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                         lnet_nid_t self, lnet_process_id_t peer_id,
  49                         int portal, __u64 xid, unsigned int offset)
  50 {
  51         int              rc;
  52         lnet_md_t         md;
  53         ENTRY;
  54
  55         LASSERT (portal != 0);
  56         CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
  57         md.start     = base;
  58         md.length    = len;
  59         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  60         md.options   = PTLRPC_MD_OPTIONS;
  61         md.user_ptr  = cbid;
  62         md.eq_handle = ptlrpc_eq_h;
  63
  64         if (unlikely(ack == LNET_ACK_REQ &&
  65                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  66                 /* don't ask for the ack to simulate failing client */
  67                 ack = LNET_NOACK_REQ;
  68         }
  69
  70         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  71         if (unlikely(rc != 0)) {
  72                 CERROR ("LNetMDBind failed: %d\n", rc);
  73                 LASSERT (rc == -ENOMEM);
  74                 RETURN (-ENOMEM);
  75         }
  76
  77         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  78                len, portal, xid, offset);
  79
  80         rc = LNetPut(self, *mdh, ack,
  81                      peer_id, portal, xid, offset, 0);
  82         if (unlikely(rc != 0)) {
  83                 int rc2;
  84                 /* We're going to get an UNLINK event when I unlink below,
  85                  * which will complete just like any other failed send, so
  86                  * I fall through and return success here! */
  87                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  88                        libcfs_id2str(peer_id), portal, xid, rc);
  89                 rc2 = LNetMDUnlink(*mdh);
  90                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  91         }
  92
  93         RETURN (0);
  94 }
  95
  96 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
  97 {
  98         int i;
  99
 100         for (i = 0; i < count; i++)
 101                 LNetMDUnlink(bd_mds[i]);
 102 }
 103
 104 #ifdef HAVE_SERVER_SUPPORT
 105 /**
 106  * Prepare bulk descriptor for specified incoming request \a req that
 107  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 108  * the bulk to be sent. Used on server-side after request was already
 109  * received.
 110  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 111  * error.
 112  */
 113 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 114                                               unsigned nfrags, unsigned max_brw,
 115                                               unsigned int type,
 116                                               unsigned portal,
 117                                               const struct ptlrpc_bulk_frag_ops
 118                                                 *ops)
 119 {
 120         struct obd_export *exp = req->rq_export;
 121         struct ptlrpc_bulk_desc *desc;
 122
 123         ENTRY;
 124         LASSERT(ptlrpc_is_bulk_op_active(type));
 125
 126         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 127         if (desc == NULL)
 128                 RETURN(NULL);
 129
 130         desc->bd_export = class_export_get(exp);
 131         desc->bd_req = req;
 132
 133         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 134         desc->bd_cbid.cbid_arg = desc;
 135
 136         /* NB we don't assign rq_bulk here; server-side requests are
 137          * re-used, and the handler frees the bulk desc explicitly. */
 138
 139         return desc;
 140 }
 141 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 142
 143 /**
 144  * Starts bulk transfer for descriptor \a desc on the server.
 145  * Returns 0 on success or error code.
 146  */
 147 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 148 {
 149         struct obd_export        *exp = desc->bd_export;
 150         lnet_nid_t                self_nid;
 151         lnet_process_id_t         peer_id;
 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         /*
 170          * Multi-Rail: get the preferred self and peer NIDs from the
 171          * request, so they are based on the route taken by the
 172          * message.
 173          */
 174         self_nid = desc->bd_req->rq_self;
 175         peer_id = desc->bd_req->rq_source;
 176
 177         /* NB total length may be 0 for a read past EOF, so we send 0
 178          * length bulks, since the client expects bulk events.
