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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_RPC
  33 #include <libcfs/linux/linux-mem.h>
  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 #include <lnet/lib-lnet.h> /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */
  41
  42 /**
  43  * Helper function. Sends \a len bytes from \a base at offset \a offset
  44  * over \a conn connection to portal \a portal.
  45  * Returns 0 on success or error code.
  46  */
  47 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
  48                         enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
  49                         lnet_nid_t self, struct lnet_process_id peer_id,
  50                         int portal, __u64 xid, unsigned int offset,
  51                         struct lnet_handle_md *bulk_cookie)
  52 {
  53         int              rc;
  54         struct lnet_md         md;
  55         ENTRY;
  56
  57         LASSERT (portal != 0);
  58         CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
  59         md.start     = base;
  60         md.length    = len;
  61         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  62         md.options   = PTLRPC_MD_OPTIONS;
  63         md.user_ptr  = cbid;
  64         md.handler   = ptlrpc_handler;
  65         LNetInvalidateMDHandle(&md.bulk_handle);
  66
  67         if (bulk_cookie) {
  68                 md.bulk_handle = *bulk_cookie;
  69                 md.options |= LNET_MD_BULK_HANDLE;
  70         }
  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 %lld, offset %u\n",
  86                len, portal, xid, offset);
  87
  88         percpu_ref_get(&ptlrpc_pending);
  89
  90         rc = LNetPut(self, *mdh, ack,
  91                      peer_id, portal, xid, offset, 0);
  92         if (unlikely(rc != 0)) {
  93                 int rc2;
  94                 /* We're going to get an UNLINK event when I unlink below,
  95                  * which will complete just like any other failed send, so
  96                  * I fall through and return success here! */
  97                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  98                        libcfs_id2str(peer_id), portal, xid, rc);
  99                 rc2 = LNetMDUnlink(*mdh);
 100                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
 101         }
 102
 103         RETURN (0);
 104 }
 105
 106 #define mdunlink_iterate_helper(mds, count) \
 107                 __mdunlink_iterate_helper(mds, count, false)
 108 static void __mdunlink_iterate_helper(struct lnet_handle_md *bd_mds,
 109                                       int count, bool discard)
 110 {
 111         int i;
 112
 113         for (i = 0; i < count; i++)
 114                 __LNetMDUnlink(bd_mds[i], discard);
 115 }
 116
 117 #ifdef HAVE_SERVER_SUPPORT
 118 /**
 119  * Prepare bulk descriptor for specified incoming request \a req that
 120  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 121  * the bulk to be sent. Used on server-side after request was already
 122  * received.
 123  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 124  * error.
 125  */
 126 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 127                                               unsigned nfrags, unsigned max_brw,
 128                                               unsigned int type,
 129                                               unsigned portal,
 130                                               const struct ptlrpc_bulk_frag_ops
 131                                                 *ops)
 132 {
 133         struct obd_export *exp = req->rq_export;
 134         struct ptlrpc_bulk_desc *desc;
 135
 136         ENTRY;
 137         LASSERT(ptlrpc_is_bulk_op_active(type));
 138
 139         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 140         if (desc == NULL)
 141                 RETURN(NULL);
 142
 143         desc->bd_export = class_export_get(exp);
 144         desc->bd_req = req;
 145
 146         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 147         desc->bd_cbid.cbid_arg = desc;
 148
 149         /* NB we don't assign rq_bulk here; server-side requests are
 150          * re-used, and the handler frees the bulk desc explicitly. */
 151
 152         return desc;
 153 }
 154 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 155
 156 /**
 157  * Starts bulk transfer for descriptor \a desc on the server.
 158  * Returns 0 on success or error code.
