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) 2012, 2013, 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 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
 105 {
 106         int i;
 107
 108         for (i = 0; i < count; i++)
 109                 LNetMDUnlink(bd_mds[i]);
 110 }
 111
 112 #ifdef HAVE_SERVER_SUPPORT
 113 /**
 114  * Prepare bulk descriptor for specified incoming request \a req that
 115  * can fit \a npages * pages. \a type is bulk type. \a portal is where
 116  * the bulk to be sent. Used on server-side after request was already
 117  * received.
 118  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 119  * error.
 120  */
 121 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 122                                               unsigned npages, unsigned max_brw,
 123                                               unsigned type, unsigned portal)
 124 {
 125         struct obd_export *exp = req->rq_export;
 126         struct ptlrpc_bulk_desc *desc;
 127
 128         ENTRY;
 129         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 130
 131         desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
 132         if (desc == NULL)
 133                 RETURN(NULL);
 134
 135         desc->bd_export = class_export_get(exp);
 136         desc->bd_req = req;
 137
 138         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 139         desc->bd_cbid.cbid_arg = desc;
 140
 141         /* NB we don't assign rq_bulk here; server-side requests are
 142          * re-used, and the handler frees the bulk desc explicitly. */
 143
 144         return desc;
 145 }
 146 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 147
 148 /**
 149  * Starts bulk transfer for descriptor \a desc on the server.
 150  * Returns 0 on success or error code.
 151  */
 152 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 153 {
 154         struct obd_export        *exp = desc->bd_export;
 155         struct ptlrpc_connection *conn = exp->exp_connection;
 156         int                       rc = 0;
 157         __u64                     xid;
 158         int                       posted_md;
 159         int                       total_md;
 160         lnet_md_t                 md;
 161         ENTRY;
 162
 163         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 164                 RETURN(0);
 165
 166         /* NB no locking required until desc is on the network */
 167         LASSERT(desc->bd_md_count == 0);
 168         LASSERT(desc->bd_type == BULK_PUT_SOURCE ||
 169                 desc->bd_type == BULK_GET_SINK);
 170
 171         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 172         LASSERT(desc->bd_cbid.cbid_arg == desc);
 173
 174         /* NB total length may be 0 for a read past EOF, so we send 0
 175          * length bulks, since the client expects bulk events.
 176          *
 177          * The client may not need all of the bulk XIDs for the RPC.  The RPC
 178          * used the XID of the highest bulk XID needed, and the server masks
 179          * off high bits to get bulk count for this RPC. LU-1431 */
 180         xid = desc->bd_req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 181         total_md = desc->bd_req->rq_xid - xid + 1;
 182
 183         desc->bd_md_count = total_md;
 184         desc->bd_failure = 0;
 185
 186         md.user_ptr = &desc->bd_cbid;
 187         md.eq_handle = ptlrpc_eq_h;
 188         md.threshold = 2; /* SENT and ACK/REPLY */
 189
 190         for (posted_md = 0; posted_md < total_md; xid++) {
 191                 md.options = PTLRPC_MD_OPTIONS;
 192
 193                 /* NB it's assumed that source and sink buffer frags are
 194                  * page-aligned. Otherwise we'd have to send client bulk
 195                  * sizes over and split server buffer accordingly */
 196                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 197                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 198                 if (rc != 0) {
 199                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 200                                exp->exp_obd->obd_name, posted_md, rc);
 201                         LASSERT(rc == -ENOMEM);
 202                         if (posted_md == 0) {
 203                                 desc->bd_md_count = 0;
 204                                 RETURN(-ENOMEM);
 205                         }
 206                         break;
 207                 }
 208                 /* Network is about to get at the memory */
 209                 if (desc->bd_type == BULK_PUT_SOURCE)
 210                         rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
 211                                      LNET_ACK_REQ, conn->c_peer,
 212                                      desc->bd_portal, xid, 0, 0);
 213                 else
 214                         rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
