lustre/ptlrpc/niobuf.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #ifndef __KERNEL__
  39 #include <liblustre.h>
  40 #endif
  41 #include <obd_support.h>
  42 #include <lustre_net.h>
  43 #include <lustre_lib.h>
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include "ptlrpc_internal.h"
  47
  48 /**
  49  * Helper function. Sends \a len bytes from \a base at offset \a offset
  50  * over \a conn connection to portal \a portal.
  51  * Returns 0 on success or error code.
  52  */
  53 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  54                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  55                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  56                          unsigned int offset)
  57 {
  58         int              rc;
  59         lnet_md_t         md;
  60         ENTRY;
  61
  62         LASSERT (portal != 0);
  63         LASSERT (conn != NULL);
  64         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  65         md.start     = base;
  66         md.length    = len;
  67         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  68         md.options   = PTLRPC_MD_OPTIONS;
  69         md.user_ptr  = cbid;
  70         md.eq_handle = ptlrpc_eq_h;
  71
  72         if (unlikely(ack == LNET_ACK_REQ &&
  73                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  74                 /* don't ask for the ack to simulate failing client */
  75                 ack = LNET_NOACK_REQ;
  76         }
  77
  78         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  79         if (unlikely(rc != 0)) {
  80                 CERROR ("LNetMDBind failed: %d\n", rc);
  81                 LASSERT (rc == -ENOMEM);
  82                 RETURN (-ENOMEM);
  83         }
  84
  85         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  86                len, portal, xid, offset);
  87
  88         rc = LNetPut (conn->c_self, *mdh, ack,
  89                       conn->c_peer, portal, xid, offset, 0);
  90         if (unlikely(rc != 0)) {
  91                 int rc2;
  92                 /* We're going to get an UNLINK event when I unlink below,
  93                  * which will complete just like any other failed send, so
  94                  * I fall through and return success here! */
  95                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  96                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  97                 rc2 = LNetMDUnlink(*mdh);
  98                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  99         }
 100
 101         RETURN (0);
 102 }
 103
 104 #ifdef HAVE_SERVER_SUPPORT
 105 /**
 106  * Prepare bulk descriptor for specified incoming request \a req that
 107  * can fit \a npages * pages. \a type is bulk type. \a portal is where
 108  * the bulk to be sent. Used on server-side after request was already
 109  * received.
 110  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 111  * error.
 112  */
 113 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 114                                               int npages, int type, int portal)
 115 {
 116         struct obd_export *exp = req->rq_export;
 117         struct ptlrpc_bulk_desc *desc;
 118
 119         ENTRY;
 120         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 121
 122         desc = new_bulk(npages, type, portal);
 123         if (desc == NULL)
 124                 RETURN(NULL);
 125
 126         desc->bd_export = class_export_get(exp);
 127         desc->bd_req = req;
 128
 129         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 130         desc->bd_cbid.cbid_arg = desc;
 131
 132         /* NB we don't assign rq_bulk here; server-side requests are
 133          * re-used, and the handler frees the bulk desc explicitly. */
 134
 135         return desc;
 136 }
 137 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 138
 139 /**
 140  * Starts bulk transfer for descriptor \a desc
 141  * Returns 0 on success or error code.