 159  */
 160 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 161 {
 162         struct obd_export        *exp = desc->bd_export;
 163         lnet_nid_t                self_nid;
 164         struct lnet_process_id    peer_id;
 165         int                       rc = 0;
 166         __u64                     mbits;
 167         int                       posted_md;
 168         int                       total_md;
 169         struct lnet_md                 md;
 170         ENTRY;
 171
 172         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 173                 RETURN(0);
 174
 175         /* NB no locking required until desc is on the network */
 176         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 177
 178         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 179         LASSERT(desc->bd_cbid.cbid_arg == desc);
 180
 181         /*
 182          * Multi-Rail: get the preferred self and peer NIDs from the
 183          * request, so they are based on the route taken by the
 184          * message.
 185          */
 186         self_nid = desc->bd_req->rq_self;
 187         peer_id = desc->bd_req->rq_source;
 188
 189         /* NB total length may be 0 for a read past EOF, so we send 0
 190          * length bulks, since the client expects bulk events.
 191          *
 192          * The client may not need all of the bulk mbits for the RPC. The RPC
 193          * used the mbits of the highest bulk mbits needed, and the server masks
 194          * off high bits to get bulk count for this RPC. LU-1431 */
 195         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 196         total_md = desc->bd_req->rq_mbits - mbits + 1;
 197         desc->bd_refs = total_md;
 198         desc->bd_failure = 0;
 199
 200         md.user_ptr = &desc->bd_cbid;
 201         md.handler = ptlrpc_handler;
 202         md.threshold = 2; /* SENT and ACK/REPLY */
 203
 204         for (posted_md = 0; posted_md < total_md; mbits++) {
 205                 md.options = PTLRPC_MD_OPTIONS;
 206
 207                 /* NB it's assumed that source and sink buffer frags are
 208                  * page-aligned. Otherwise we'd have to send client bulk
 209                  * sizes over and split server buffer accordingly */
 210                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 211                 rc = LNetMDBind(&md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 212                 if (rc != 0) {
 213                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 214                                exp->exp_obd->obd_name, posted_md, rc);
 215                         LASSERT(rc == -ENOMEM);
 216                         if (posted_md == 0) {
 217                                 desc->bd_md_count = 0;
 218                                 RETURN(-ENOMEM);
 219                         }
 220                         break;
 221                 }
 222                 percpu_ref_get(&ptlrpc_pending);
 223
 224                 /* sanity.sh 224c: lets skip last md */
 225                 if (posted_md == desc->bd_md_max_brw - 1)
 226                         OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 227                                              CFS_FAIL_PTLRPC_OST_BULK_CB2);
 228
 229                 /* Network is about to get at the memory */
 230                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 231                         rc = LNetPut(self_nid, desc->bd_mds[posted_md],
 232                                      LNET_ACK_REQ, peer_id,
 233                                      desc->bd_portal, mbits, 0, 0);
 234                 else
 235                         rc = LNetGet(self_nid, desc->bd_mds[posted_md],
 236                                      peer_id, desc->bd_portal, mbits, 0, false);
 237
 238                 posted_md++;
 239                 if (rc != 0) {
 240                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 241                                "rc = %d\n", exp->exp_obd->obd_name,
 242                                libcfs_id2str(peer_id), desc->bd_portal,
 243                                mbits, rc);
 244                         break;
 245                 }
 246         }
 247
 248         if (rc != 0) {
 249                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 250                  * event this creates will signal completion with failure,
 251                  * so we return SUCCESS here! */
 252                 spin_lock(&desc->bd_lock);
 253                 desc->bd_refs -= total_md - posted_md;
 254                 spin_unlock(&desc->bd_lock);
 255                 LASSERT(desc->bd_refs >= 0);
 256
 257                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 258                 RETURN(0);
 259         }
 260
 261         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 262                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 263                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