 179          *
 180          * The client may not need all of the bulk mbits for the RPC. The RPC
 181          * used the mbits of the highest bulk mbits needed, and the server masks
 182          * off high bits to get bulk count for this RPC. LU-1431 */
 183         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 184         total_md = desc->bd_req->rq_mbits - mbits + 1;
 185
 186         desc->bd_md_count = total_md;
 187         desc->bd_failure = 0;
 188
 189         md.user_ptr = &desc->bd_cbid;
 190         md.eq_handle = ptlrpc_eq_h;
 191         md.threshold = 2; /* SENT and ACK/REPLY */
 192
 193         for (posted_md = 0; posted_md < total_md; mbits++) {
 194                 md.options = PTLRPC_MD_OPTIONS;
 195
 196                 /* NB it's assumed that source and sink buffer frags are
 197                  * page-aligned. Otherwise we'd have to send client bulk
 198                  * sizes over and split server buffer accordingly */
 199                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 200                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 201                 if (rc != 0) {
 202                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 203                                exp->exp_obd->obd_name, posted_md, rc);
 204                         LASSERT(rc == -ENOMEM);
 205                         if (posted_md == 0) {
 206                                 desc->bd_md_count = 0;
 207                                 RETURN(-ENOMEM);
 208                         }
 209                         break;
 210                 }
 211
 212                 /* LU-6441: last md is not sent and desc->bd_md_count == 1 */
 213                 if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 214                                          CFS_FAIL_ONCE) &&
 215                     total_md > 1 && posted_md == total_md - 1) {
 216                         posted_md++;
 217                         continue;
 218                 }
 219
 220                 /* Network is about to get at the memory */
 221                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 222                         rc = LNetPut(self_nid, desc->bd_mds[posted_md],
 223                                      LNET_ACK_REQ, peer_id,
 224                                      desc->bd_portal, mbits, 0, 0);
 225                 else
 226                         rc = LNetGet(self_nid, desc->bd_mds[posted_md],
 227                                      peer_id, desc->bd_portal, mbits, 0);
 228
 229                 posted_md++;
 230                 if (rc != 0) {
 231                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 232                                "rc = %d\n", exp->exp_obd->obd_name,
 233                                libcfs_id2str(peer_id), desc->bd_portal,
 234                                mbits, rc);
 235                         break;
 236                 }
 237         }
 238
 239         if (rc != 0) {
 240                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 241                  * event this creates will signal completion with failure,
 242                  * so we return SUCCESS here! */
 243                 spin_lock(&desc->bd_lock);
 244                 desc->bd_md_count -= total_md - posted_md;
 245                 spin_unlock(&desc->bd_lock);
 246                 LASSERT(desc->bd_md_count >= 0);
 247
 248                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 249                 RETURN(0);
 250         }
 251
 252         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 253                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 254                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
 255                mbits - posted_md, mbits - 1);
 256
 257         RETURN(0);
 258 }
 259
 260 /**
 261  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 262  * serialises with completion callback)
 263  */
 264 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 265 {
 266         struct l_wait_info       lwi;
 267         int                      rc;
 268
 269         LASSERT(!in_interrupt());           /* might sleep */
 270
 271         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 272                 return;                         /* never started */
 273
 274         /* We used to poison the pages with 0xab here because we did not want to
 275          * send any meaningful data over the wire for evicted clients (bug 9297)
 276          * However, this is no longer safe now that we use the page cache on the
 277          * OSS (bug 20560) */
 278
 279         /* The unlink ensures the callback happens ASAP and is the last
 280          * one.  If it fails, it must be because completion just happened,
 281          * but we must still l_wait_event() in this case, to give liblustre
 282          * a chance to run server_bulk_callback()*/
 283         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 284
 285         for (;;) {
 286                 /* Network access will complete in finite time but the HUGE
 287                  * timeout lets us CWARN for visibility of sluggish NALs */
 288                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 289                                            cfs_time_seconds(1), NULL, NULL);
 290                 rc = l_wait_event(desc->bd_waitq,
 291                                   !ptlrpc_server_bulk_active(desc), &lwi);
 292                 if (rc == 0)
 293                         return;
 294
 295                 LASSERT(rc == -ETIMEDOUT);
 296                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 297         }
 298 }
 299 #endif /* HAVE_SERVER_SUPPORT */
 300
 301 /**
 302  * Register bulk at the sender for later transfer.