 215                                      conn->c_peer, desc->bd_portal, xid, 0);
 216
 217                 posted_md++;
 218                 if (rc != 0) {
 219                         CERROR("%s: failed bulk transfer with %s:%u x"LPU64": "
 220                                "rc = %d\n", exp->exp_obd->obd_name,
 221                                libcfs_id2str(conn->c_peer), desc->bd_portal,
 222                                xid, rc);
 223                         break;
 224                 }
 225         }
 226
 227         if (rc != 0) {
 228                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 229                  * event this creates will signal completion with failure,
 230                  * so we return SUCCESS here! */
 231                 spin_lock(&desc->bd_lock);
 232                 desc->bd_md_count -= total_md - posted_md;
 233                 spin_unlock(&desc->bd_lock);
 234                 LASSERT(desc->bd_md_count >= 0);
 235
 236                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 237                 RETURN(0);
 238         }
 239
 240         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 241                "id %s xid "LPX64"-"LPX64"\n", desc->bd_iov_count,
 242                desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
 243                xid - posted_md, xid - 1);
 244
 245         RETURN(0);
 246 }
 247 EXPORT_SYMBOL(ptlrpc_start_bulk_transfer);
 248
 249 /**
 250  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 251  * serialises with completion callback)
 252  */
 253 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 254 {
 255         struct l_wait_info       lwi;
 256         int                      rc;
 257
 258         LASSERT(!in_interrupt());           /* might sleep */
 259
 260         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 261                 return;                         /* never started */
 262
 263         /* We used to poison the pages with 0xab here because we did not want to
 264          * send any meaningful data over the wire for evicted clients (bug 9297)
 265          * However, this is no longer safe now that we use the page cache on the
 266          * OSS (bug 20560) */
 267
 268         /* The unlink ensures the callback happens ASAP and is the last
 269          * one.  If it fails, it must be because completion just happened,
 270          * but we must still l_wait_event() in this case, to give liblustre
 271          * a chance to run server_bulk_callback()*/
 272         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_count);
 273
 274         for (;;) {
 275                 /* Network access will complete in finite time but the HUGE
 276                  * timeout lets us CWARN for visibility of sluggish NALs */
 277                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 278                                            cfs_time_seconds(1), NULL, NULL);
 279                 rc = l_wait_event(desc->bd_waitq,
 280                                   !ptlrpc_server_bulk_active(desc), &lwi);
 281                 if (rc == 0)
 282                         return;
 283
 284                 LASSERT(rc == -ETIMEDOUT);
 285                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 286         }
 287 }
 288 EXPORT_SYMBOL(ptlrpc_abort_bulk);
 289 #endif /* HAVE_SERVER_SUPPORT */
 290
 291 /**
 292  * Register bulk at the sender for later transfer.
 293  * Returns 0 on success or error code.
 294  */
 295 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 296 {
 297         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 298         lnet_process_id_t peer;
 299         int rc = 0;
 300         int rc2;
 301         int posted_md;
 302         int total_md;
 303         __u64 xid;
 304         lnet_handle_me_t  me_h;
 305         lnet_md_t         md;
 306         ENTRY;
 307
 308         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 309                 RETURN(0);
 310
 311         /* NB no locking required until desc is on the network */
 312         LASSERT(desc->bd_nob > 0);
 313         LASSERT(desc->bd_md_count == 0);
 314         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 315         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 316         LASSERT(desc->bd_req != NULL);
 317         LASSERT(desc->bd_type == BULK_PUT_SINK ||
 318                 desc->bd_type == BULK_GET_SOURCE);
 319
 320         /* cleanup the state of the bulk for it will be reused */
 321         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 322                 desc->bd_nob_transferred = 0;
 323         else
 324                 LASSERT(desc->bd_nob_transferred == 0);
 325
 326         desc->bd_failure = 0;
 327
 328         peer = desc->bd_import->imp_connection->c_peer;
 329
 330         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 331         LASSERT(desc->bd_cbid.cbid_arg == desc);
 332
 333         /* An XID is only used for a single request from the client.