 142  */
 143 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 144 {
 145         struct ptlrpc_connection *conn = desc->bd_export->exp_connection;
 146         int                       rc;
 147         int                       rc2;
 148         lnet_md_t                 md;
 149         __u64                     xid;
 150         ENTRY;
 151
 152         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 153                 RETURN(0);
 154
 155         /* NB no locking required until desc is on the network */
 156         LASSERT (!desc->bd_network_rw);
 157         LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
 158                  desc->bd_type == BULK_GET_SINK);
 159         desc->bd_success = 0;
 160
 161         md.user_ptr = &desc->bd_cbid;
 162         md.eq_handle = ptlrpc_eq_h;
 163         md.threshold = 2; /* SENT and ACK/REPLY */
 164         md.options = PTLRPC_MD_OPTIONS;
 165         ptlrpc_fill_bulk_md(&md, desc);
 166
 167         LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
 168         LASSERT (desc->bd_cbid.cbid_arg == desc);
 169
 170         /* NB total length may be 0 for a read past EOF, so we send a 0
 171          * length bulk, since the client expects a bulk event. */
 172
 173         rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_md_h);
 174         if (rc != 0) {
 175                 CERROR("LNetMDBind failed: %d\n", rc);
 176                 LASSERT (rc == -ENOMEM);
 177                 RETURN(-ENOMEM);
 178         }
 179
 180         /* Client's bulk and reply matchbits are the same */
 181         xid = desc->bd_req->rq_xid;
 182         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 183                "id %s xid "LPX64"\n", desc->bd_iov_count,
 184                desc->bd_nob, desc->bd_portal,
 185                libcfs_id2str(conn->c_peer), xid);
 186
 187         /* Network is about to get at the memory */
 188         desc->bd_network_rw = 1;
 189
 190         if (desc->bd_type == BULK_PUT_SOURCE)
 191                 rc = LNetPut (conn->c_self, desc->bd_md_h, LNET_ACK_REQ,
 192                               conn->c_peer, desc->bd_portal, xid, 0, 0);
 193         else
 194                 rc = LNetGet (conn->c_self, desc->bd_md_h,
 195                               conn->c_peer, desc->bd_portal, xid, 0);
 196
 197         if (rc != 0) {
 198                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 199                  * event this creates will signal completion with failure,
 200                  * so we return SUCCESS here! */
 201                 CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
 202                        libcfs_id2str(conn->c_peer), desc->bd_portal, xid, rc);
 203                 rc2 = LNetMDUnlink(desc->bd_md_h);
 204                 LASSERT (rc2 == 0);
 205         }
 206
 207         RETURN(0);
 208 }
 209 EXPORT_SYMBOL(ptlrpc_start_bulk_transfer);
 210
 211 /**
 212  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 213  * serialises with completion callback)
 214  */
 215 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 216 {
 217         struct l_wait_info       lwi;
 218         int                      rc;
 219
 220         LASSERT(!cfs_in_interrupt());           /* might sleep */
 221
 222         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 223                 return;                         /* never started */
 224
 225         /* We used to poison the pages with 0xab here because we did not want to
 226          * send any meaningful data over the wire for evicted clients (bug 9297)
 227          * However, this is no longer safe now that we use the page cache on the
 228          * OSS (bug 20560) */
 229
 230         /* The unlink ensures the callback happens ASAP and is the last
 231          * one.  If it fails, it must be because completion just happened,
 232          * but we must still l_wait_event() in this case, to give liblustre
 233          * a chance to run server_bulk_callback()*/
 234
 235         LNetMDUnlink(desc->bd_md_h);
 236
 237         for (;;) {
 238                 /* Network access will complete in finite time but the HUGE
 239                  * timeout lets us CWARN for visibility of sluggish NALs */
 240                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 241                                            cfs_time_seconds(1), NULL, NULL);
 242                 rc = l_wait_event(desc->bd_waitq,
 243                                   !ptlrpc_server_bulk_active(desc), &lwi);
 244                 if (rc == 0)
 245                         return;
 246
 247                 LASSERT(rc == -ETIMEDOUT);
 248                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 249         }
 250 }
 251 EXPORT_SYMBOL(ptlrpc_abort_bulk);
 252 #endif /* HAVE_SERVER_SUPPORT */
 253
 254 /**
 255  * Register bulk for later transfer
 256  * Returns 0 on success or error code.