 303  * Returns 0 on success or error code.
 304  */
 305 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 306 {
 307         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 308         lnet_process_id_t peer;
 309         int rc = 0;
 310         int rc2;
 311         int posted_md;
 312         int total_md;
 313         __u64 mbits;
 314         lnet_handle_me_t  me_h;
 315         lnet_md_t         md;
 316         ENTRY;
 317
 318         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 319                 RETURN(0);
 320
 321         /* NB no locking required until desc is on the network */
 322         LASSERT(desc->bd_nob > 0);
 323         LASSERT(desc->bd_md_count == 0);
 324         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 325         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 326         LASSERT(desc->bd_req != NULL);
 327         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 328
 329         /* cleanup the state of the bulk for it will be reused */
 330         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 331                 desc->bd_nob_transferred = 0;
 332         else
 333                 LASSERT(desc->bd_nob_transferred == 0);
 334
 335         desc->bd_failure = 0;
 336
 337         peer = desc->bd_import->imp_connection->c_peer;
 338
 339         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 340         LASSERT(desc->bd_cbid.cbid_arg == desc);
 341
 342         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 343         /* rq_mbits is matchbits of the final bulk */
 344         mbits = req->rq_mbits - total_md + 1;
 345
 346         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 347                  "first mbits = x%llu, last mbits = x%llu\n",
 348                  mbits, req->rq_mbits);
 349         LASSERTF(!(desc->bd_registered &&
 350                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 351                  mbits != desc->bd_last_mbits,
 352                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 353                  desc->bd_registered, mbits, desc->bd_last_mbits);
 354
 355         desc->bd_registered = 1;
 356         desc->bd_last_mbits = mbits;
 357         desc->bd_md_count = total_md;
 358         md.user_ptr = &desc->bd_cbid;
 359         md.eq_handle = ptlrpc_eq_h;
 360         md.threshold = 1;                       /* PUT or GET */
 361
 362         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 363                 md.options = PTLRPC_MD_OPTIONS |
 364                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 365                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 366                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 367
 368                 rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 369                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 370                 if (rc != 0) {
 371                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 372                                desc->bd_import->imp_obd->obd_name, mbits,
 373                                posted_md, rc);
 374                         break;
 375                 }
 376
 377                 /* About to let the network at it... */
 378                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 379                                   &desc->bd_mds[posted_md]);
 380                 if (rc != 0) {
 381                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 382                                desc->bd_import->imp_obd->obd_name, mbits,
 383                                posted_md, rc);
 384                         rc2 = LNetMEUnlink(me_h);
 385                         LASSERT(rc2 == 0);
 386                         break;
 387                 }
 388         }
 389
 390         if (rc != 0) {
 391                 LASSERT(rc == -ENOMEM);
 392                 spin_lock(&desc->bd_lock);
 393                 desc->bd_md_count -= total_md - posted_md;
 394                 spin_unlock(&desc->bd_lock);
 395                 LASSERT(desc->bd_md_count >= 0);
 396                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 397                 req->rq_status = -ENOMEM;
 398                 RETURN(-ENOMEM);
 399         }
 400
 401         spin_lock(&desc->bd_lock);
 402         /* Holler if peer manages to touch buffers before he knows the mbits */
 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                "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
 411                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 412                desc->bd_iov_count, desc->bd_nob,
 413                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 414
 415         RETURN(0);
 416 }
 417
 418 /**
 419  * Disconnect a bulk desc from the network. Idempotent. Not
 420  * thread-safe (i.e. only interlocks with completion callback).