 334          * For retried bulk transfers, a new XID will be allocated in
 335          * in ptlrpc_check_set() if it needs to be resent, so it is not
 336          * using the same RDMA match bits after an error.
 337          *
 338          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 339          * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
 340         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 341         LASSERTF(!(desc->bd_registered &&
 342                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 343                  xid != desc->bd_last_xid,
 344                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 345                  desc->bd_registered, xid, desc->bd_last_xid);
 346
 347         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 348         desc->bd_registered = 1;
 349         desc->bd_last_xid = xid;
 350         desc->bd_md_count = total_md;
 351         md.user_ptr = &desc->bd_cbid;
 352         md.eq_handle = ptlrpc_eq_h;
 353         md.threshold = 1;                       /* PUT or GET */
 354
 355         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 356                 md.options = PTLRPC_MD_OPTIONS |
 357                              ((desc->bd_type == BULK_GET_SOURCE) ?
 358                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 359                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 360
 361                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 362                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 363                 if (rc != 0) {
 364                         CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
 365                                desc->bd_import->imp_obd->obd_name, xid,
 366                                posted_md, rc);
 367                         break;
 368                 }
 369
 370                 /* About to let the network at it... */
 371                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 372                                   &desc->bd_mds[posted_md]);
 373                 if (rc != 0) {
 374                         CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
 375                                desc->bd_import->imp_obd->obd_name, xid,
 376                                posted_md, rc);
 377                         rc2 = LNetMEUnlink(me_h);
 378                         LASSERT(rc2 == 0);
 379                         break;
 380                 }
 381         }
 382
 383         if (rc != 0) {
 384                 LASSERT(rc == -ENOMEM);
 385                 spin_lock(&desc->bd_lock);
 386                 desc->bd_md_count -= total_md - posted_md;
 387                 spin_unlock(&desc->bd_lock);
 388                 LASSERT(desc->bd_md_count >= 0);
 389                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 390                 req->rq_status = -ENOMEM;
 391                 RETURN(-ENOMEM);
 392         }
 393
 394         /* Set rq_xid to matchbits of the final bulk so that server can
 395          * infer the number of bulks that were prepared */
 396         req->rq_xid = --xid;
 397         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 398                  "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n",
 399                  desc->bd_last_xid, req->rq_xid);
 400
 401         spin_lock(&desc->bd_lock);
 402         /* Holler if peer manages to touch buffers before he knows the xid */
 403         if (desc->bd_md_count != total_md)
 404                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 405                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 406                       total_md - desc->bd_md_count);
 407         spin_unlock(&desc->bd_lock);
 408
 409         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 410                "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
 411                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 412                desc->bd_iov_count, desc->bd_nob,
 413                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 414
 415         RETURN(0);
 416 }
 417 EXPORT_SYMBOL(ptlrpc_register_bulk);
 418
 419 /**
 420  * Disconnect a bulk desc from the network. Idempotent. Not
 421  * thread-safe (i.e. only interlocks with completion callback).
 422  * Returns 1 on success or 0 if network unregistration failed for whatever
 423  * reason.