 264                mbits - posted_md, mbits - 1);
 265
 266         RETURN(0);
 267 }
 268
 269 /**
 270  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 271  * serialises with completion callback)
 272  */
 273 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 274 {
 275         LASSERT(!in_interrupt());           /* might sleep */
 276
 277         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 278                 return;                         /* never started */
 279
 280         /* We used to poison the pages with 0xab here because we did not want to
 281          * send any meaningful data over the wire for evicted clients (bug 9297)
 282          * However, this is no longer safe now that we use the page cache on the
 283          * OSS (bug 20560) */
 284
 285         /* The unlink ensures the callback happens ASAP and is the last
 286          * one.  If it fails, it must be because completion just happened,
 287          * but we must still wait_event_idle_timeout() in this case, to give
 288          * us a chance to run server_bulk_callback()
 289          */
 290         __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true);
 291
 292         for (;;) {
 293                 /* Network access will complete in finite time but the HUGE
 294                  * timeout lets us CWARN for visibility of sluggish NALs */
 295                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 296
 297                 while (seconds > 0 &&
 298                        wait_event_idle_timeout(desc->bd_waitq,
 299                                                !ptlrpc_server_bulk_active(desc),
 300                                                cfs_time_seconds(1)) == 0)
 301                         seconds -= 1;
 302                 if (seconds > 0)
 303                         return;
 304
 305                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 306         }
 307 }
 308 #endif /* HAVE_SERVER_SUPPORT */
 309
 310 /**
 311  * Register bulk at the sender for later transfer.
 312  * Returns 0 on success or error code.
 313  */
 314 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 315 {
 316         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 317         struct lnet_processid peer;
 318         int rc = 0;
 319         int posted_md;
 320         int total_md;
 321         __u64 mbits;
 322         struct lnet_me *me;
 323         struct lnet_md md;
 324         ENTRY;
 325
 326         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 327                 RETURN(0);
 328
 329         /* NB no locking required until desc is on the network */
 330         LASSERT(desc->bd_nob > 0);
 331         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 332         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 333         LASSERT(desc->bd_req != NULL);
 334         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 335
 336         /* cleanup the state of the bulk for it will be reused */
 337         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 338                 desc->bd_nob_transferred = 0;
 339         else if (desc->bd_nob_transferred != 0)
 340                 /* If the network failed after an RPC was sent, this condition
 341                  * could happen.  Rather than assert (was here before), return
 342                  * an EIO error. */
 343                 RETURN(-EIO);
 344
 345         desc->bd_failure = 0;
 346
 347         peer.pid = desc->bd_import->imp_connection->c_peer.pid;
 348         lnet_nid4_to_nid(desc->bd_import->imp_connection->c_peer.nid,
 349                       &peer.nid);
 350
 351         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 352         LASSERT(desc->bd_cbid.cbid_arg == desc);
 353
 354         total_md = desc->bd_md_count;
 355         /* rq_mbits is matchbits of the final bulk */
 356         mbits = req->rq_mbits - desc->bd_md_count + 1;
 357
 358         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 359                  "first mbits = x%llu, last mbits = x%llu\n",
 360                  mbits, req->rq_mbits);
 361         LASSERTF(!(desc->bd_registered &&
 362                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 363                  mbits != desc->bd_last_mbits,
 364                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 365                  desc->bd_registered, mbits, desc->bd_last_mbits);
 366
 367         desc->bd_registered = 1;
 368         desc->bd_last_mbits = mbits;
 369         desc->bd_refs = total_md;
 370         md.user_ptr = &desc->bd_cbid;
 371         md.handler = ptlrpc_handler;
 372         md.threshold = 1;                       /* PUT or GET */
 373
 374         for (posted_md = 0; posted_md < desc->bd_md_count;
 375              posted_md++, mbits++) {
 376                 md.options = PTLRPC_MD_OPTIONS |
 377                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 378                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 379                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 380
