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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_RPC
  34 #include <obd_support.h>
  35 #include <lustre_net.h>
  36 #include <lustre_lib.h>
  37 #include <obd.h>
  38 #include <obd_class.h>
  39 #include "ptlrpc_internal.h"
  40
  41 /**
  42  * Helper function. Sends \a len bytes from \a base at offset \a offset
  43  * over \a conn connection to portal \a portal.
  44  * Returns 0 on success or error code.
  45  */
  46 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  47                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  48                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  49                          unsigned int offset)
  50 {
  51         int              rc;
  52         lnet_md_t         md;
  53         ENTRY;
  54
  55         LASSERT (portal != 0);
  56         LASSERT (conn != NULL);
  57         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  58         md.start     = base;
  59         md.length    = len;
  60         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  61         md.options   = PTLRPC_MD_OPTIONS;
  62         md.user_ptr  = cbid;
  63         md.eq_handle = ptlrpc_eq_h;
  64
  65         if (unlikely(ack == LNET_ACK_REQ &&
  66                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  67                 /* don't ask for the ack to simulate failing client */
  68                 ack = LNET_NOACK_REQ;
  69         }
  70
  71         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  72         if (unlikely(rc != 0)) {
  73                 CERROR ("LNetMDBind failed: %d\n", rc);
  74                 LASSERT (rc == -ENOMEM);
  75                 RETURN (-ENOMEM);
  76         }
  77
  78         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  79                len, portal, xid, offset);
  80
  81         rc = LNetPut (conn->c_self, *mdh, ack,
  82                       conn->c_peer, portal, xid, offset, 0);
  83         if (unlikely(rc != 0)) {
  84                 int rc2;
  85                 /* We're going to get an UNLINK event when I unlink below,
  86                  * which will complete just like any other failed send, so
  87                  * I fall through and return success here! */
  88                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  89                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  90                 rc2 = LNetMDUnlink(*mdh);
  91                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  92         }
  93
  94         RETURN (0);
  95 }
  96
  97 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
  98 {
  99         int i;
 100
 101         for (i = 0; i < count; i++)
 102                 LNetMDUnlink(bd_mds[i]);
 103 }
 104
 105 #ifdef HAVE_SERVER_SUPPORT
 106 /**
 107  * Prepare bulk descriptor for specified incoming request \a req that
 108  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 109  * the bulk to be sent. Used on server-side after request was already
 110  * received.
 111  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 112  * error.
 113  */
 114 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 115                                               unsigned nfrags, unsigned max_brw,
 116                                               unsigned int type,
 117                                               unsigned portal,
 118                                               const struct ptlrpc_bulk_frag_ops
 119                                                 *ops)
 120 {
 121         struct obd_export *exp = req->rq_export;
 122         struct ptlrpc_bulk_desc *desc;
 123
 124         ENTRY;
 125         LASSERT(ptlrpc_is_bulk_op_active(type));
 126
 127         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 128         if (desc == NULL)
 129                 RETURN(NULL);
 130
 131         desc->bd_export = class_export_get(exp);
 132         desc->bd_req = req;
 133
 134         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 135         desc->bd_cbid.cbid_arg = desc;
 136
 137         /* NB we don't assign rq_bulk here; server-side requests are
 138          * re-used, and the handler frees the bulk desc explicitly. */
 139
 140         return desc;
 141 }
 142 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 143
 144 /**
 145  * Starts bulk transfer for descriptor \a desc on the server.
 146  * Returns 0 on success or error code.
 147  */
 148 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 149 {
 150         struct obd_export        *exp = desc->bd_export;
 151         struct ptlrpc_connection *conn = exp->exp_connection;
 152         int                       rc = 0;
 153         __u64                     mbits;
 154         int                       posted_md;
 155         int                       total_md;
 156         lnet_md_t                 md;
 157         ENTRY;
 158
 159         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 160                 RETURN(0);
 161
 162         /* NB no locking required until desc is on the network */
 163         LASSERT(desc->bd_md_count == 0);
 164         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 165
 166         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 167         LASSERT(desc->bd_cbid.cbid_arg == desc);
 168
 169         /* NB total length may be 0 for a read past EOF, so we send 0
 170          * length bulks, since the client expects bulk events.