 424  */
 425 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 426 {
 427         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 428         struct l_wait_info       lwi;
 429         int                      rc;
 430         ENTRY;
 431
 432         LASSERT(!in_interrupt());     /* might sleep */
 433
 434         /* Let's setup deadline for reply unlink. */
 435         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 436             async && req->rq_bulk_deadline == 0)
 437                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 438
 439         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 440                 RETURN(1);                              /* never registered */
 441
 442         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 443
 444         /* the unlink ensures the callback happens ASAP and is the last
 445          * one.  If it fails, it must be because completion just happened,
 446          * but we must still l_wait_event() in this case to give liblustre
 447          * a chance to run client_bulk_callback() */
 448         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 449
 450         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 451                 RETURN(1);                              /* never registered */
 452
 453         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 454         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 455
 456         /* Do not wait for unlink to finish. */
 457         if (async)
 458                 RETURN(0);
 459
 460         for (;;) {
 461 #ifdef __KERNEL__
 462                 /* The wq argument is ignored by user-space wait_event macros */
 463                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 464                                         &req->rq_set->set_waitq :
 465                                         &req->rq_reply_waitq;
 466 #endif
 467                 /* Network access will complete in finite time but the HUGE
 468                  * timeout lets us CWARN for visibility of sluggish NALs */
 469                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 470                                            cfs_time_seconds(1), NULL, NULL);
 471                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 472                 if (rc == 0) {
 473                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 474                         RETURN(1);
 475                 }
 476
 477                 LASSERT(rc == -ETIMEDOUT);
 478                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 479                           desc);
 480         }
 481         RETURN(0);
 482 }
 483 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
 484
 485 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 486 {
 487         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 488         struct ptlrpc_service           *svc = svcpt->scp_service;
 489         int service_time = max_t(int, cfs_time_current_sec() -
 490                                  req->rq_arrival_time.tv_sec, 1);
 491
 492         if (!(flags & PTLRPC_REPLY_EARLY) &&
 493             (req->rq_type != PTL_RPC_MSG_ERR) &&
 494             (req->rq_reqmsg != NULL) &&
 495             !(lustre_msg_get_flags(req->rq_reqmsg) &
 496               (MSG_RESENT | MSG_REPLAY |
 497                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 498                 /* early replies, errors and recovery requests don't count
 499                  * toward our service time estimate */
 500                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 501
 502                 if (oldse != 0) {
 503                         DEBUG_REQ(D_ADAPTTO, req,
 504                                   "svc %s changed estimate from %d to %d",
 505                                   svc->srv_name, oldse,
 506                                   at_get(&svcpt->scp_at_estimate));
 507                 }
 508         }
 509         /* Report actual service time for client latency calc */
 510         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 511         /* Report service time estimate for future client reqs, but report 0
 512          * (to be ignored by client) if it's a error reply during recovery.
 513          * (bz15815) */
 514         if (req->rq_type == PTL_RPC_MSG_ERR &&
 515             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 516                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 517         else
 518                 lustre_msg_set_timeout(req->rq_repmsg,
 519                                        at_get(&svcpt->scp_at_estimate));
 520
 521         if (req->rq_reqmsg &&
 522             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 523                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 524                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 525                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 526                        lustre_msg_is_v1(req->rq_reqmsg),
 527                        lustre_msg_get_magic(req->rq_reqmsg),
 528                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 529         }
 530 }
 531
 532 /**
 533  * Send request reply from request \a req reply buffer.
 534  * \a flags defines reply types
 535  * Returns 0 on sucess or error code
 536  */
 537 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 538 {
 539         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 540         struct ptlrpc_connection  *conn;
 541         int                        rc;
 542
 543         /* We must already have a reply buffer (only ptlrpc_error() may be
 544          * called without one). The reply generated by sptlrpc layer (e.g.
 545          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 546          * have a request buffer which is either the actual (swabbed) incoming
 547          * request, or a saved copy if this is a req saved in
 548          * target_queue_final_reply().
 549          */
 550         LASSERT (req->rq_no_reply == 0);
 551         LASSERT (req->rq_reqbuf != NULL);
 552         LASSERT (rs != NULL);
 553         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 554         LASSERT (req->rq_repmsg != NULL);
 555         LASSERT (req->rq_repmsg == rs->rs_msg);
 556         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 557         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 558
 559         /* There may be no rq_export during failover */
 560
 561         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 562                      req->rq_export->exp_obd->obd_fail)) {
 563                 /* Failed obd's only send ENODEV */
 564                 req->rq_type = PTL_RPC_MSG_ERR;
 565                 req->rq_status = -ENODEV;
 566                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 567                        req->rq_export->exp_obd->obd_minor);
 568         }
 569
 570         /* In order to keep interoprability with the client (< 2.3) which
 571          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 572          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 573          * reply buffer on client will be overflow.
 574          *
 575          * XXX Remove this whenver we drop the interoprability with such client.