 421  * Returns 1 on success or 0 if network unregistration failed for whatever
 422  * reason.
 423  */
 424 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 425 {
 426         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 427         struct l_wait_info       lwi;
 428         int                      rc;
 429         ENTRY;
 430
 431         LASSERT(!in_interrupt());     /* might sleep */
 432
 433         /* Let's setup deadline for reply unlink. */
 434         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 435             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 436                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 437
 438         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 439                 RETURN(1);                              /* never registered */
 440
 441         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 442
 443         /* the unlink ensures the callback happens ASAP and is the last
 444          * one.  If it fails, it must be because completion just happened,
 445          * but we must still l_wait_event() in this case to give liblustre
 446          * a chance to run client_bulk_callback() */
 447         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 448
 449         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 450                 RETURN(1);                              /* never registered */
 451
 452         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 453         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 454
 455         /* Do not wait for unlink to finish. */
 456         if (async)
 457                 RETURN(0);
 458
 459         for (;;) {
 460                 /* The wq argument is ignored by user-space wait_event macros */
 461                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 462                                         &req->rq_set->set_waitq :
 463                                         &req->rq_reply_waitq;
 464                 /* Network access will complete in finite time but the HUGE
 465                  * timeout lets us CWARN for visibility of sluggish NALs */
 466                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 467                                            cfs_time_seconds(1), NULL, NULL);
 468                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 469                 if (rc == 0) {
 470                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 471                         RETURN(1);
 472                 }
 473
 474                 LASSERT(rc == -ETIMEDOUT);
 475                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 476                           desc);
 477         }
 478         RETURN(0);
 479 }
 480
 481 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 482 {
 483         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 484         struct ptlrpc_service           *svc = svcpt->scp_service;
 485         int service_time = max_t(int, cfs_time_current_sec() -
 486                                  req->rq_arrival_time.tv_sec, 1);
 487
 488         if (!(flags & PTLRPC_REPLY_EARLY) &&
 489             (req->rq_type != PTL_RPC_MSG_ERR) &&
 490             (req->rq_reqmsg != NULL) &&
 491             !(lustre_msg_get_flags(req->rq_reqmsg) &
 492               (MSG_RESENT | MSG_REPLAY |
 493                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 494                 /* early replies, errors and recovery requests don't count
 495                  * toward our service time estimate */
 496                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 497
 498                 if (oldse != 0) {
 499                         DEBUG_REQ(D_ADAPTTO, req,
 500                                   "svc %s changed estimate from %d to %d",
 501                                   svc->srv_name, oldse,
 502                                   at_get(&svcpt->scp_at_estimate));
 503                 }
 504         }
 505         /* Report actual service time for client latency calc */
 506         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 507         /* Report service time estimate for future client reqs, but report 0
 508          * (to be ignored by client) if it's an error reply during recovery.
 509          * (bz15815) */
 510         if (req->rq_type == PTL_RPC_MSG_ERR &&
 511             (req->rq_export == NULL ||
 512              req->rq_export->exp_obd->obd_recovering)) {
 513                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 514         } else {
 515                 __u32 timeout;
 516
 517                 if (req->rq_export && req->rq_reqmsg != NULL &&
 518                     (flags & PTLRPC_REPLY_EARLY) &&
 519                     lustre_msg_get_flags(req->rq_reqmsg) &
 520                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))
 521                         timeout = cfs_time_current_sec() -
 522                                 req->rq_arrival_time.tv_sec +
 523                                 min(at_extra,
 524                                     req->rq_export->exp_obd->
 525                                     obd_recovery_timeout / 4);
 526                 else
 527                         timeout = at_get(&svcpt->scp_at_estimate);
 528                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 529         }
 530
 531         if (req->rq_reqmsg &&
 532             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 533                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 534                        "req_flags=%#x magic=%x/%x len=%d\n",
 535                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 536                        lustre_msg_get_magic(req->rq_reqmsg),
 537                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 538         }
 539 }
 540
 541 /**
 542  * Send request reply from request \a req reply buffer.
