin_unlock_irqrestore (&desc->bd_lock, flags);
 502
 503         LASSERT (desc->bd_req == req);     /* NB bd_req NULL until registered */
 504
 505         /* NB...
 506          * 1. If the MD unlink is successful, the ME gets unlinked too.
 507          * 2. Since client-side bulk only gets a single event and a
 508          * .. threshold of 1.  If the MD was inuse at the first link
 509          * .. attempt, the callback is due any minute, and the MD/ME will
 510          * .. unlink themselves.
 511          */
 512         rc = PtlMDUnlink (desc->bd_md_h);
 513         switch (rc) {
 514         default:
 515                 CERROR("PtlMDUnlink returned %d\n", rc);
 516                 LBUG ();
 517         case PTL_OK:                          /* Won the race with completion */
 518                 LASSERT (!desc->bd_complete);   /* Callback hasn't happened */
 519                 desc->bd_network_rw = 0;
 520                 return;
 521         case PTL_MD_INUSE:                  /* MD is being accessed right now */
 522                 for (;;) {
 523                         /* Network access will complete in finite time but the
 524                          * timeout lets us CERROR for visibility */
 525                         if (desc->bd_req->rq_set != NULL)
 526                                 wq = &req->rq_set->set_waitq;
 527                         else
 528                                 wq = &req->rq_wait_for_rep;
 529                         lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL);
 530                         rc = l_wait_event(*wq, ptlrpc_bulk_complete(desc), &lwi);
 531                         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 532                         if (rc == 0)
 533                                 break;
 534                         CERROR ("Unexpectedly long timeout: desc %p\n", desc);
 535                         LBUG();
 536                 }
 537                 /* Fall through */
 538         case PTL_INV_MD:                     /* Lost the race with completion */
 539                 LASSERT (desc->bd_complete);/* Callback has run to completion */
 540                 LASSERT (!desc->bd_network_rw);
 541                 return;
 542         }
 543 }
 544
 545 int ptlrpc_reply(struct ptlrpc_request *req)
 546 {
 547         unsigned long flags;
 548         int rc;
 549
 550         /* We must already have a reply buffer (only ptlrpc_error() may be
 551          * called without one).  We must also have a request buffer which
 552          * is either the actual (swabbed) incoming request, or a saved copy
 553          * if this is a req saved in target_queue_final_reply(). */
 554         LASSERT (req->rq_repmsg != NULL);
 555         LASSERT (req->rq_reqmsg != NULL);
 556
 557         /* FIXME: we need to increment the count of handled events */
 558         if (req->rq_type != PTL_RPC_MSG_ERR)
 559                 req->rq_type = PTL_RPC_MSG_REPLY;
 560
 561         req->rq_repmsg->status = req->rq_status;
 562         req->rq_repmsg->opc = req->rq_reqmsg->opc;
 563
 564         init_waitqueue_head(&req->rq_wait_for_rep);
 565         rc = ptl_send_buf(req, req->rq_connection, req->rq_svc->srv_rep_portal);
 566         if (rc != 0) {
 567                 /* Do what the callback handler would have done */
 568                 OBD_FREE (req->rq_repmsg, req->rq_replen);
 569
 570                 spin_lock_irqsave (&req->rq_lock, flags);
 571                 req->rq_want_ack = 0;
 572                 spin_unlock_irqrestore (&req->rq_lock, flags);
 573         }
 574         return rc;
 575 }
 576
 577 int ptlrpc_error(struct ptlrpc_request *req)
 578 {
 579         int rc;
 580         ENTRY;
 581
 582         if (!req->rq_repmsg) {
 583                 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen,
 584                                      &req->rq_repmsg);
 585                 if (rc)
 586                         RETURN(rc);
 587         }
 588
 589
 590         req->rq_type = PTL_RPC_MSG_ERR;
 591
 592         rc = ptlrpc_reply(req);
 593         RETURN(rc);
 594 }
 595
 596 int ptl_send_rpc(struct ptlrpc_request *request)
 597 {
 598         int rc;
 599         int rc2;
 600         unsigned long flags;
 601         ptl_process_id_t source_id;
 602         ptl_handle_me_t  reply_me_h;
 603         ENTRY;
 604
 605         LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 606
 607         /* If this is a re-transmit, we're required to have disengaged
 608          * cleanly from the previous attempt */
 609         LASSERT (!request->rq_receiving_reply);
 610
 611         if (request->rq_bulk != NULL) {
 612                 rc = ptlrpc_register_bulk (request);
 613                 if (rc != 0)
 614                         RETURN(rc);
 615         }
 616
 617         request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
 618
 619         source_id.nid = request->rq_connection->c_peer.peer_nid;
 620         source_id.pid = PTL_PID_ANY;
 621
 622         LASSERT (request->rq_replen != 0);
 623         OBD_ALLOC(request->rq_repmsg, request->rq_replen);
 624         if (request->rq_repmsg == NULL) {
 625                 LBUG();
 626                 RETURN(-ENOMEM);
 627         }
 628