 576          */
 577         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 578                                            sizeof(struct ptlrpc_body_v2), 1);
 579
 580         if (req->rq_type != PTL_RPC_MSG_ERR)
 581                 req->rq_type = PTL_RPC_MSG_REPLY;
 582
 583         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 584         lustre_msg_set_status(req->rq_repmsg,
 585                               ptlrpc_status_hton(req->rq_status));
 586         lustre_msg_set_opc(req->rq_repmsg,
 587                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 588
 589         target_pack_pool_reply(req);
 590
 591         ptlrpc_at_set_reply(req, flags);
 592
 593         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 594                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 595         else
 596                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 597
 598         if (unlikely(conn == NULL)) {
 599                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 600                 return -ENOTCONN;
 601         }
 602         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 603
 604         rc = sptlrpc_svc_wrap_reply(req);
 605         if (unlikely(rc))
 606                 goto out;
 607
 608         req->rq_sent = cfs_time_current_sec();
 609
 610         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 611                            (rs->rs_difficult && !rs->rs_no_ack) ?
 612                            LNET_ACK_REQ : LNET_NOACK_REQ,
 613                            &rs->rs_cb_id, conn,
 614                            ptlrpc_req2svc(req)->srv_rep_portal,
 615                            req->rq_xid, req->rq_reply_off);
 616 out:
 617         if (unlikely(rc != 0))
 618                 ptlrpc_req_drop_rs(req);
 619         ptlrpc_connection_put(conn);
 620         return rc;
 621 }
 622 EXPORT_SYMBOL(ptlrpc_send_reply);
 623
 624 int ptlrpc_reply (struct ptlrpc_request *req)
 625 {
 626         if (req->rq_no_reply)
 627                 return 0;
 628         else
 629                 return (ptlrpc_send_reply(req, 0));
 630 }
 631 EXPORT_SYMBOL(ptlrpc_reply);
 632
 633 /**
 634  * For request \a req send an error reply back. Create empty
 635  * reply buffers if necessary.
 636  */
 637 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 638 {
 639         int rc;
 640         ENTRY;
 641
 642         if (req->rq_no_reply)
 643                 RETURN(0);
 644
 645         if (!req->rq_repmsg) {
 646                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 647                 if (rc)
 648                         RETURN(rc);
 649         }
 650
 651         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 652             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 653             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 654                 req->rq_type = PTL_RPC_MSG_ERR;
 655
 656         rc = ptlrpc_send_reply(req, may_be_difficult);
 657         RETURN(rc);
 658 }
 659 EXPORT_SYMBOL(ptlrpc_send_error);
 660
 661 int ptlrpc_error(struct ptlrpc_request *req)
 662 {
 663         return ptlrpc_send_error(req, 0);
 664 }
 665 EXPORT_SYMBOL(ptlrpc_error);
 666
 667 /**
 668  * Send request \a request.
 669  * if \a noreply is set, don't expect any reply back and don't set up
 670  * reply buffers.
 671  * Returns 0 on success or error code.