 381                 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
 382                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 383                         rc = -ENOMEM;
 384                 } else {
 385                         me = LNetMEAttach(desc->bd_portal, &peer, mbits, 0,
 386                                   LNET_UNLINK, LNET_INS_AFTER);
 387                         rc = PTR_ERR_OR_ZERO(me);
 388                 }
 389                 if (rc != 0) {
 390                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 391                                desc->bd_import->imp_obd->obd_name, mbits,
 392                                posted_md, rc);
 393                         break;
 394                 }
 395                 percpu_ref_get(&ptlrpc_pending);
 396
 397                 /* About to let the network at it... */
 398                 rc = LNetMDAttach(me, &md, LNET_UNLINK,
 399                                   &desc->bd_mds[posted_md]);
 400                 if (rc != 0) {
 401                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 402                                desc->bd_import->imp_obd->obd_name, mbits,
 403                                posted_md, rc);
 404                         break;
 405                 }
 406         }
 407
 408         if (rc != 0) {
 409                 LASSERT(rc == -ENOMEM);
 410                 spin_lock(&desc->bd_lock);
 411                 desc->bd_refs -= total_md - posted_md;
 412                 spin_unlock(&desc->bd_lock);
 413                 LASSERT(desc->bd_refs >= 0);
 414                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 415                 req->rq_status = -ENOMEM;
 416                 desc->bd_registered = 0;
 417                 RETURN(-ENOMEM);
 418         }
 419
 420         spin_lock(&desc->bd_lock);
 421         /* Holler if peer manages to touch buffers before he knows the mbits */
 422         if (desc->bd_refs != total_md)
 423                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 424                       desc->bd_import->imp_obd->obd_name, libcfs_idstr(&peer),
 425                       total_md - desc->bd_refs);
 426         spin_unlock(&desc->bd_lock);
 427
 428         CDEBUG(D_NET,
 429                "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
 430                desc->bd_refs,
 431                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 432                desc->bd_iov_count, desc->bd_nob,
 433                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 434
 435         RETURN(0);
 436 }
 437
 438 /**
 439  * Disconnect a bulk desc from the network. Idempotent. Not
 440  * thread-safe (i.e. only interlocks with completion callback).
 441  * Returns 1 on success or 0 if network unregistration failed for whatever
 442  * reason.
 443  */
 444 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 445 {
 446         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 447         ENTRY;
 448
 449         LASSERT(!in_interrupt());     /* might sleep */
 450
 451         if (desc)
 452                 desc->bd_registered = 0;
 453
 454         /* Let's setup deadline for reply unlink. */
 455         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 456             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 457                 req->rq_bulk_deadline = ktime_get_real_seconds() +
 458                                         PTLRPC_REQ_LONG_UNLINK;
 459
 460         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 461                 RETURN(1);                              /* never registered */
 462
 463         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 464
 465         /* the unlink ensures the callback happens ASAP and is the last
 466          * one.  If it fails, it must be because completion just happened,
 467          * but we must still wait_event_idle_timeout() in this case to give
 468          * us a chance to run client_bulk_callback()
 469          */
 470         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 471
 472         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 473                 RETURN(1);                              /* never registered */
 474
 475         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 476         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 477
 478         /* Do not wait for unlink to finish. */
 479         if (async)
 480                 RETURN(0);
 481
 482         for (;;) {
 483                 /* The wq argument is ignored by user-space wait_event macros */
 484                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 485                                         &req->rq_set->set_waitq :
 486                                         &req->rq_reply_waitq;
 487                 /*
 488                  * Network access will complete in finite time but the HUGE
 489                  * timeout lets us CWARN for visibility of sluggish NALs.