 171          *
 172          * The client may not need all of the bulk mbits for the RPC. The RPC
 173          * used the mbits of the highest bulk mbits needed, and the server masks
 174          * off high bits to get bulk count for this RPC. LU-1431 */
 175         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 176         total_md = desc->bd_req->rq_mbits - mbits + 1;
 177
 178         desc->bd_md_count = total_md;
 179         desc->bd_failure = 0;
 180
 181         md.user_ptr = &desc->bd_cbid;
 182         md.eq_handle = ptlrpc_eq_h;
 183         md.threshold = 2; /* SENT and ACK/REPLY */
 184
 185         for (posted_md = 0; posted_md < total_md; mbits++) {
 186                 md.options = PTLRPC_MD_OPTIONS;
 187
 188                 /* NB it's assumed that source and sink buffer frags are
 189                  * page-aligned. Otherwise we'd have to send client bulk
 190                  * sizes over and split server buffer accordingly */
 191                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 192                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 193                 if (rc != 0) {
 194                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 195                                exp->exp_obd->obd_name, posted_md, rc);
 196                         LASSERT(rc == -ENOMEM);
 197                         if (posted_md == 0) {
 198                                 desc->bd_md_count = 0;
 199                                 RETURN(-ENOMEM);
 200                         }
 201                         break;
 202                 }
 203
 204                 /* LU-6441: last md is not sent and desc->bd_md_count == 1 */
 205                 if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 206                                          CFS_FAIL_ONCE) &&
 207                     total_md > 1 && posted_md == total_md - 1) {
 208                         posted_md++;
 209                         continue;
 210                 }
 211
 212                 /* Network is about to get at the memory */
 213                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 214                         rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
 215                                      LNET_ACK_REQ, conn->c_peer,
 216                                      desc->bd_portal, mbits, 0, 0);
 217                 else
 218                         rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
 219                                      conn->c_peer, desc->bd_portal, mbits, 0);
 220
 221                 posted_md++;
 222                 if (rc != 0) {
 223                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 224                                "rc = %d\n", exp->exp_obd->obd_name,
 225                                libcfs_id2str(conn->c_peer), desc->bd_portal,
 226                                mbits, rc);
 227                         break;
 228                 }
 229         }
 230
 231         if (rc != 0) {
 232                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 233                  * event this creates will signal completion with failure,
 234                  * so we return SUCCESS here! */
 235                 spin_lock(&desc->bd_lock);
 236                 desc->bd_md_count -= total_md - posted_md;
 237                 spin_unlock(&desc->bd_lock);
 238                 LASSERT(desc->bd_md_count >= 0);
 239
 240                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 241                 RETURN(0);
 242         }
 243
 244         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 245                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 246                desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
 247                mbits - posted_md, mbits - 1);
 248
 249         RETURN(0);
 250 }
 251
 252 /**
 253  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 254  * serialises with completion callback)
 255  */
 256 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 257 {
 258         struct l_wait_info       lwi;
 259         int                      rc;
 260
 261         LASSERT(!in_interrupt());           /* might sleep */
 262
 263         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 264                 return;                         /* never started */
 265
 266         /* We used to poison the pages with 0xab here because we did not want to
 267          * send any meaningful data over the wire for evicted clients (bug 9297)
 268          * However, this is no longer safe now that we use the page cache on the
 269          * OSS (bug 20560) */
 270
 271         /* The unlink ensures the callback happens ASAP and is the last
 272          * one.  If it fails, it must be because completion just happened,
 273          * but we must still l_wait_event() in this case, to give liblustre
 274          * a chance to run server_bulk_callback()*/
 275         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 276
 277         for (;;) {
 278                 /* Network access will complete in finite time but the HUGE
 279                  * timeout lets us CWARN for visibility of sluggish NALs */
 280                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 281                                            cfs_time_seconds(1), NULL, NULL);
 282                 rc = l_wait_event(desc->bd_waitq,
 283                                   !ptlrpc_server_bulk_active(desc), &lwi);
 284                 if (rc == 0)
 285                         return;
 286
 287                 LASSERT(rc == -ETIMEDOUT);
 288                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 289         }
 290 }
 291 #endif /* HAVE_SERVER_SUPPORT */
 292
 293 /**
 294  * Register bulk at the sender for later transfer.