 672  */
 673 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 674 {
 675         int rc;
 676         int rc2;
 677         int mpflag = 0;
 678         struct ptlrpc_connection *connection;
 679         lnet_handle_me_t  reply_me_h;
 680         lnet_md_t         reply_md;
 681         struct obd_device *obd = request->rq_import->imp_obd;
 682         ENTRY;
 683
 684         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 685                 RETURN(0);
 686
 687         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 688         LASSERT(request->rq_wait_ctx == 0);
 689
 690         /* If this is a re-transmit, we're required to have disengaged
 691          * cleanly from the previous attempt */
 692         LASSERT(!request->rq_receiving_reply);
 693
 694         if (request->rq_import->imp_obd &&
 695             request->rq_import->imp_obd->obd_fail) {
 696                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 697                        request->rq_import->imp_obd->obd_name);
 698                 /* this prevents us from waiting in ptlrpc_queue_wait */
 699                 spin_lock(&request->rq_lock);
 700                 request->rq_err = 1;
 701                 spin_unlock(&request->rq_lock);
 702                 request->rq_status = -ENODEV;
 703                 RETURN(-ENODEV);
 704         }
 705
 706         connection = request->rq_import->imp_connection;
 707
 708         lustre_msg_set_handle(request->rq_reqmsg,
 709                               &request->rq_import->imp_remote_handle);
 710         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 711         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 712                                 request->rq_import->imp_conn_cnt);
 713         lustre_msghdr_set_flags(request->rq_reqmsg,
 714                                 request->rq_import->imp_msghdr_flags);
 715
 716         if (request->rq_resend)
 717                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 718
 719         if (request->rq_memalloc)
 720                 mpflag = cfs_memory_pressure_get_and_set();
 721
 722         rc = sptlrpc_cli_wrap_request(request);
 723         if (rc)
 724                 GOTO(out, rc);
 725
 726         /* bulk register should be done after wrap_request() */
 727         if (request->rq_bulk != NULL) {
 728                 rc = ptlrpc_register_bulk (request);
 729                 if (rc != 0)
 730                         GOTO(out, rc);
 731         }
 732
 733         if (!noreply) {
 734                 LASSERT (request->rq_replen != 0);
 735                 if (request->rq_repbuf == NULL) {
 736                         LASSERT(request->rq_repdata == NULL);
 737                         LASSERT(request->rq_repmsg == NULL);
 738                         rc = sptlrpc_cli_alloc_repbuf(request,
 739                                                       request->rq_replen);
 740                         if (rc) {
 741                                 /* this prevents us from looping in
 742                                  * ptlrpc_queue_wait */
 743                                 spin_lock(&request->rq_lock);
 744                                 request->rq_err = 1;
 745                                 spin_unlock(&request->rq_lock);
 746                                 request->rq_status = rc;
 747                                 GOTO(cleanup_bulk, rc);
 748                         }
 749                 } else {
 750                         request->rq_repdata = NULL;
 751                         request->rq_repmsg = NULL;
 752                 }
 753
 754                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 755                                   connection->c_peer, request->rq_xid, 0,
 756                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 757                 if (rc != 0) {
 758                         CERROR("LNetMEAttach failed: %d\n", rc);
 759                         LASSERT (rc == -ENOMEM);
 760                         GOTO(cleanup_bulk, rc = -ENOMEM);
 761                 }
 762         }
 763
 764         spin_lock(&request->rq_lock);
 765         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 766         request->rq_receiving_reply = !noreply;
 767         /* We are responsible for unlinking the reply buffer */
 768         request->rq_must_unlink = !noreply;
 769         /* Clear any flags that may be present from previous sends. */
 770         request->rq_replied = 0;
 771         request->rq_err = 0;
 772         request->rq_timedout = 0;
 773         request->rq_net_err = 0;
 774         request->rq_resend = 0;
 775         request->rq_restart = 0;
 776         request->rq_reply_truncate = 0;
 777         spin_unlock(&request->rq_lock);
 778
 779         if (!noreply) {
 780                 reply_md.start     = request->rq_repbuf;
 781                 reply_md.length    = request->rq_repbuf_len;
 782                 /* Allow multiple early replies */
 783                 reply_md.threshold = LNET_MD_THRESH_INF;
 784                 /* Manage remote for early replies */
 785                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 786                         LNET_MD_MANAGE_REMOTE |
 787                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 788                 reply_md.user_ptr  = &request->rq_reply_cbid;
 789                 reply_md.eq_handle = ptlrpc_eq_h;