 629         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni->pni_ni_h,
 630                          request->rq_reply_portal, /* XXX FIXME bug 249 */
 631                          source_id, request->rq_xid, 0, PTL_UNLINK,
 632                          PTL_INS_AFTER, &reply_me_h);
 633         if (rc != PTL_OK) {
 634                 CERROR("PtlMEAttach failed: %d\n", rc);
 635                 LASSERT (rc == PTL_NOSPACE);
 636                 LBUG();
 637                 GOTO(cleanup, rc = -ENOMEM);
 638         }
 639
 640         request->rq_reply_md.start = request->rq_repmsg;
 641         request->rq_reply_md.length = request->rq_replen;
 642         request->rq_reply_md.threshold = 1;
 643         request->rq_reply_md.options = PTL_MD_OP_PUT;
 644         request->rq_reply_md.user_ptr = request;
 645         request->rq_reply_md.eventq =
 646                 request->rq_connection->c_peer.peer_ni->pni_reply_in_eq_h;
 647
 648         rc = PtlMDAttach(reply_me_h, request->rq_reply_md,
 649                          PTL_UNLINK, &request->rq_reply_md_h);
 650         if (rc != PTL_OK) {
 651                 CERROR("PtlMDAttach failed: %d\n", rc);
 652                 LASSERT (rc == PTL_NOSPACE);
 653                 LBUG();
 654                 GOTO(cleanup2, rc -ENOMEM);
 655         }
 656
 657         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 658                ", portal %u on %s\n",
 659                request->rq_replen, request->rq_xid,
 660                request->rq_reply_portal,
 661                request->rq_connection->c_peer.peer_ni->pni_name);
 662
 663         ptlrpc_request_addref(request);        /* 1 ref for the SENT callback */
 664
 665         spin_lock_irqsave (&request->rq_lock, flags);
 666         request->rq_receiving_reply = 1;
 667         /* Clear any flags that may be present from previous sends. */
 668         request->rq_replied = 0;
 669         request->rq_err = 0;
 670         request->rq_timedout = 0;
 671         request->rq_resend = 0;
 672         request->rq_restart = 0;
 673         spin_unlock_irqrestore (&request->rq_lock, flags);
 674
 675         request->rq_sent = LTIME_S(CURRENT_TIME);
 676         ptlrpc_pinger_sending_on_import(request->rq_import);
 677         rc = ptl_send_buf(request, request->rq_connection,
 678                           request->rq_request_portal);
 679         if (rc == 0)
 680                 RETURN(rc);
 681
 682         spin_lock_irqsave (&request->rq_lock, flags);
 683         request->rq_receiving_reply = 0;
 684         spin_unlock_irqrestore (&request->rq_lock, flags);
 685         ptlrpc_req_finished (request);          /* drop callback ref */
 686  cleanup2:
 687         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 688          * nobody apart from the PUT's target has the right nid+XID to
 689          * access the reply buffer. */
 690         rc2 = PtlMEUnlink(reply_me_h);
 691         LASSERT (rc2 == PTL_OK);
 692  cleanup:
 693         OBD_FREE(request->rq_repmsg, request->rq_replen);
 694         request->rq_repmsg = NULL;
 695         return rc;
 696 }
 697
 698 void ptlrpc_link_svc_me(struct ptlrpc_request_buffer_desc *rqbd)
 699 {
 700         struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
 701         struct ptlrpc_service *service = srv_ni->sni_service;
 702         static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
 703         int rc;
 704         ptl_md_t dummy;
 705         ptl_handle_md_t md_h;
 706
 707         LASSERT(atomic_read(&rqbd->rqbd_refcount) == 0);
 708
 709         CDEBUG(D_NET, "PtlMEAttach: portal %d on %s h %lx."LPX64"\n",
 710                service->srv_req_portal, srv_ni->sni_ni->pni_name,
 711                srv_ni->sni_ni->pni_ni_h.nal_idx,
 712                srv_ni->sni_ni->pni_ni_h.cookie);
 713
 714         /* Attach the leading ME on which we build the ring */
 715         rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
 716                          match_id, 0, ~0,
 717                          PTL_UNLINK, PTL_INS_AFTER, &rqbd->rqbd_me_h);
 718         if (rc != PTL_OK) {
 719                 CERROR("PtlMEAttach failed: %d\n", rc);
 720                 /* BUG 1191 */
 721                 LBUG();
 722         }
 723
 724         dummy.start      = rqbd->rqbd_buffer;
 725         dummy.length     = service->srv_buf_size;
 726         dummy.max_size   = service->srv_max_req_size;
 727         dummy.threshold  = PTL_MD_THRESH_INF;
 728         dummy.options    = PTL_MD_OP_PUT | PTL_MD_MAX_SIZE | PTL_MD_AUTO_UNLINK;
 729         dummy.user_ptr   = rqbd;
 730         dummy.eventq     = srv_ni->sni_eq_h;
 731
 732         atomic_inc(&srv_ni->sni_nrqbds_receiving);
 733         atomic_set(&rqbd->rqbd_refcount, 1);   /* 1 ref for portals */
 734
 735         rc = PtlMDAttach(rqbd->rqbd_me_h, dummy, PTL_UNLINK, &md_h);
 736         if (rc != PTL_OK) {
 737                 CERROR("PtlMDAttach failed: %d\n", rc);
 738                 LASSERT (rc == PTL_NOSPACE);
 739                 LBUG();
 740                 /* BUG 1191 */
 741                 PtlMEUnlink (rqbd->rqbd_me_h);
 742                 atomic_set(&rqbd->rqbd_refcount, 0);
 743                 atomic_dec(&srv_ni->sni_nrqbds_receiving);
 744         }
 745 }