 295  * Returns 0 on success or error code.
 296  */
 297 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 298 {
 299         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 300         lnet_process_id_t peer;
 301         int rc = 0;
 302         int rc2;
 303         int posted_md;
 304         int total_md;
 305         __u64 mbits;
 306         lnet_handle_me_t  me_h;
 307         lnet_md_t         md;
 308         ENTRY;
 309
 310         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 311                 RETURN(0);
 312
 313         /* NB no locking required until desc is on the network */
 314         LASSERT(desc->bd_nob > 0);
 315         LASSERT(desc->bd_md_count == 0);
 316         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 317         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 318         LASSERT(desc->bd_req != NULL);
 319         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 320
 321         /* cleanup the state of the bulk for it will be reused */
 322         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 323                 desc->bd_nob_transferred = 0;
 324         else
 325                 LASSERT(desc->bd_nob_transferred == 0);
 326
 327         desc->bd_failure = 0;
 328
 329         peer = desc->bd_import->imp_connection->c_peer;
 330
 331         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 332         LASSERT(desc->bd_cbid.cbid_arg == desc);
 333
 334         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 335         /* rq_mbits is matchbits of the final bulk */
 336         mbits = req->rq_mbits - total_md + 1;
 337
 338         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 339                  "first mbits = x%llu, last mbits = x%llu\n",
 340                  mbits, req->rq_mbits);
 341         LASSERTF(!(desc->bd_registered &&
 342                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 343                  mbits != desc->bd_last_mbits,
 344                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 345                  desc->bd_registered, mbits, desc->bd_last_mbits);
 346
 347         desc->bd_registered = 1;
 348         desc->bd_last_mbits = mbits;
 349         desc->bd_md_count = total_md;
 350         md.user_ptr = &desc->bd_cbid;
 351         md.eq_handle = ptlrpc_eq_h;
 352         md.threshold = 1;                       /* PUT or GET */
 353
 354         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 355                 md.options = PTLRPC_MD_OPTIONS |
 356                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 357                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 358                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 359
 360                 rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 361                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 362                 if (rc != 0) {
 363                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 364                                desc->bd_import->imp_obd->obd_name, mbits,
 365                                posted_md, rc);
 366                         break;
 367                 }
 368
 369                 /* About to let the network at it... */
 370                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 371                                   &desc->bd_mds[posted_md]);
 372                 if (rc != 0) {
 373                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 374                                desc->bd_import->imp_obd->obd_name, mbits,
 375                                posted_md, rc);
 376                         rc2 = LNetMEUnlink(me_h);
 377                         LASSERT(rc2 == 0);
 378                         break;
 379                 }
 380         }
 381
 382         if (rc != 0) {
 383                 LASSERT(rc == -ENOMEM);
 384                 spin_lock(&desc->bd_lock);
 385                 desc->bd_md_count -= total_md - posted_md;
 386                 spin_unlock(&desc->bd_lock);
 387                 LASSERT(desc->bd_md_count >= 0);
 388                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 389                 req->rq_status = -ENOMEM;
 390                 RETURN(-ENOMEM);
 391         }
 392
 393         spin_lock(&desc->bd_lock);
 394         /* Holler if peer manages to touch buffers before he knows the mbits */
 395         if (desc->bd_md_count != total_md)
 396                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 397                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 398                       total_md - desc->bd_md_count);
 399         spin_unlock(&desc->bd_lock);
 400
 401         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 402                "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
 403                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 404                desc->bd_iov_count, desc->bd_nob,
 405                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 406
 407         RETURN(0);
 408 }
 409
 410 /**
 411  * Disconnect a bulk desc from the network. Idempotent. Not
 412  * thread-safe (i.e. only interlocks with completion callback).