 257  */
 258 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 259 {
 260         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 261         lnet_process_id_t peer;
 262         int rc;
 263         int rc2;
 264         lnet_handle_me_t  me_h;
 265         lnet_md_t         md;
 266         ENTRY;
 267
 268         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 269                 RETURN(0);
 270
 271         /* NB no locking required until desc is on the network */
 272         LASSERT (desc->bd_nob > 0);
 273         LASSERT (!desc->bd_network_rw);
 274         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 275         LASSERT (desc->bd_req != NULL);
 276         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 277                  desc->bd_type == BULK_GET_SOURCE);
 278
 279         desc->bd_success = 0;
 280
 281         peer = desc->bd_import->imp_connection->c_peer;
 282
 283         md.user_ptr = &desc->bd_cbid;
 284         md.eq_handle = ptlrpc_eq_h;
 285         md.threshold = 1;                       /* PUT or GET */
 286         md.options = PTLRPC_MD_OPTIONS |
 287                      ((desc->bd_type == BULK_GET_SOURCE) ?
 288                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 289         ptlrpc_fill_bulk_md(&md, desc);
 290
 291         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 292         LASSERT (desc->bd_cbid.cbid_arg == desc);
 293
 294         /* XXX Registering the same xid on retried bulk makes my head
 295          * explode trying to understand how the original request's bulk
 296          * might interfere with the retried request -eeb
 297          * On the other hand replaying with the same xid is fine, since
 298          * we are guaranteed old request have completed. -green */
 299         LASSERTF(!(desc->bd_registered &&
 300                  req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 301                  req->rq_xid != desc->bd_last_xid,
 302                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 303                  desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 304         desc->bd_registered = 1;
 305         desc->bd_last_xid = req->rq_xid;
 306
 307         rc = LNetMEAttach(desc->bd_portal, peer,
 308                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 309         if (rc != 0) {
 310                 CERROR("LNetMEAttach failed: %d\n", rc);
 311                 LASSERT (rc == -ENOMEM);
 312                 RETURN (-ENOMEM);
 313         }
 314
 315         /* About to let the network at it... */
 316         desc->bd_network_rw = 1;
 317         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 318         if (rc != 0) {
 319                 CERROR("LNetMDAttach failed: %d\n", rc);
 320                 LASSERT (rc == -ENOMEM);
 321                 desc->bd_network_rw = 0;
 322                 rc2 = LNetMEUnlink (me_h);
 323                 LASSERT (rc2 == 0);
 324                 RETURN (-ENOMEM);
 325         }
 326
 327         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", "
 328                "portal %u\n",
 329                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 330                desc->bd_iov_count, desc->bd_nob,
 331                req->rq_xid, desc->bd_portal);
 332         RETURN(0);
 333 }
 334 EXPORT_SYMBOL(ptlrpc_register_bulk);
 335
 336 /**
 337  * Disconnect a bulk desc from the network. Idempotent. Not
 338  * thread-safe (i.e. only interlocks with completion callback).
 339  * Returns 1 on success or 0 if network unregistration failed for whatever
 340  * reason.