 543  * \a flags defines reply types
 544  * Returns 0 on success or error code
 545  */
 546 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 547 {
 548         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 549         struct ptlrpc_connection  *conn;
 550         int                        rc;
 551
 552         /* We must already have a reply buffer (only ptlrpc_error() may be
 553          * called without one). The reply generated by sptlrpc layer (e.g.
 554          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 555          * have a request buffer which is either the actual (swabbed) incoming
 556          * request, or a saved copy if this is a req saved in
 557          * target_queue_final_reply().
 558          */
 559         LASSERT (req->rq_no_reply == 0);
 560         LASSERT (req->rq_reqbuf != NULL);
 561         LASSERT (rs != NULL);
 562         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 563         LASSERT (req->rq_repmsg != NULL);
 564         LASSERT (req->rq_repmsg == rs->rs_msg);
 565         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 566         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 567
 568         /* There may be no rq_export during failover */
 569
 570         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 571                      req->rq_export->exp_obd->obd_fail)) {
 572                 /* Failed obd's only send ENODEV */
 573                 req->rq_type = PTL_RPC_MSG_ERR;
 574                 req->rq_status = -ENODEV;
 575                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 576                        req->rq_export->exp_obd->obd_minor);
 577         }
 578
 579         /* In order to keep interoprability with the client (< 2.3) which
 580          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 581          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 582          * reply buffer on client will be overflow.
 583          *
 584          * XXX Remove this whenver we drop the interoprability with such client.
 585          */
 586         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 587                                            sizeof(struct ptlrpc_body_v2), 1);
 588
 589         if (req->rq_type != PTL_RPC_MSG_ERR)
 590                 req->rq_type = PTL_RPC_MSG_REPLY;
 591
 592         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 593         lustre_msg_set_status(req->rq_repmsg,
 594                               ptlrpc_status_hton(req->rq_status));
 595         lustre_msg_set_opc(req->rq_repmsg,
 596                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 597
 598         target_pack_pool_reply(req);
 599
 600         ptlrpc_at_set_reply(req, flags);
 601
 602         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 603                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 604         else
 605                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 606
 607         if (unlikely(conn == NULL)) {
 608                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 609                 return -ENOTCONN;
 610         }
 611         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 612
 613         rc = sptlrpc_svc_wrap_reply(req);
 614         if (unlikely(rc))
 615                 goto out;
 616
 617         req->rq_sent = cfs_time_current_sec();
 618
 619         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 620                           (rs->rs_difficult && !rs->rs_no_ack) ?
 621                           LNET_ACK_REQ : LNET_NOACK_REQ,
 622                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 623                           ptlrpc_req2svc(req)->srv_rep_portal,
 624                           req->rq_xid, req->rq_reply_off);
 625 out:
 626         if (unlikely(rc != 0))
 627                 ptlrpc_req_drop_rs(req);
 628         ptlrpc_connection_put(conn);
 629         return rc;
 630 }
 631
 632 int ptlrpc_reply (struct ptlrpc_request *req)
 633 {
 634         if (req->rq_no_reply)
 635                 return 0;
 636         else
 637                 return (ptlrpc_send_reply(req, 0));
 638 }
 639
 640 /**
 641  * For request \a req send an error reply back. Create empty
 642  * reply buffers if necessary.
 643  */
 644 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 645 {
 646         int rc;
 647         ENTRY;
 648
 649         if (req->rq_no_reply)
 650                 RETURN(0);
 651
 652         if (!req->rq_repmsg) {
 653                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 654                 if (rc)
 655                         RETURN(rc);
 656         }
 657
 658         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 659             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 660             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 661                 req->rq_type = PTL_RPC_MSG_ERR;
 662
 663         rc = ptlrpc_send_reply(req, may_be_difficult);
 664         RETURN(rc);
 665 }
 666
 667 int ptlrpc_error(struct ptlrpc_request *req)
 668 {
 669         return ptlrpc_send_error(req, 0);
 670 }
 671
 672 /**
 673  * Send request \a request.
