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