 341  */
 342 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 343 {
 344         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 345         cfs_waitq_t             *wq;
 346         struct l_wait_info       lwi;
 347         int                      rc;
 348         ENTRY;
 349
 350         LASSERT(!cfs_in_interrupt());     /* might sleep */
 351
 352         /* Let's setup deadline for reply unlink. */
 353         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 354             async && req->rq_bulk_deadline == 0)
 355                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 356
 357         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 358                 RETURN(1);                    /* never registered */
 359
 360         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 361
 362         /* the unlink ensures the callback happens ASAP and is the last
 363          * one.  If it fails, it must be because completion just happened,
 364          * but we must still l_wait_event() in this case to give liblustre
 365          * a chance to run client_bulk_callback() */
 366
 367         LNetMDUnlink(desc->bd_md_h);
 368
 369         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 370                 RETURN(1);                    /* never registered */
 371
 372         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 373         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 374
 375         /* Do not wait for unlink to finish. */
 376         if (async)
 377                 RETURN(0);
 378
 379         if (req->rq_set != NULL)
 380                 wq = &req->rq_set->set_waitq;
 381         else
 382                 wq = &req->rq_reply_waitq;
 383
 384         for (;;) {
 385                 /* Network access will complete in finite time but the HUGE
 386                  * timeout lets us CWARN for visibility of sluggish NALs */
 387                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 388                                            cfs_time_seconds(1), NULL, NULL);
 389                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 390                 if (rc == 0) {
 391                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 392                         RETURN(1);
 393                 }
 394
 395                 LASSERT(rc == -ETIMEDOUT);
 396                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 397                           desc);
 398         }
 399         RETURN(0);
 400 }
 401 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
 402
 403 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 404 {
 405         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 406         struct ptlrpc_service           *svc = svcpt->scp_service;
 407         int service_time = max_t(int, cfs_time_current_sec() -
 408                                  req->rq_arrival_time.tv_sec, 1);
 409
 410         if (!(flags & PTLRPC_REPLY_EARLY) &&
 411             (req->rq_type != PTL_RPC_MSG_ERR) &&
 412             (req->rq_reqmsg != NULL) &&
 413             !(lustre_msg_get_flags(req->rq_reqmsg) &
 414               (MSG_RESENT | MSG_REPLAY |
 415                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 416                 /* early replies, errors and recovery requests don't count
 417                  * toward our service time estimate */
 418                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 419
 420                 if (oldse != 0) {
 421                         DEBUG_REQ(D_ADAPTTO, req,
 422                                   "svc %s changed estimate from %d to %d",
 423                                   svc->srv_name, oldse,
 424                                   at_get(&svcpt->scp_at_estimate));
 425                 }
 426         }
 427         /* Report actual service time for client latency calc */
 428         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 429         /* Report service time estimate for future client reqs, but report 0
 430          * (to be ignored by client) if it's a error reply during recovery.
 431          * (bz15815) */
 432         if (req->rq_type == PTL_RPC_MSG_ERR &&
 433             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 434                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 435         else
 436                 lustre_msg_set_timeout(req->rq_repmsg,
 437                                        at_get(&svcpt->scp_at_estimate));
 438
 439         if (req->rq_reqmsg &&
 440             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 441                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 442                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 443                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 444                        lustre_msg_is_v1(req->rq_reqmsg),
 445                        lustre_msg_get_magic(req->rq_reqmsg),
 446                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 447         }
 448 }
 449
 450 /**
 451  * Send request reply from request \a req reply buffer.
 452  * \a flags defines reply types
 453  * Returns 0 on sucess or error code
 454  */
 455 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 456 {
 457         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 458         struct ptlrpc_connection  *conn;
 459         int                        rc;
 460
 461         /* We must already have a reply buffer (only ptlrpc_error() may be
 462          * called without one). The reply generated by sptlrpc layer (e.g.
 463          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 464          * have a request buffer which is either the actual (swabbed) incoming
 465          * request, or a saved copy if this is a req saved in
 466          * target_queue_final_reply().
 467          */
 468         LASSERT (req->rq_no_reply == 0);
 469         LASSERT (req->rq_reqbuf != NULL);
 470         LASSERT (rs != NULL);
 471         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 472         LASSERT (req->rq_repmsg != NULL);
 473         LASSERT (req->rq_repmsg == rs->rs_msg);
 474         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 475         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 476
 477         /* There may be no rq_export during failover */
 478
 479         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 480                      req->rq_export->exp_obd->obd_fail)) {
 481                 /* Failed obd's only send ENODEV */
 482                 req->rq_type = PTL_RPC_MSG_ERR;
 483                 req->rq_status = -ENODEV;
 484                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 485                        req->rq_export->exp_obd->obd_minor);
 486         }
 487
 488         /* In order to keep interoprability with the client (< 2.3) which
 489          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 490          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 491          * reply buffer on client will be overflow.
 492          *
 493          * XXX Remove this whenver we drop the interoprability with such client.