 674  * if \a noreply is set, don't expect any reply back and don't set up
 675  * reply buffers.
 676  * Returns 0 on success or error code.
 677  */
 678 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 679 {
 680         int rc;
 681         int rc2;
 682         int mpflag = 0;
 683         struct ptlrpc_connection *connection;
 684         lnet_handle_me_t  reply_me_h;
 685         lnet_md_t         reply_md;
 686         struct obd_import *imp = request->rq_import;
 687         struct obd_device *obd = imp->imp_obd;
 688         ENTRY;
 689
 690         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 691                 RETURN(0);
 692
 693         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 694         LASSERT(request->rq_wait_ctx == 0);
 695
 696         /* If this is a re-transmit, we're required to have disengaged
 697          * cleanly from the previous attempt */
 698         LASSERT(!request->rq_receiving_reply);
 699         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 700                 (imp->imp_state == LUSTRE_IMP_FULL)));
 701
 702         if (unlikely(obd != NULL && obd->obd_fail)) {
 703                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 704                         obd->obd_name);
 705                 /* this prevents us from waiting in ptlrpc_queue_wait */
 706                 spin_lock(&request->rq_lock);
 707                 request->rq_err = 1;
 708                 spin_unlock(&request->rq_lock);
 709                 request->rq_status = -ENODEV;
 710                 RETURN(-ENODEV);
 711         }
 712
 713         connection = imp->imp_connection;
 714
 715         lustre_msg_set_handle(request->rq_reqmsg,
 716                               &imp->imp_remote_handle);
 717         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 718         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 719                                 imp->imp_conn_cnt);
 720         lustre_msghdr_set_flags(request->rq_reqmsg,
 721                                 imp->imp_msghdr_flags);
 722
 723         /* If it's the first time to resend the request for EINPROGRESS,
 724          * we need to allocate a new XID (see after_reply()), it's different
 725          * from the resend for reply timeout. */
 726         if (request->rq_nr_resend != 0 &&
 727             list_empty(&request->rq_unreplied_list)) {
 728                 __u64 min_xid = 0;
 729                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 730                  * reconstruction */
 731                 spin_lock(&imp->imp_lock);
 732                 ptlrpc_assign_next_xid_nolock(request);
 733                 min_xid = ptlrpc_known_replied_xid(imp);
 734                 spin_unlock(&imp->imp_lock);
 735
 736                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 737                 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
 738                           "resend on EINPROGRESS");
 739         }
 740
 741         if (request->rq_bulk != NULL) {
 742                 ptlrpc_set_bulk_mbits(request);
 743                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 744         }
 745
 746         if (list_empty(&request->rq_unreplied_list) ||
 747             request->rq_xid <= imp->imp_known_replied_xid) {
 748                 DEBUG_REQ(D_ERROR, request, "xid: %llu, replied: %llu, "
 749                           "list_empty:%d\n", request->rq_xid,
 750                           imp->imp_known_replied_xid,
 751                           list_empty(&request->rq_unreplied_list));
 752                 LBUG();
 753         }
 754
 755         /** For enabled AT all request should have AT_SUPPORT in the
 756          * FULL import state when OBD_CONNECT_AT is set */
 757         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 758                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 759                 !(imp->imp_connect_data.ocd_connect_flags &
 760                 OBD_CONNECT_AT));
 761
 762         if (request->rq_resend) {
 763                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 764                 if (request->rq_resend_cb != NULL)
 765                         request->rq_resend_cb(request, &request->rq_async_args);
 766         }
 767         if (request->rq_memalloc)
 768                 mpflag = cfs_memory_pressure_get_and_set();
 769
 770         rc = sptlrpc_cli_wrap_request(request);