 413  * Returns 1 on success or 0 if network unregistration failed for whatever
 414  * reason.
 415  */
 416 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 417 {
 418         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 419         struct l_wait_info       lwi;
 420         int                      rc;
 421         ENTRY;
 422
 423         LASSERT(!in_interrupt());     /* might sleep */
 424
 425         /* Let's setup deadline for reply unlink. */
 426         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 427             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 428                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 429
 430         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 431                 RETURN(1);                              /* never registered */
 432
 433         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 434
 435         /* the unlink ensures the callback happens ASAP and is the last
 436          * one.  If it fails, it must be because completion just happened,
 437          * but we must still l_wait_event() in this case to give liblustre
 438          * a chance to run client_bulk_callback() */
 439         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 440
 441         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 442                 RETURN(1);                              /* never registered */
 443
 444         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 445         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 446
 447         /* Do not wait for unlink to finish. */
 448         if (async)
 449                 RETURN(0);
 450
 451         for (;;) {
 452                 /* The wq argument is ignored by user-space wait_event macros */
 453                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 454                                         &req->rq_set->set_waitq :
 455                                         &req->rq_reply_waitq;
 456                 /* Network access will complete in finite time but the HUGE
 457                  * timeout lets us CWARN for visibility of sluggish NALs */
 458                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 459                                            cfs_time_seconds(1), NULL, NULL);
 460                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 461                 if (rc == 0) {
 462                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 463                         RETURN(1);
 464                 }
 465
 466                 LASSERT(rc == -ETIMEDOUT);
 467                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 468                           desc);
 469         }
 470         RETURN(0);
 471 }
 472
 473 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 474 {
 475         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 476         struct ptlrpc_service           *svc = svcpt->scp_service;
 477         int service_time = max_t(int, cfs_time_current_sec() -
 478                                  req->rq_arrival_time.tv_sec, 1);
 479
 480         if (!(flags & PTLRPC_REPLY_EARLY) &&
 481             (req->rq_type != PTL_RPC_MSG_ERR) &&
 482             (req->rq_reqmsg != NULL) &&
 483             !(lustre_msg_get_flags(req->rq_reqmsg) &
 484               (MSG_RESENT | MSG_REPLAY |
 485                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 486                 /* early replies, errors and recovery requests don't count
 487                  * toward our service time estimate */
 488                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 489
 490                 if (oldse != 0) {
 491                         DEBUG_REQ(D_ADAPTTO, req,
 492                                   "svc %s changed estimate from %d to %d",
 493                                   svc->srv_name, oldse,
 494                                   at_get(&svcpt->scp_at_estimate));
 495                 }
 496         }
 497         /* Report actual service time for client latency calc */
 498         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 499         /* Report service time estimate for future client reqs, but report 0
 500          * (to be ignored by client) if it's an error reply during recovery.
 501          * (bz15815) */
 502         if (req->rq_type == PTL_RPC_MSG_ERR &&
 503             (req->rq_export == NULL ||
 504              req->rq_export->exp_obd->obd_recovering)) {
 505                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 506         } else {
 507                 __u32 timeout;
 508
 509                 if (req->rq_export && req->rq_reqmsg != NULL &&
 510                     (flags & PTLRPC_REPLY_EARLY) &&
 511                     lustre_msg_get_flags(req->rq_reqmsg) &
 512                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))
 513                         timeout = cfs_time_current_sec() -
 514                                 req->rq_arrival_time.tv_sec +
 515                                 min(at_extra,
 516                                     req->rq_export->exp_obd->
 517                                     obd_recovery_timeout / 4);
 518                 else
 519                         timeout = at_get(&svcpt->scp_at_estimate);
 520                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 521         }
 522
 523         if (req->rq_reqmsg &&
 524             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 525                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 526                        "req_flags=%#x magic=%x/%x len=%d\n",
 527                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 528                        lustre_msg_get_magic(req->rq_reqmsg),
 529                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 530         }
 531 }
 532
 533 /**
 534  * Send request reply from request \a req reply buffer.