 494          */
 495         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 496                                            sizeof(struct ptlrpc_body_v2), 1);
 497
 498         if (req->rq_type != PTL_RPC_MSG_ERR)
 499                 req->rq_type = PTL_RPC_MSG_REPLY;
 500
 501         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 502         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 503         lustre_msg_set_opc(req->rq_repmsg,
 504                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 505
 506         target_pack_pool_reply(req);
 507
 508         ptlrpc_at_set_reply(req, flags);
 509
 510         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 511                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 512         else
 513                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 514
 515         if (unlikely(conn == NULL)) {
 516                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 517                 return -ENOTCONN;
 518         }
 519         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 520
 521         rc = sptlrpc_svc_wrap_reply(req);
 522         if (unlikely(rc))
 523                 goto out;
 524
 525         req->rq_sent = cfs_time_current_sec();
 526
 527         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 528                            (rs->rs_difficult && !rs->rs_no_ack) ?
 529                            LNET_ACK_REQ : LNET_NOACK_REQ,
 530                            &rs->rs_cb_id, conn,
 531                            ptlrpc_req2svc(req)->srv_rep_portal,
 532                            req->rq_xid, req->rq_reply_off);
 533 out:
 534         if (unlikely(rc != 0))
 535                 ptlrpc_req_drop_rs(req);
 536         ptlrpc_connection_put(conn);
 537         return rc;
 538 }
 539 EXPORT_SYMBOL(ptlrpc_send_reply);
 540
 541 int ptlrpc_reply (struct ptlrpc_request *req)
 542 {
 543         if (req->rq_no_reply)
 544                 return 0;
 545         else
 546                 return (ptlrpc_send_reply(req, 0));
 547 }
 548 EXPORT_SYMBOL(ptlrpc_reply);
 549
 550 /**
 551  * For request \a req send an error reply back. Create empty
 552  * reply buffers if necessary.
 553  */
 554 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 555 {
 556         int rc;
 557         ENTRY;
 558
 559         if (req->rq_no_reply)
 560                 RETURN(0);
 561
 562         if (!req->rq_repmsg) {
 563                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 564                 if (rc)
 565                         RETURN(rc);
 566         }
 567
 568         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 569             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 570             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 571                 req->rq_type = PTL_RPC_MSG_ERR;
 572
 573         rc = ptlrpc_send_reply(req, may_be_difficult);
 574         RETURN(rc);
 575 }
 576 EXPORT_SYMBOL(ptlrpc_send_error);
 577
 578 int ptlrpc_error(struct ptlrpc_request *req)
 579 {
 580         return ptlrpc_send_error(req, 0);
 581 }
 582 EXPORT_SYMBOL(ptlrpc_error);
 583
 584 /**
 585  * Send request \a request.
 586  * if \a noreply is set, don't expect any reply back and don't set up
 587  * reply buffers.
 588  * Returns 0 on success or error code.