 771         if (rc == -ENOMEM)
 772                 /* set rq_sent so that this request is treated
 773                  * as a delayed send in the upper layers */
 774                 request->rq_sent = cfs_time_current_sec();
 775         if (rc)
 776                 GOTO(out, rc);
 777
 778         /* bulk register should be done after wrap_request() */
 779         if (request->rq_bulk != NULL) {
 780                 rc = ptlrpc_register_bulk (request);
 781                 if (rc != 0)
 782                         GOTO(out, rc);
 783         }
 784
 785         if (!noreply) {
 786                 LASSERT (request->rq_replen != 0);
 787                 if (request->rq_repbuf == NULL) {
 788                         LASSERT(request->rq_repdata == NULL);
 789                         LASSERT(request->rq_repmsg == NULL);
 790                         rc = sptlrpc_cli_alloc_repbuf(request,
 791                                                       request->rq_replen);
 792                         if (rc) {
 793                                 /* this prevents us from looping in
 794                                  * ptlrpc_queue_wait */
 795                                 spin_lock(&request->rq_lock);
 796                                 request->rq_err = 1;
 797                                 spin_unlock(&request->rq_lock);
 798                                 request->rq_status = rc;
 799                                 GOTO(cleanup_bulk, rc);
 800                         }
 801                 } else {
 802                         request->rq_repdata = NULL;
 803                         request->rq_repmsg = NULL;
 804                 }
 805
 806                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 807                                   connection->c_peer, request->rq_xid, 0,
 808                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 809                 if (rc != 0) {
 810                         CERROR("LNetMEAttach failed: %d\n", rc);
 811                         LASSERT (rc == -ENOMEM);
 812                         GOTO(cleanup_bulk, rc = -ENOMEM);
 813                 }
 814         }
 815
 816         spin_lock(&request->rq_lock);
 817         /* We are responsible for unlinking the reply buffer */
 818         request->rq_reply_unlinked = noreply;
 819         request->rq_receiving_reply = !noreply;
 820         /* Clear any flags that may be present from previous sends. */
 821         request->rq_req_unlinked = 0;
 822         request->rq_replied = 0;
 823         request->rq_err = 0;
 824         request->rq_timedout = 0;
 825         request->rq_net_err = 0;
 826         request->rq_resend = 0;
 827         request->rq_restart = 0;
 828         request->rq_reply_truncated = 0;
 829         spin_unlock(&request->rq_lock);
 830
 831         if (!noreply) {
 832                 reply_md.start     = request->rq_repbuf;
 833                 reply_md.length    = request->rq_repbuf_len;
 834                 /* Allow multiple early replies */
 835                 reply_md.threshold = LNET_MD_THRESH_INF;
 836                 /* Manage remote for early replies */
 837                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 838                         LNET_MD_MANAGE_REMOTE |
 839                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 840                 reply_md.user_ptr  = &request->rq_reply_cbid;
 841                 reply_md.eq_handle = ptlrpc_eq_h;
 842
 843                 /* We must see the unlink callback to set rq_reply_unlinked,
 844                  * so we can't auto-unlink */
 845                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 846                                   &request->rq_reply_md_h);
 847                 if (rc != 0) {
 848                         CERROR("LNetMDAttach failed: %d\n", rc);
 849                         LASSERT (rc == -ENOMEM);
 850                         spin_lock(&request->rq_lock);
 851                         /* ...but the MD attach didn't succeed... */
 852                         request->rq_receiving_reply = 0;
 853                         spin_unlock(&request->rq_lock);
 854                         GOTO(cleanup_me, rc = -ENOMEM);
 855                 }
 856
 857                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
 858                        ", portal %u\n",
 859                        request->rq_repbuf_len, request->rq_xid,
 860                        request->rq_reply_portal);