 790
 791                 /* We must see the unlink callback to unset rq_must_unlink,
 792                    so we can't auto-unlink */
 793                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 794                                   &request->rq_reply_md_h);
 795                 if (rc != 0) {
 796                         CERROR("LNetMDAttach failed: %d\n", rc);
 797                         LASSERT (rc == -ENOMEM);
 798                         spin_lock(&request->rq_lock);
 799                         /* ...but the MD attach didn't succeed... */
 800                         request->rq_receiving_reply = 0;
 801                         spin_unlock(&request->rq_lock);
 802                         GOTO(cleanup_me, rc = -ENOMEM);
 803                 }
 804
 805                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 806                        ", portal %u\n",
 807                        request->rq_repbuf_len, request->rq_xid,
 808                        request->rq_reply_portal);
 809         }
 810
 811         /* add references on request for request_out_callback */
 812         ptlrpc_request_addref(request);
 813         if (obd->obd_svc_stats != NULL)
 814                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 815                         cfs_atomic_read(&request->rq_import->imp_inflight));
 816
 817         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 818
 819         do_gettimeofday(&request->rq_arrival_time);
 820         request->rq_sent = cfs_time_current_sec();
 821         /* We give the server rq_timeout secs to process the req, and
 822            add the network latency for our local timeout. */
 823         request->rq_deadline = request->rq_sent + request->rq_timeout +
 824                 ptlrpc_at_get_net_latency(request);
 825
 826         ptlrpc_pinger_sending_on_import(request->rq_import);
 827
 828         DEBUG_REQ(D_INFO, request, "send flg=%x",
 829                   lustre_msg_get_flags(request->rq_reqmsg));
 830         rc = ptl_send_buf(&request->rq_req_md_h,
 831                           request->rq_reqbuf, request->rq_reqdata_len,
 832                           LNET_NOACK_REQ, &request->rq_req_cbid,
 833                           connection,
 834                           request->rq_request_portal,
 835                           request->rq_xid, 0);
 836         if (rc == 0)
 837                 GOTO(out, rc);
 838
 839         ptlrpc_req_finished(request);
 840         if (noreply)
 841                 GOTO(out, rc);
 842
 843  cleanup_me:
 844         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 845          * nobody apart from the PUT's target has the right nid+XID to
 846          * access the reply buffer. */
 847         rc2 = LNetMEUnlink(reply_me_h);
 848         LASSERT (rc2 == 0);
 849         /* UNLINKED callback called synchronously */
 850         LASSERT(!request->rq_receiving_reply);
 851
 852  cleanup_bulk:
 853         /* We do sync unlink here as there was no real transfer here so
 854          * the chance to have long unlink to sluggish net is smaller here. */
 855         ptlrpc_unregister_bulk(request, 0);
 856  out:
 857         if (request->rq_memalloc)
 858                 cfs_memory_pressure_restore(mpflag);
 859         return rc;
 860 }
 861 EXPORT_SYMBOL(ptl_send_rpc);
 862
 863 /**
 864  * Register request buffer descriptor for request receiving.
 865  */
 866 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 867 {
 868         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 869         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 870         int                       rc;
 871         lnet_md_t                 md;
 872         lnet_handle_me_t          me_h;
 873
 874         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 875                service->srv_req_portal);
 876
 877         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 878                 return (-ENOMEM);
 879
 880         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 881          * which means buffer can only be attached on local CPT, and LND
 882          * threads can find it by grabbing a local lock */
 883         rc = LNetMEAttach(service->srv_req_portal,
 884                           match_id, 0, ~0, LNET_UNLINK,
 885                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 886                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 887         if (rc != 0) {
 888                 CERROR("LNetMEAttach failed: %d\n", rc);
 889                 return (-ENOMEM);
 890         }
 891
 892         LASSERT(rqbd->rqbd_refcount == 0);
 893         rqbd->rqbd_refcount = 1;
 894
 895         md.start     = rqbd->rqbd_buffer;
 896         md.length    = service->srv_buf_size;
 897         md.max_size  = service->srv_max_req_size;
 898         md.threshold = LNET_MD_THRESH_INF;
 899         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 900         md.user_ptr  = &rqbd->rqbd_cbid;
 901         md.eq_handle = ptlrpc_eq_h;
 902
 903         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 904         if (rc == 0)
 905                 return (0);
 906
 907         CERROR("LNetMDAttach failed: %d; \n", rc);
 908         LASSERT (rc == -ENOMEM);
 909         rc = LNetMEUnlink (me_h);
 910         LASSERT (rc == 0);
 911         rqbd->rqbd_refcount = 0;
 912
 913         return (-ENOMEM);
 914 }