 589  */
 590 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 591 {
 592         int rc;
 593         int rc2;
 594         int mpflag = 0;
 595         struct ptlrpc_connection *connection;
 596         lnet_handle_me_t  reply_me_h;
 597         lnet_md_t         reply_md;
 598         struct obd_device *obd = request->rq_import->imp_obd;
 599         ENTRY;
 600
 601         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 602                 RETURN(0);
 603
 604         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 605         LASSERT(request->rq_wait_ctx == 0);
 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_import->imp_obd &&
 612             request->rq_import->imp_obd->obd_fail) {
 613                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 614                        request->rq_import->imp_obd->obd_name);
 615                 /* this prevents us from waiting in ptlrpc_queue_wait */
 616                 request->rq_err = 1;
 617                 request->rq_status = -ENODEV;
 618                 RETURN(-ENODEV);
 619         }
 620
 621         connection = request->rq_import->imp_connection;
 622
 623         lustre_msg_set_handle(request->rq_reqmsg,
 624                               &request->rq_import->imp_remote_handle);
 625         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 626         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 627                                 request->rq_import->imp_conn_cnt);
 628         lustre_msghdr_set_flags(request->rq_reqmsg,
 629                                 request->rq_import->imp_msghdr_flags);
 630
 631         if (request->rq_resend)
 632                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 633
 634         if (request->rq_memalloc)
 635                 mpflag = cfs_memory_pressure_get_and_set();
 636
 637         rc = sptlrpc_cli_wrap_request(request);
 638         if (rc)
 639                 GOTO(out, rc);
 640
 641         /* bulk register should be done after wrap_request() */
 642         if (request->rq_bulk != NULL) {
 643                 rc = ptlrpc_register_bulk (request);
 644                 if (rc != 0)
 645                         GOTO(out, rc);
 646         }
 647
 648         if (!noreply) {
 649                 LASSERT (request->rq_replen != 0);
 650                 if (request->rq_repbuf == NULL) {
 651                         LASSERT(request->rq_repdata == NULL);
 652                         LASSERT(request->rq_repmsg == NULL);
 653                         rc = sptlrpc_cli_alloc_repbuf(request,
 654                                                       request->rq_replen);
 655                         if (rc) {
 656                                 /* this prevents us from looping in
 657                                  * ptlrpc_queue_wait */
 658                                 request->rq_err = 1;
 659                                 request->rq_status = rc;
 660                                 GOTO(cleanup_bulk, rc);
 661                         }
 662                 } else {
 663                         request->rq_repdata = NULL;
 664                         request->rq_repmsg = NULL;
 665                 }
 666
 667                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 668                                   connection->c_peer, request->rq_xid, 0,
 669                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 670                 if (rc != 0) {
 671                         CERROR("LNetMEAttach failed: %d\n", rc);
 672                         LASSERT (rc == -ENOMEM);
 673                         GOTO(cleanup_bulk, rc = -ENOMEM);
 674                 }
 675         }
 676
 677         spin_lock(&request->rq_lock);
 678         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 679         request->rq_receiving_reply = !noreply;
 680         /* We are responsible for unlinking the reply buffer */
 681         request->rq_must_unlink = !noreply;
 682         /* Clear any flags that may be present from previous sends. */
 683         request->rq_replied = 0;
 684         request->rq_err = 0;
 685         request->rq_timedout = 0;
 686         request->rq_net_err = 0;
 687         request->rq_resend = 0;
 688         request->rq_restart = 0;
 689         request->rq_reply_truncate = 0;
 690         spin_unlock(&request->rq_lock);
 691
 692         if (!noreply) {
 693                 reply_md.start     = request->rq_repbuf;
 694                 reply_md.length    = request->rq_repbuf_len;
 695                 /* Allow multiple early replies */
 696                 reply_md.threshold = LNET_MD_THRESH_INF;
 697                 /* Manage remote for early replies */
 698                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 699                         LNET_MD_MANAGE_REMOTE |
 700                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 701                 reply_md.user_ptr  = &request->rq_reply_cbid;
 702                 reply_md.eq_handle = ptlrpc_eq_h;
 703
 704                 /* We must see the unlink callback to unset rq_must_unlink,
 705                    so we can't auto-unlink */
