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         char str[PTL_NALFMT_SIZE];
  43
  44         LASSERT (portal != 0);
  45         LASSERT (conn != NULL);
  46         CDEBUG (D_INFO, "conn=%p ni %s nid "LPX64" (%s) on %s\n",
  47                 conn, conn->c_peer.peer_ni->pni_name,
  48                 conn->c_peer.peer_nid,
  49                 portals_nid2str(conn->c_peer.peer_ni->pni_number,
  50                                 conn->c_peer.peer_nid, str),
  51                 conn->c_peer.peer_ni->pni_name);
  52
  53         request->rq_req_md.user_ptr = request;
  54
  55         switch (request->rq_type) {
  56         case PTL_RPC_MSG_REQUEST:
  57                 request->rq_reqmsg->type = request->rq_type;
  58                 request->rq_req_md.start = request->rq_reqmsg;
  59                 request->rq_req_md.length = request->rq_reqlen;
  60                 request->rq_req_md.eventq =
  61                         conn->c_peer.peer_ni->pni_request_out_eq_h;
  62                 LASSERT (!request->rq_want_ack);
  63                 break;
  64         case PTL_RPC_MSG_ERR:
  65         case PTL_RPC_MSG_REPLY:
  66                 request->rq_repmsg->type = request->rq_type;
  67                 request->rq_req_md.start = request->rq_repmsg;
  68                 request->rq_req_md.length = request->rq_replen;
  69                 request->rq_req_md.eventq =
  70                         conn->c_peer.peer_ni->pni_reply_out_eq_h;
  71                 break;
  72         default:
  73                 LBUG();
  74                 return -1; /* notreached */
  75         }
  76         if (request->rq_want_ack) {
  77                 request->rq_req_md.threshold = 2; /* SENT and ACK */
  78                 ack_req = PTL_ACK_REQ;
  79         } else {
  80                 request->rq_req_md.threshold = 1;
  81                 ack_req = PTL_NOACK_REQ;
  82         }
  83         request->rq_req_md.options = PTL_MD_OP_PUT;
  84         request->rq_req_md.user_ptr = request;
  85
  86         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
  87                 request->rq_req_md.options |= PTL_MD_ACK_DISABLE;
  88                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
  89         }
  90
  91         /* NB if the send fails, we back out of the send and return
  92          * failure; it's down to the caller to handle missing callbacks */
  93
  94         rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h, request->rq_req_md,
  95                        &md_h);
  96         if (rc != PTL_OK) {
  97                 CERROR("PtlMDBind failed: %d\n", rc);
  98                 LASSERT (rc == PTL_NOSPACE);
  99                 RETURN (-ENOMEM);
 100         }
 101         if (request->rq_type != PTL_RPC_MSG_REQUEST)
 102                 memcpy(&request->rq_reply_md_h, &md_h, sizeof(md_h));
 103
 104         remote_id.nid = conn->c_peer.peer_nid;
 105         remote_id.pid = 0;
 106
 107         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
 108                request->rq_req_md.length, portal, request->rq_xid);
 109
 110         rc = PtlPut(md_h, ack_req, remote_id, portal, 0, request->rq_xid, 0, 0);
 111         if (rc != PTL_OK) {
 112                 CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n",
 113                        remote_id.nid, portal, request->rq_xid, rc);
 114                 rc2 = PtlMDUnlink(md_h);
 115                 LASSERT (rc2 == PTL_OK);
 116                 RETURN ((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 117         }
 118
 119         return 0;
 120 }
 121
 122 static inline ptl_kiov_t *
 123 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
 124 {
 125         ptl_kiov_t *iov;
 126
 127         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov))
 128                 return (desc->bd_iov);
 129
 130         OBD_ALLOC (iov, desc->bd_page_count * sizeof (*iov));
 131         if (iov == NULL)
 132                 LBUG();
 133
 134         return (iov);
 135 }
 136
 137 static inline void
 138 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, ptl_kiov_t *iov)
 139 {
 140         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (*iov))
 141                 return;
 142
 143         OBD_FREE (iov, desc->bd_page_count * sizeof (*iov));