 535  * \a flags defines reply types
 536  * Returns 0 on success or error code
 537  */
 538 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 539 {
 540         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 541         struct ptlrpc_connection  *conn;
 542         int                        rc;
 543
 544         /* We must already have a reply buffer (only ptlrpc_error() may be
 545          * called without one). The reply generated by sptlrpc layer (e.g.
 546          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 547          * have a request buffer which is either the actual (swabbed) incoming
 548          * request, or a saved copy if this is a req saved in
 549          * target_queue_final_reply().
 550          */
 551         LASSERT (req->rq_no_reply == 0);
 552         LASSERT (req->rq_reqbuf != NULL);
 553         LASSERT (rs != NULL);
 554         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 555         LASSERT (req->rq_repmsg != NULL);
 556         LASSERT (req->rq_repmsg == rs->rs_msg);
 557         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 558         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 559
 560         /* There may be no rq_export during failover */
 561
 562         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 563                      req->rq_export->exp_obd->obd_fail)) {
 564                 /* Failed obd's only send ENODEV */
 565                 req->rq_type = PTL_RPC_MSG_ERR;
 566                 req->rq_status = -ENODEV;
 567                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 568                        req->rq_export->exp_obd->obd_minor);
 569         }
 570
 571         /* In order to keep interoprability with the client (< 2.3) which
 572          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 573          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 574          * reply buffer on client will be overflow.
 575          *
 576          * XXX Remove this whenver we drop the interoprability with such client.
 577          */
 578         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 579                                            sizeof(struct ptlrpc_body_v2), 1);
 580
 581         if (req->rq_type != PTL_RPC_MSG_ERR)
 582                 req->rq_type = PTL_RPC_MSG_REPLY;
 583
 584         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 585         lustre_msg_set_status(req->rq_repmsg,
 586                               ptlrpc_status_hton(req->rq_status));
 587         lustre_msg_set_opc(req->rq_repmsg,
 588                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 589
 590         target_pack_pool_reply(req);
 591
 592         ptlrpc_at_set_reply(req, flags);
 593
 594         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 595                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 596         else
 597                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 598
 599         if (unlikely(conn == NULL)) {
 600                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 601                 return -ENOTCONN;
 602         }
 603         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 604
 605         rc = sptlrpc_svc_wrap_reply(req);
 606         if (unlikely(rc))
 607                 goto out;
 608
 609         req->rq_sent = cfs_time_current_sec();
 610
 611         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 612                            (rs->rs_difficult && !rs->rs_no_ack) ?
 613                            LNET_ACK_REQ : LNET_NOACK_REQ,
 614                            &rs->rs_cb_id, conn,
 615                            ptlrpc_req2svc(req)->srv_rep_portal,
 616                            req->rq_xid, req->rq_reply_off);
 617 out:
 618         if (unlikely(rc != 0))
 619                 ptlrpc_req_drop_rs(req);
 620         ptlrpc_connection_put(conn);
 621         return rc;
 622 }
 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
 632 /**
 633  * For request \a req send an error reply back. Create empty
 634  * reply buffers if necessary.