 861         }
 862
 863         /* add references on request for request_out_callback */
 864         ptlrpc_request_addref(request);
 865         if (obd != NULL && obd->obd_svc_stats != NULL)
 866                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 867                         atomic_read(&imp->imp_inflight));
 868
 869         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 870
 871         do_gettimeofday(&request->rq_sent_tv);
 872         request->rq_sent = cfs_time_current_sec();
 873         /* We give the server rq_timeout secs to process the req, and
 874            add the network latency for our local timeout. */
 875         request->rq_deadline = request->rq_sent + request->rq_timeout +
 876                 ptlrpc_at_get_net_latency(request);
 877
 878         ptlrpc_pinger_sending_on_import(imp);
 879
 880         DEBUG_REQ(D_INFO, request, "send flg=%x",
 881                   lustre_msg_get_flags(request->rq_reqmsg));
 882         rc = ptl_send_buf(&request->rq_req_md_h,
 883                           request->rq_reqbuf, request->rq_reqdata_len,
 884                           LNET_NOACK_REQ, &request->rq_req_cbid,
 885                           LNET_NID_ANY, connection->c_peer,
 886                           request->rq_request_portal,
 887                           request->rq_xid, 0);
 888         if (likely(rc == 0))
 889                 GOTO(out, rc);
 890
 891         request->rq_req_unlinked = 1;
 892         ptlrpc_req_finished(request);
 893         if (noreply)
 894                 GOTO(out, rc);
 895
 896  cleanup_me:
 897         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 898          * nobody apart from the PUT's target has the right nid+XID to
 899          * access the reply buffer. */
 900         rc2 = LNetMEUnlink(reply_me_h);
 901         LASSERT (rc2 == 0);
 902         /* UNLINKED callback called synchronously */
 903         LASSERT(!request->rq_receiving_reply);
 904
 905  cleanup_bulk:
 906         /* We do sync unlink here as there was no real transfer here so
 907          * the chance to have long unlink to sluggish net is smaller here. */
 908         ptlrpc_unregister_bulk(request, 0);
 909  out:
 910         if (request->rq_memalloc)
 911                 cfs_memory_pressure_restore(mpflag);
 912         return rc;
 913 }
 914 EXPORT_SYMBOL(ptl_send_rpc);
 915
 916 /**
 917  * Register request buffer descriptor for request receiving.
 918  */
 919 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 920 {
 921         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 922         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 923         int                       rc;
 924         lnet_md_t                 md;
 925         lnet_handle_me_t          me_h;
 926
 927         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 928                service->srv_req_portal);
 929
 930         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 931                 return (-ENOMEM);
 932
 933         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 934          * which means buffer can only be attached on local CPT, and LND
 935          * threads can find it by grabbing a local lock */
 936         rc = LNetMEAttach(service->srv_req_portal,
 937                           match_id, 0, ~0, LNET_UNLINK,
 938                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 939                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 940         if (rc != 0) {
 941                 CERROR("LNetMEAttach failed: %d\n", rc);
 942                 return (-ENOMEM);
 943         }
 944
 945         LASSERT(rqbd->rqbd_refcount == 0);
 946         rqbd->rqbd_refcount = 1;
 947
 948         md.start     = rqbd->rqbd_buffer;
 949         md.length    = service->srv_buf_size;
 950         md.max_size  = service->srv_max_req_size;
 951         md.threshold = LNET_MD_THRESH_INF;
 952         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 953         md.user_ptr  = &rqbd->rqbd_cbid;
 954         md.eq_handle = ptlrpc_eq_h;
 955
 956         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 957         if (rc == 0)
 958                 return (0);
 959
 960         CERROR("LNetMDAttach failed: %d; \n", rc);
 961         LASSERT (rc == -ENOMEM);
 962         rc = LNetMEUnlink (me_h);
 963         LASSERT (rc == 0);
 964         rqbd->rqbd_refcount = 0;
 965
 966         return (-ENOMEM);
 967 }