 144 }
 145
 146 int ptlrpc_bulk_put(struct ptlrpc_bulk_desc *desc)
 147 {
 148         int rc;
 149         int rc2;
 150         struct ptlrpc_peer *peer;
 151         struct list_head *tmp, *next;
 152         ptl_process_id_t remote_id;
 153         ptl_kiov_t *iov;
 154         __u64 xid;
 155         ENTRY;
 156
 157         /* NB no locking required until desc is on the network */
 158         LASSERT (!desc->bd_network_rw);
 159         LASSERT (desc->bd_type == BULK_PUT_SOURCE);
 160         desc->bd_complete = 0;
 161
 162         iov = ptlrpc_get_bulk_iov (desc);
 163         if (iov == NULL)
 164                 RETURN (-ENOMEM);
 165
 166         peer = &desc->bd_export->exp_connection->c_peer;
 167
 168         desc->bd_md.start = iov;
 169         desc->bd_md.niov = 0;
 170         desc->bd_md.length = 0;
 171         desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_source_eq_h;
 172         desc->bd_md.threshold = 2; /* SENT and ACK */
 173         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV;
 174         desc->bd_md.user_ptr = desc;
 175
 176         desc->bd_callback_count = 2;
 177
 178         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 179                 struct ptlrpc_bulk_page *bulk;
 180                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 181
 182                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 183
 184                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 185                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 186                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 187
 188                 LASSERT (iov[desc->bd_md.niov].kiov_offset +
 189                          iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE);
 190                 desc->bd_md.niov++;
 191                 desc->bd_md.length += bulk->bp_buflen;
 192         }
 193
 194         /* NB total length may be 0 for a read past EOF, so we send a 0
 195          * length bulk, since the client expects a bulk event. */
 196         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 197
 198         rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md,
 199                        &desc->bd_md_h);
 200
 201         ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/
 202
 203         if (rc != PTL_OK) {
 204                 CERROR("PtlMDBind failed: %d\n", rc);
 205                 LASSERT (rc == PTL_NOSPACE);
 206                 RETURN(-ENOMEM);
 207         }
 208
 209         /* Client's bulk and reply matchbits are the same */
 210         xid = desc->bd_req->rq_xid;
 211         remote_id.nid = peer->peer_nid;
 212         remote_id.pid = 0;
 213
 214         CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d on %s "
 215                "nid "LPX64" pid %d xid "LPX64"\n",
 216                desc->bd_md.niov, desc->bd_md.length,
 217                desc->bd_portal, peer->peer_ni->pni_name,
 218                remote_id.nid, remote_id.pid, xid);
 219
 220         desc->bd_network_rw = 1;
 221         rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
 222                     desc->bd_portal, 0, xid, 0, 0);
 223         if (rc != PTL_OK) {
 224                 desc->bd_network_rw = 0;
 225                 CERROR("PtlPut("LPU64", %d, "LPX64") failed: %d\n",
 226                        remote_id.nid, desc->bd_portal, xid, rc);
 227                 rc2 = PtlMDUnlink(desc->bd_md_h);
 228                 LASSERT (rc2 == PTL_OK);
 229                 RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 230         }
 231
 232         RETURN(0);
 233 }
 234
 235 int ptlrpc_bulk_get(struct ptlrpc_bulk_desc *desc)
 236 {
 237         int rc;
 238         int rc2;
 239         struct ptlrpc_peer *peer;
 240         struct list_head *tmp, *next;
 241         ptl_process_id_t remote_id;
 242         ptl_kiov_t *iov;
 243         __u64 xid;
 244         ENTRY;
 245
 246         /* NB no locking required until desc is on the network */
 247         LASSERT (!desc->bd_network_rw);
 248         LASSERT (desc->bd_type == BULK_GET_SINK);
 249         desc->bd_complete = 0;
 250
 251         iov = ptlrpc_get_bulk_iov (desc);
 252         if (iov == NULL)
 253                 RETURN(-ENOMEM);
 254