 635  */
 636 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 637 {
 638         int rc;
 639         ENTRY;
 640
 641         if (req->rq_no_reply)
 642                 RETURN(0);
 643
 644         if (!req->rq_repmsg) {
 645                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 646                 if (rc)
 647                         RETURN(rc);
 648         }
 649
 650         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 651             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 652             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 653                 req->rq_type = PTL_RPC_MSG_ERR;
 654
 655         rc = ptlrpc_send_reply(req, may_be_difficult);
 656         RETURN(rc);
 657 }
 658
 659 int ptlrpc_error(struct ptlrpc_request *req)
 660 {
 661         return ptlrpc_send_error(req, 0);
 662 }
 663
 664 /**
 665  * Send request \a request.
 666  * if \a noreply is set, don't expect any reply back and don't set up
 667  * reply buffers.
 668  * Returns 0 on success or error code.
 669  */
 670 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 671 {
 672         int rc;
 673         int rc2;
 674         int mpflag = 0;
 675         struct ptlrpc_connection *connection;
 676         lnet_handle_me_t  reply_me_h;
 677         lnet_md_t         reply_md;
 678         struct obd_import *imp = request->rq_import;
 679         struct obd_device *obd = imp->imp_obd;
 680         ENTRY;
 681
 682         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 683                 RETURN(0);
 684
 685         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 686         LASSERT(request->rq_wait_ctx == 0);
 687
 688         /* If this is a re-transmit, we're required to have disengaged
 689          * cleanly from the previous attempt */
 690         LASSERT(!request->rq_receiving_reply);
 691         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 692                 (imp->imp_state == LUSTRE_IMP_FULL)));
 693
 694         if (unlikely(obd != NULL && obd->obd_fail)) {
 695                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 696                         obd->obd_name);
 697                 /* this prevents us from waiting in ptlrpc_queue_wait */
 698                 spin_lock(&request->rq_lock);
 699                 request->rq_err = 1;
 700                 spin_unlock(&request->rq_lock);
 701                 request->rq_status = -ENODEV;
 702                 RETURN(-ENODEV);
 703         }
 704
 705         connection = imp->imp_connection;
 706
 707         lustre_msg_set_handle(request->rq_reqmsg,
 708                               &imp->imp_remote_handle);
 709         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 710         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 711                                 imp->imp_conn_cnt);
 712         lustre_msghdr_set_flags(request->rq_reqmsg,
 713                                 imp->imp_msghdr_flags);
 714
 715         /* If it's the first time to resend the request for EINPROGRESS,
 716          * we need to allocate a new XID (see after_reply()), it's different
 717          * from the resend for reply timeout. */
 718         if (request->rq_nr_resend != 0 &&
 719             list_empty(&request->rq_unreplied_list)) {
 720                 __u64 min_xid = 0;
 721                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 722                  * reconstruction */
 723                 spin_lock(&imp->imp_lock);
 724                 ptlrpc_assign_next_xid_nolock(request);
 725                 min_xid = ptlrpc_known_replied_xid(imp);
 726                 spin_unlock(&imp->imp_lock);
 727
 728                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 729                 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
 730                           "resend on EINPROGRESS");
 731         }
 732
 733         if (request->rq_bulk != NULL) {
 734                 ptlrpc_set_bulk_mbits(request);
 735                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 736         }
 737
 738         if (list_empty(&request->rq_unreplied_list) ||
 739             request->rq_xid <= imp->imp_known_replied_xid) {
 740                 DEBUG_REQ(D_ERROR, request, "xid: %llu, replied: %llu, "
 741                           "list_empty:%d\n", request->rq_xid,
 742                           imp->imp_known_replied_xid,
 743                           list_empty(&request->rq_unreplied_list));
 744                 LBUG();
 745         }
 746
 747         /** For enabled AT all request should have AT_SUPPORT in the
 748          * FULL import state when OBD_CONNECT_AT is set */