 706                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 707                                   &request->rq_reply_md_h);
 708                 if (rc != 0) {
 709                         CERROR("LNetMDAttach failed: %d\n", rc);
 710                         LASSERT (rc == -ENOMEM);
 711                         spin_lock(&request->rq_lock);
 712                         /* ...but the MD attach didn't succeed... */
 713                         request->rq_receiving_reply = 0;
 714                         spin_unlock(&request->rq_lock);
 715                         GOTO(cleanup_me, rc = -ENOMEM);
 716                 }
 717
 718                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 719                        ", portal %u\n",
 720                        request->rq_repbuf_len, request->rq_xid,
 721                        request->rq_reply_portal);
 722         }
 723
 724         /* add references on request for request_out_callback */
 725         ptlrpc_request_addref(request);
 726         if (obd->obd_svc_stats != NULL)
 727                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 728                         cfs_atomic_read(&request->rq_import->imp_inflight));
 729
 730         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 731
 732         cfs_gettimeofday(&request->rq_arrival_time);
 733         request->rq_sent = cfs_time_current_sec();
 734         /* We give the server rq_timeout secs to process the req, and
 735            add the network latency for our local timeout. */
 736         request->rq_deadline = request->rq_sent + request->rq_timeout +
 737                 ptlrpc_at_get_net_latency(request);
 738
 739         ptlrpc_pinger_sending_on_import(request->rq_import);
 740
 741         DEBUG_REQ(D_INFO, request, "send flg=%x",
 742                   lustre_msg_get_flags(request->rq_reqmsg));
 743         rc = ptl_send_buf(&request->rq_req_md_h,
 744                           request->rq_reqbuf, request->rq_reqdata_len,
 745                           LNET_NOACK_REQ, &request->rq_req_cbid,
 746                           connection,
 747                           request->rq_request_portal,
 748                           request->rq_xid, 0);
 749         if (rc == 0)
 750                 GOTO(out, rc);
 751
 752         ptlrpc_req_finished(request);
 753         if (noreply)
 754                 GOTO(out, rc);
 755
 756  cleanup_me:
 757         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 758          * nobody apart from the PUT's target has the right nid+XID to
 759          * access the reply buffer. */
 760         rc2 = LNetMEUnlink(reply_me_h);
 761         LASSERT (rc2 == 0);
 762         /* UNLINKED callback called synchronously */
 763         LASSERT(!request->rq_receiving_reply);
 764
 765  cleanup_bulk:
 766         /* We do sync unlink here as there was no real transfer here so
 767          * the chance to have long unlink to sluggish net is smaller here. */
 768         ptlrpc_unregister_bulk(request, 0);
 769  out:
 770         if (request->rq_memalloc)
 771                 cfs_memory_pressure_restore(mpflag);
 772         return rc;
 773 }
 774 EXPORT_SYMBOL(ptl_send_rpc);
 775
 776 /**
 777  * Register request buffer descriptor for request receiving.
 778  */
 779 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 780 {
 781         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 782         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 783         int                       rc;
 784         lnet_md_t                 md;
 785         lnet_handle_me_t          me_h;
 786
 787         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 788                service->srv_req_portal);
 789
 790         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 791                 return (-ENOMEM);
 792
 793         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 794          * which means buffer can only be attached on local CPT, and LND
 795          * threads can find it by grabbing a local lock */
 796         rc = LNetMEAttach(service->srv_req_portal,
 797                           match_id, 0, ~0, LNET_UNLINK,
 798                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 799                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 800         if (rc != 0) {
 801                 CERROR("LNetMEAttach failed: %d\n", rc);
 802                 return (-ENOMEM);
 803         }
 804
 805         LASSERT(rqbd->rqbd_refcount == 0);
 806         rqbd->rqbd_refcount = 1;
 807
 808         md.start     = rqbd->rqbd_buffer;
 809         md.length    = service->srv_buf_size;
 810         md.max_size  = service->srv_max_req_size;
 811         md.threshold = LNET_MD_THRESH_INF;
 812         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 813         md.user_ptr  = &rqbd->rqbd_cbid;
 814         md.eq_handle = ptlrpc_eq_h;
 815
 816         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 817         if (rc == 0)
 818                 return (0);
 819
 820         CERROR("LNetMDAttach failed: %d; \n", rc);
 821         LASSERT (rc == -ENOMEM);
 822         rc = LNetMEUnlink (me_h);
 823         LASSERT (rc == 0);
 824         rqbd->rqbd_refcount = 0;
 825
 826         return (-ENOMEM);
 827 }