 255         peer = &desc->bd_export->exp_connection->c_peer;
 256
 257         desc->bd_md.start = iov;
 258         desc->bd_md.niov = 0;
 259         desc->bd_md.length = 0;
 260         desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_sink_eq_h;
 261         desc->bd_md.threshold = 2; /* SENT and REPLY */
 262         desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV;
 263         desc->bd_md.user_ptr = desc;
 264
 265         desc->bd_callback_count = 2;
 266
 267         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 268                 struct ptlrpc_bulk_page *bulk;
 269                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 270
 271                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 272
 273                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 274                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 275                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 276
 277                 LASSERT (iov[desc->bd_md.niov].kiov_offset +
 278                          iov[desc->bd_md.niov].kiov_len <= PAGE_SIZE);
 279                 desc->bd_md.niov++;
 280                 desc->bd_md.length += bulk->bp_buflen;
 281         }
 282
 283         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 284         LASSERT(desc->bd_md.niov != 0);
 285
 286         rc = PtlMDBind(peer->peer_ni->pni_ni_h, desc->bd_md, &desc->bd_md_h);
 287
 288         ptlrpc_put_bulk_iov(desc, iov); /*move down to reduce latency to send*/
 289
 290         if (rc != PTL_OK) {
 291                 CERROR("PtlMDBind failed: %d\n", rc);
 292                 LASSERT (rc == PTL_NOSPACE);
 293                 RETURN(-ENOMEM);
 294         }
 295
 296         /* Client's bulk and reply matchbits are the same */
 297         xid = desc->bd_req->rq_xid;
 298         remote_id.nid = desc->bd_export->exp_connection->c_peer.peer_nid;
 299         remote_id.pid = 0;
 300
 301         CDEBUG(D_NET, "Fetching %u pages %u bytes from portal %d on %s "
 302                "nid "LPX64" pid %d xid "LPX64"\n",
 303                desc->bd_md.niov, desc->bd_md.length, desc->bd_portal,
 304                peer->peer_ni->pni_name, remote_id.nid, remote_id.pid,
 305                xid);
 306
 307         desc->bd_network_rw = 1;
 308         rc = PtlGet(desc->bd_md_h, remote_id, desc->bd_portal, 0,
 309                     xid, 0);
 310         if (rc != PTL_OK) {
 311                 desc->bd_network_rw = 0;
 312                 CERROR("PtlGet("LPU64", %d, "LPX64") failed: %d\n",
 313                        remote_id.nid, desc->bd_portal, xid, rc);
 314                 rc2 = PtlMDUnlink(desc->bd_md_h);
 315                 LASSERT (rc2 == PTL_OK);
 316                 RETURN((rc == PTL_NOSPACE) ? -ENOMEM : -ECOMM);
 317         }
 318
 319         RETURN(0);
 320 }
 321
 322 void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
 323 {
 324         /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 325          * serialises with completion callback) */
 326         unsigned long      flags;
 327         struct l_wait_info lwi;
 328         int                callback_count;
 329         int                rc;
 330
 331         LASSERT (!in_interrupt ());             /* might sleep */
 332
 333         /* NB. server-side bulk gets 2 events, so we have to keep trying to
 334          * unlink the MD until all callbacks have happened, or
 335          * PtlMDUnlink() returns OK or INVALID */
 336  again:
 337         spin_lock_irqsave (&desc->bd_lock, flags);
 338         if (!desc->bd_network_rw) {
 339                 /* completed or never even registered. NB holding bd_lock
 340                  * guarantees callback has completed if it ran. */
 341                 spin_unlock_irqrestore (&desc->bd_lock, flags);
 342                 return;
 343         }
 344
 345         /* sample callback count while we have the lock */
 346         callback_count = desc->bd_callback_count;
 347         spin_unlock_irqrestore (&desc->bd_lock, flags);
 348
 349         rc = PtlMDUnlink (desc->bd_md_h);
 350         switch (rc) {
 351         default:
 352                 CERROR("PtlMDUnlink returned %d\n", rc);
 353                 LBUG ();
 354         case PTL_OK:                    /* Won the race with the network */