 749         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 750                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 751                 !(imp->imp_connect_data.ocd_connect_flags &
 752                 OBD_CONNECT_AT));
 753
 754         if (request->rq_resend) {
 755                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 756                 if (request->rq_resend_cb != NULL)
 757                         request->rq_resend_cb(request, &request->rq_async_args);
 758         }
 759         if (request->rq_memalloc)
 760                 mpflag = cfs_memory_pressure_get_and_set();
 761
 762         rc = sptlrpc_cli_wrap_request(request);
 763         if (rc == -ENOMEM)
 764                 /* set rq_sent so that this request is treated
 765                  * as a delayed send in the upper layers */
 766                 request->rq_sent = cfs_time_current_sec();
 767         if (rc)
 768                 GOTO(out, rc);
 769
 770         /* bulk register should be done after wrap_request() */
 771         if (request->rq_bulk != NULL) {
 772                 rc = ptlrpc_register_bulk (request);
 773                 if (rc != 0)
 774                         GOTO(out, rc);
 775         }
 776
 777         if (!noreply) {
 778                 LASSERT (request->rq_replen != 0);
 779                 if (request->rq_repbuf == NULL) {
 780                         LASSERT(request->rq_repdata == NULL);
 781                         LASSERT(request->rq_repmsg == NULL);
 782                         rc = sptlrpc_cli_alloc_repbuf(request,
 783                                                       request->rq_replen);
 784                         if (rc) {
 785                                 /* this prevents us from looping in
 786                                  * ptlrpc_queue_wait */
 787                                 spin_lock(&request->rq_lock);
 788                                 request->rq_err = 1;
 789                                 spin_unlock(&request->rq_lock);
 790                                 request->rq_status = rc;
 791                                 GOTO(cleanup_bulk, rc);
 792                         }
 793                 } else {
 794                         request->rq_repdata = NULL;
 795                         request->rq_repmsg = NULL;
 796                 }
 797
 798                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 799                                   connection->c_peer, request->rq_xid, 0,
 800                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 801                 if (rc != 0) {
 802                         CERROR("LNetMEAttach failed: %d\n", rc);
 803                         LASSERT (rc == -ENOMEM);
 804                         GOTO(cleanup_bulk, rc = -ENOMEM);
 805                 }
 806         }
 807
 808         spin_lock(&request->rq_lock);
 809         /* We are responsible for unlinking the reply buffer */
 810         request->rq_reply_unlinked = noreply;
 811         request->rq_receiving_reply = !noreply;
 812         /* Clear any flags that may be present from previous sends. */
 813         request->rq_req_unlinked = 0;
 814         request->rq_replied = 0;
 815         request->rq_err = 0;
 816         request->rq_timedout = 0;
 817         request->rq_net_err = 0;
 818         request->rq_resend = 0;
 819         request->rq_restart = 0;
 820         request->rq_reply_truncated = 0;
 821         spin_unlock(&request->rq_lock);
 822
 823         if (!noreply) {
 824                 reply_md.start     = request->rq_repbuf;
 825                 reply_md.length    = request->rq_repbuf_len;
 826                 /* Allow multiple early replies */
 827                 reply_md.threshold = LNET_MD_THRESH_INF;
 828                 /* Manage remote for early replies */
 829                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 830                         LNET_MD_MANAGE_REMOTE |
 831                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 832                 reply_md.user_ptr  = &request->rq_reply_cbid;
 833                 reply_md.eq_handle = ptlrpc_eq_h;
 834
 835                 /* We must see the unlink callback to set rq_reply_unlinked,
 836                  * so we can't auto-unlink */
 837                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 838                                   &request->rq_reply_md_h);
 839                 if (rc != 0) {
 840                         CERROR("LNetMDAttach failed: %d\n", rc);
 841                         LASSERT (rc == -ENOMEM);
 842                         spin_lock(&request->rq_lock);
 843                         /* ...but the MD attach didn't succeed... */
 844                         request->rq_receiving_reply = 0;
 845                         spin_unlock(&request->rq_lock);