 355                 LASSERT (!desc->bd_complete); /* Not all callbacks ran */
 356                 desc->bd_network_rw = 0;
 357                 return;
 358
 359         case PTL_MD_INUSE:              /* MD is being accessed right now */
 360                 for (;;) {
 361                         /* Network access will complete in finite time but the
 362                          * timeout lets us CERROR for visibility */
 363                         lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL);
 364                         rc = l_wait_event(desc->bd_waitq,
 365                                           desc->bd_callback_count !=
 366                                           callback_count, &lwi);
 367                         if (rc == -ETIMEDOUT) {
 368                                 CERROR("Unexpectedly long timeout: desc %p\n",
 369                                        desc);
 370                                 continue;
 371                         }
 372                         LASSERT (rc == 0);
 373                         break;
 374                 }
 375                 /* go back and try again... */
 376                 goto again;
 377
 378         case PTL_INV_MD:            /* Lost the race with completion */
 379                 LASSERT (desc->bd_complete);    /* Callbacks all ran */
 380                 LASSERT (!desc->bd_network_rw);
 381                 return;
 382         }
 383 }
 384
 385 int ptlrpc_register_bulk (struct ptlrpc_request *req)
 386 {
 387         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 388         struct ptlrpc_peer *peer;
 389         struct list_head *tmp, *next;
 390         int rc;
 391         int rc2;
 392         ptl_kiov_t *iov;
 393         ptl_process_id_t source_id;
 394         ENTRY;
 395
 396         /* NB no locking required until desc is on the network */
 397         LASSERT (!desc->bd_network_rw);
 398         LASSERT (desc->bd_page_count <= PTL_MD_MAX_PAGES);
 399         LASSERT (desc->bd_req != NULL);
 400         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 401                  desc->bd_type == BULK_GET_SOURCE);
 402
 403         desc->bd_complete = 0;
 404
 405         iov = ptlrpc_get_bulk_iov (desc);
 406         if (iov == NULL)
 407                 return (-ENOMEM);
 408
 409         peer = &desc->bd_import->imp_connection->c_peer;
 410
 411         desc->bd_md.start = iov;
 412         desc->bd_md.niov = 0;
 413         desc->bd_md.length = 0;
 414         desc->bd_md.threshold = 1;
 415         desc->bd_md.user_ptr = desc;
 416
 417         if (desc->bd_type == BULK_GET_SOURCE) {
 418                 desc->bd_md.options = PTL_MD_OP_GET | PTL_MD_KIOV;
 419                 desc->bd_md.eventq = peer->peer_ni->pni_bulk_get_source_eq_h;
 420         } else {
 421                 desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_KIOV;
 422                 desc->bd_md.eventq = peer->peer_ni->pni_bulk_put_sink_eq_h;
 423         }
 424
 425         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 426                 struct ptlrpc_bulk_page *bulk;
 427                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 428
 429                 LASSERT(desc->bd_md.niov < desc->bd_page_count);
 430
 431                 iov[desc->bd_md.niov].kiov_page = bulk->bp_page;
 432                 iov[desc->bd_md.niov].kiov_len = bulk->bp_buflen;
 433                 iov[desc->bd_md.niov].kiov_offset = bulk->bp_pageoffset;
 434
 435                 LASSERT (bulk->bp_pageoffset + bulk->bp_buflen <= PAGE_SIZE);
 436                 desc->bd_md.niov++;
 437                 desc->bd_md.length += bulk->bp_buflen;
 438         }
 439
 440         LASSERT(desc->bd_md.niov == desc->bd_page_count);
 441         LASSERT(desc->bd_md.niov != 0);
 442
 443         /* XXX Registering the same xid on retried bulk makes my head
 444          * explode trying to understand how the original request's bulk
 445          * might interfere with the retried request -eeb */
 446         LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
 447         desc->bd_registered = 1;
 448         desc->bd_last_xid = desc->bd_last_xid;
 449
 450         source_id.nid = desc->bd_import->imp_connection->c_peer.peer_nid;
 451         source_id.pid = PTL_PID_ANY;
 452
 453         rc = PtlMEAttach(peer->peer_ni->pni_ni_h,
 454                          desc->bd_portal, source_id, req->rq_xid, 0,
 455                          PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
 456
 457         if (rc != PTL_OK) {
 458                 CERROR("PtlMEAttach failed: %d\n", rc);
 459                 LASSERT (rc == PTL_NOSPACE);
 460                 GOTO(out, rc = -ENOMEM);
 461         }
 462
 463         /* About to let the network at it... */
 464         desc->bd_network_rw = 1;
 465         rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
 466                          &desc->bd_md_h);
 467         if (rc != PTL_OK) {
 468                 CERROR("PtlMDAttach failed: %d\n", rc);
 469                 LASSERT (rc == PTL_NOSPACE);
 470                 desc->bd_network_rw = 0;
 471                 rc2 = PtlMEUnlink (desc->bd_me_h);
 472                 LASSERT (rc2 == PTL_OK);
 473                 GOTO(out, rc = -ENOMEM);
 474         }
 475         rc = 0;
 476
 477         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
 478                "portal %u on %s\n",
 479                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 480                desc->bd_md.niov, desc->bd_md.length,
 481                req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
 482
 483  out:
 484         ptlrpc_put_bulk_iov (desc, iov);
 485         RETURN(rc);
 486 }
 487
 488 void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
 489 {
 490         /* Disconnect a bulk desc from the network. Idempotent. Not
 491          * thread-safe (i.e. only interlocks with completion callback). */
 492         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 493         wait_queue_head_t       *wq;
 494         unsigned long            flags;
 495         struct l_wait_info       lwi;
 496         int                      rc;
 497
 498         LASSERT (!in_interrupt ());             /* might sleep */
 499
 500         spin_lock_irqsave (&desc->bd_lock, flags);
 501         if (!desc->bd_network_rw) {     /* completed or never even registered */
 502                 spin_unlock_irqrestore (&desc->bd_lock, flags);
 503                 return;
 504         }
 505         spin_unlock_irqrestore (&desc->bd_lock, flags);
 506
 507         LASSERT (desc->bd_req == req);     /* NB bd_req NULL until registered */
 508
 509         /* NB...
 510          * 1. If the MD unlink is successful, the ME gets unlinked too.
 511          * 2. Since client-side bulk only gets a single event and a
 512          * .. threshold of 1.  If the MD was inuse at the first link
 513          * .. attempt, the callback is due any minute, and the MD/ME will
 514          * .. unlink themselves.
 515          */
 516         rc = PtlMDUnlink (desc->bd_md_h);
 517         switch (rc) {
 518         default:
 519                 CERROR("PtlMDUnlink returned %d\n", rc);
 520                 LBUG ();
 521         case PTL_OK:                          /* Won the race with completion */
 522                 LASSERT (!desc->bd_complete);   /* Callback hasn't happened */
 523                 desc->bd_network_rw = 0;
 524                 return;
 525         case PTL_MD_INUSE:                  /* MD is being accessed right now */
 526                 for (;;) {
 527                         /* Network access will complete in finite time but the
 528                          * timeout lets us CERROR for visibility */
 529                         if (desc->bd_req->rq_set != NULL)
 530                                 wq = &req->rq_set->set_waitq;
 531                         else
 532                                 wq = &req->rq_reply_waitq;
 533                         lwi = LWI_TIMEOUT (10 * HZ, NULL, NULL);
 534                         rc = l_wait_event(*wq, ptlrpc_bulk_complete(desc), &lwi);
 535                         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 536                         if (rc == 0)
 537                                 break;
 538                         CERROR ("Unexpectedly long timeout: desc %p\n", desc);
 539                         LBUG();
 540                 }
 541                 /* Fall through */
 542         case PTL_INV_MD:                     /* Lost the race with completion */