 846                         GOTO(cleanup_me, rc = -ENOMEM);
 847                 }
 848
 849                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
 850                        ", portal %u\n",
 851                        request->rq_repbuf_len, request->rq_xid,
 852                        request->rq_reply_portal);
 853         }
 854
 855         /* add references on request for request_out_callback */
 856         ptlrpc_request_addref(request);
 857         if (obd != NULL && obd->obd_svc_stats != NULL)
 858                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 859                         atomic_read(&imp->imp_inflight));
 860
 861         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 862
 863         do_gettimeofday(&request->rq_sent_tv);
 864         request->rq_sent = cfs_time_current_sec();
 865         /* We give the server rq_timeout secs to process the req, and
 866            add the network latency for our local timeout. */
 867         request->rq_deadline = request->rq_sent + request->rq_timeout +
 868                 ptlrpc_at_get_net_latency(request);
 869
 870         ptlrpc_pinger_sending_on_import(imp);
 871
 872         DEBUG_REQ(D_INFO, request, "send flg=%x",
 873                   lustre_msg_get_flags(request->rq_reqmsg));
 874         rc = ptl_send_buf(&request->rq_req_md_h,
 875                           request->rq_reqbuf, request->rq_reqdata_len,
 876                           LNET_NOACK_REQ, &request->rq_req_cbid,
 877                           connection,
 878                           request->rq_request_portal,
 879                           request->rq_xid, 0);
 880         if (likely(rc == 0))
 881                 GOTO(out, rc);
 882
 883         request->rq_req_unlinked = 1;
 884         ptlrpc_req_finished(request);
 885         if (noreply)
 886                 GOTO(out, rc);
 887
 888  cleanup_me:
 889         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 890          * nobody apart from the PUT's target has the right nid+XID to
 891          * access the reply buffer. */
 892         rc2 = LNetMEUnlink(reply_me_h);
 893         LASSERT (rc2 == 0);
 894         /* UNLINKED callback called synchronously */
 895         LASSERT(!request->rq_receiving_reply);
 896
 897  cleanup_bulk:
 898         /* We do sync unlink here as there was no real transfer here so
 899          * the chance to have long unlink to sluggish net is smaller here. */
 900         ptlrpc_unregister_bulk(request, 0);
 901  out:
 902         if (request->rq_memalloc)
 903                 cfs_memory_pressure_restore(mpflag);
 904         return rc;
 905 }
 906 EXPORT_SYMBOL(ptl_send_rpc);
 907
 908 /**
 909  * Register request buffer descriptor for request receiving.
 910  */
 911 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 912 {
 913         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 914         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 915         int                       rc;
 916         lnet_md_t                 md;
 917         lnet_handle_me_t          me_h;
 918
 919         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 920                service->srv_req_portal);
 921
 922         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 923                 return (-ENOMEM);
 924
 925         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 926          * which means buffer can only be attached on local CPT, and LND
 927          * threads can find it by grabbing a local lock */
 928         rc = LNetMEAttach(service->srv_req_portal,
 929                           match_id, 0, ~0, LNET_UNLINK,
 930                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 931                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 932         if (rc != 0) {
 933                 CERROR("LNetMEAttach failed: %d\n", rc);
 934                 return (-ENOMEM);
 935         }
 936
 937         LASSERT(rqbd->rqbd_refcount == 0);
 938         rqbd->rqbd_refcount = 1;
 939
 940         md.start     = rqbd->rqbd_buffer;
 941         md.length    = service->srv_buf_size;
 942         md.max_size  = service->srv_max_req_size;
 943         md.threshold = LNET_MD_THRESH_INF;
 944         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 945         md.user_ptr  = &rqbd->rqbd_cbid;
 946         md.eq_handle = ptlrpc_eq_h;
 947
 948         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 949         if (rc == 0)
 950                 return (0);
 951
 952         CERROR("LNetMDAttach failed: %d; \n", rc);
 953         LASSERT (rc == -ENOMEM);
 954         rc = LNetMEUnlink (me_h);
 955         LASSERT (rc == 0);
 956         rqbd->rqbd_refcount = 0;
 957
 958         return (-ENOMEM);
 959 }