 543                 LASSERT (desc->bd_complete);/* Callback has run to completion */
 544                 LASSERT (!desc->bd_network_rw);
 545                 return;
 546         }
 547 }
 548
 549 int ptlrpc_reply(struct ptlrpc_request *req)
 550 {
 551         unsigned long flags;
 552         int rc;
 553
 554         /* We must already have a reply buffer (only ptlrpc_error() may be
 555          * called without one).  We must also have a request buffer which
 556          * is either the actual (swabbed) incoming request, or a saved copy
 557          * if this is a req saved in target_queue_final_reply(). */
 558         LASSERT (req->rq_repmsg != NULL);
 559         LASSERT (req->rq_reqmsg != NULL);
 560
 561         /* FIXME: we need to increment the count of handled events */
 562         if (req->rq_type != PTL_RPC_MSG_ERR)
 563                 req->rq_type = PTL_RPC_MSG_REPLY;
 564
 565         req->rq_repmsg->status = req->rq_status;
 566         req->rq_repmsg->opc = req->rq_reqmsg->opc;
 567
 568         init_waitqueue_head(&req->rq_reply_waitq);
 569         rc = ptl_send_buf(req, req->rq_connection, req->rq_svc->srv_rep_portal);
 570         if (rc != 0) {
 571                 /* Do what the callback handler would have done */
 572                 OBD_FREE (req->rq_repmsg, req->rq_replen);
 573
 574                 spin_lock_irqsave (&req->rq_lock, flags);
 575                 req->rq_want_ack = 0;
 576                 spin_unlock_irqrestore (&req->rq_lock, flags);
 577         }
 578         return rc;
 579 }
 580
 581 int ptlrpc_error(struct ptlrpc_request *req)
 582 {
 583         int rc;
 584         ENTRY;
 585
 586         if (!req->rq_repmsg) {
 587                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 588                 if (rc)
 589                         RETURN(rc);
 590         }
 591
 592
 593         req->rq_type = PTL_RPC_MSG_ERR;
 594
 595         rc = ptlrpc_reply(req);
 596         RETURN(rc);
 597 }
 598
 599 int ptl_send_rpc(struct ptlrpc_request *request)
 600 {
 601         int rc;
 602         int rc2;
 603         unsigned long flags;
 604         ptl_process_id_t source_id;
 605         ptl_handle_me_t  reply_me_h;
 606         ENTRY;
 607
 608         LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 609
 610         /* If this is a re-transmit, we're required to have disengaged
 611          * cleanly from the previous attempt */
 612         LASSERT (!request->rq_receiving_reply);
 613
 614         if (request->rq_bulk != NULL) {
 615                 rc = ptlrpc_register_bulk (request);
 616                 if (rc != 0)
 617                         RETURN(rc);
 618         }
 619
 620         request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
 621         request->rq_reqmsg->conn_cnt = request->rq_import->imp_conn_cnt;
 622
 623         source_id.nid = request->rq_connection->c_peer.peer_nid;
 624         source_id.pid = PTL_PID_ANY;
 625
 626         LASSERT (request->rq_replen != 0);
 627         if (request->rq_repmsg == NULL)
 628                 OBD_ALLOC(request->rq_repmsg, request->rq_replen);
 629         if (request->rq_repmsg == NULL) {
 630                 LBUG();
 631                 RETURN(-ENOMEM);
 632         }
 633
 634         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni->pni_ni_h,
 635                          request->rq_reply_portal, /* XXX FIXME bug 249 */
 636                          source_id, request->rq_xid, 0, PTL_UNLINK,
 637                          PTL_INS_AFTER, &reply_me_h);
 638         if (rc != PTL_OK) {
 639                 CERROR("PtlMEAttach failed: %d\n", rc);
 640                 LASSERT (rc == PTL_NOSPACE);
 641                 LBUG();
 642                 GOTO(cleanup, rc = -ENOMEM);
 643         }
 644
 645         request->rq_reply_md.start = request->rq_repmsg;
 646         request->rq_reply_md.length = request->rq_replen;
 647         request->rq_reply_md.threshold = 1;
 648         request->rq_reply_md.options = PTL_MD_OP_PUT;
 649         request->rq_reply_md.user_ptr = request;
 650         request->rq_reply_md.eventq =
 651                 request->rq_connection->c_peer.peer_ni->pni_reply_in_eq_h;
 652
 653         rc = PtlMDAttach(reply_me_h, request->rq_reply_md,
 654                          PTL_UNLINK, &request->rq_reply_md_h);
 655         if (rc != PTL_OK) {
 656                 CERROR("PtlMDAttach failed: %d\n", rc);
 657                 LASSERT (rc == PTL_NOSPACE);
 658                 LBUG();
 659                 GOTO(cleanup2, rc -ENOMEM);
 660         }
 661
 662         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 663                ", portal %u on %s\n",
 664                request->rq_replen, request->rq_xid,
 665                request->rq_reply_portal,
 666                request->rq_connection->c_peer.peer_ni->pni_name);
 667
 668         ptlrpc_request_addref(request);        /* 1 ref for the SENT callback */
 669
 670         spin_lock_irqsave (&request->rq_lock, flags);
 671         request->rq_receiving_reply = 1;
 672         /* Clear any flags that may be present from previous sends. */
 673         request->rq_replied = 0;
 674         request->rq_err = 0;
 675         request->rq_timedout = 0;
 676         request->rq_resend = 0;
 677         request->rq_restart = 0;
 678         spin_unlock_irqrestore (&request->rq_lock, flags);
 679
 680         request->rq_sent = LTIME_S(CURRENT_TIME);
 681         ptlrpc_pinger_sending_on_import(request->rq_import);
 682         rc = ptl_send_buf(request, request->rq_connection,
 683                           request->rq_request_portal);
 684         if (rc == 0) {
 685                 ptlrpc_lprocfs_rpc_sent(request);
 686                 RETURN(rc);
 687         }
 688
 689         spin_lock_irqsave (&request->rq_lock, flags);
 690         request->rq_receiving_reply = 0;
 691         spin_unlock_irqrestore (&request->rq_lock, flags);
 692         ptlrpc_req_finished (request);          /* drop callback ref */
 693  cleanup2:
 694         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 695          * nobody apart from the PUT's target has the right nid+XID to
 696          * access the reply buffer. */
 697         rc2 = PtlMEUnlink(reply_me_h);
 698         LASSERT (rc2 == PTL_OK);
 699  cleanup:
 700         OBD_FREE(request->rq_repmsg, request->rq_replen);
 701         request->rq_repmsg = NULL;
 702         return rc;
 703 }
 704
 705 void ptlrpc_link_svc_me(struct ptlrpc_request_buffer_desc *rqbd)
 706 {
 707         struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
 708         struct ptlrpc_service *service = srv_ni->sni_service;
 709         static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
 710         int rc;
 711         ptl_md_t dummy;
 712         ptl_handle_md_t md_h;
 713
 714         LASSERT(atomic_read(&rqbd->rqbd_refcount) == 0);
 715
 716         CDEBUG(D_NET, "PtlMEAttach: portal %d on %s h %lx."LPX64"\n",
 717                service->srv_req_portal, srv_ni->sni_ni->pni_name,
 718                srv_ni->sni_ni->pni_ni_h.nal_idx,
 719                srv_ni->sni_ni->pni_ni_h.cookie);
 720
 721         /* Attach the leading ME on which we build the ring */
 722         rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
 723                          match_id, 0, ~0,
 724                          PTL_UNLINK, PTL_INS_AFTER, &rqbd->rqbd_me_h);
 725         if (rc != PTL_OK) {
 726                 CERROR("PtlMEAttach failed: %d\n", rc);
 727                 /* BUG 1191 */
 728                 LBUG();
 729         }
 730
 731         dummy.start      = rqbd->rqbd_buffer;
 732         dummy.length     = service->srv_buf_size;
 733         dummy.max_size   = service->srv_max_req_size;
 734         dummy.threshold  = PTL_MD_THRESH_INF;
 735         dummy.options    = PTL_MD_OP_PUT | PTL_MD_MAX_SIZE | PTL_MD_AUTO_UNLINK;
 736         dummy.user_ptr   = rqbd;
 737         dummy.eventq     = srv_ni->sni_eq_h;
 738
 739         atomic_inc(&srv_ni->sni_nrqbds_receiving);
 740         atomic_set(&rqbd->rqbd_refcount, 1);   /* 1 ref for portals */
 741
 742         rc = PtlMDAttach(rqbd->rqbd_me_h, dummy, PTL_UNLINK, &md_h);
 743         if (rc != PTL_OK) {
 744                 CERROR("PtlMDAttach failed: %d\n", rc);
 745                 LASSERT (rc == PTL_NOSPACE);
 746                 LBUG();
 747                 /* BUG 1191 */
 748                 PtlMEUnlink (rqbd->rqbd_me_h);
 749                 atomic_set(&rqbd->rqbd_refcount, 0);
 750                 atomic_dec(&srv_ni->sni_nrqbds_receiving);
 751         }
 752 }