 490                  */
 491                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 492
 493                 while (seconds > 0 &&
 494                        wait_event_idle_timeout(*wq,
 495                                                !ptlrpc_client_bulk_active(req),
 496                                                cfs_time_seconds(1)) == 0)
 497                         seconds -= 1;
 498                 if (seconds > 0) {
 499                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 500                         RETURN(1);
 501                 }
 502
 503                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 504                           desc);
 505         }
 506         RETURN(0);
 507 }
 508
 509 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 510 {
 511         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 512         struct ptlrpc_service           *svc = svcpt->scp_service;
 513         timeout_t service_timeout;
 514
 515         service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
 516                                              req->rq_arrival_time.tv_sec, 1,
 517                                   (AT_OFF ? obd_timeout * 3 / 2 : at_max));
 518         if (!(flags & PTLRPC_REPLY_EARLY) &&
 519             (req->rq_type != PTL_RPC_MSG_ERR) &&
 520             (req->rq_reqmsg != NULL) &&
 521             !(lustre_msg_get_flags(req->rq_reqmsg) &
 522               (MSG_RESENT | MSG_REPLAY |
 523                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 524                 /* early replies, errors and recovery requests don't count
 525                  * toward our service time estimate
 526                  */
 527                 timeout_t oldse = at_measured(&svcpt->scp_at_estimate,
 528                                               service_timeout);
 529
 530                 if (oldse != 0) {
 531                         DEBUG_REQ(D_ADAPTTO, req,
 532                                   "svc %s changed estimate from %d to %d",
 533                                   svc->srv_name, oldse,
 534                                   at_get(&svcpt->scp_at_estimate));
 535                 }
 536         }
 537         /* Report actual service time for client latency calc */
 538         lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
 539         /* Report service time estimate for future client reqs, but report 0
 540          * (to be ignored by client) if it's an error reply during recovery.
 541          * b=15815
 542          */
 543         if (req->rq_type == PTL_RPC_MSG_ERR &&
 544             (req->rq_export == NULL ||
 545              req->rq_export->exp_obd->obd_recovering)) {
 546                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 547         } else {
 548                 timeout_t timeout;
 549
 550                 if (req->rq_export && req->rq_reqmsg != NULL &&
 551                     (flags & PTLRPC_REPLY_EARLY) &&
 552                     lustre_msg_get_flags(req->rq_reqmsg) &
 553                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 554                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 555
 556                         timeout = ktime_get_real_seconds() -
 557                                   req->rq_arrival_time.tv_sec +
 558                                   min_t(timeout_t, at_extra,
 559                                         exp_obd->obd_recovery_timeout / 4);
 560                 } else {
 561                         timeout = at_get(&svcpt->scp_at_estimate);
 562                 }
 563                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 564         }
 565
 566         if (req->rq_reqmsg &&
 567             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 568                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 569                        "req_flags=%#x magic=%x/%x len=%d\n",
 570                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 571                        lustre_msg_get_magic(req->rq_reqmsg),
 572                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 573         }
 574 }
 575
 576 /**
 577  * Send request reply from request \a req reply buffer.
 578  * \a flags defines reply types
 579  * Returns 0 on success or error code
 580  */
 581 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 582 {
 583         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 584         struct ptlrpc_connection  *conn;
 585         int                        rc;
 586
 587         /* We must already have a reply buffer (only ptlrpc_error() may be
 588          * called without one). The reply generated by sptlrpc layer (e.g.
 589          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 590          * have a request buffer which is either the actual (swabbed) incoming
 591          * request, or a saved copy if this is a req saved in
 592          * target_queue_final_reply().
 593          */
 594         LASSERT (req->rq_no_reply == 0);
 595         LASSERT (req->rq_reqbuf != NULL);
 596         LASSERT (rs != NULL);
 597         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 598         LASSERT (req->rq_repmsg != NULL);
 599         LASSERT (req->rq_repmsg == rs->rs_msg);
 600         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 601         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 602
 603         /* There may be no rq_export during failover */
 604
 605         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 606                      req->rq_export->exp_obd->obd_fail)) {
 607                 /* Failed obd's only send ENODEV */
 608                 req->rq_type = PTL_RPC_MSG_ERR;
 609                 req->rq_status = -ENODEV;
 610                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 611                        req->rq_export->exp_obd->obd_minor);
 612         }
 613
 614         if (req->rq_type != PTL_RPC_MSG_ERR)
 615                 req->rq_type = PTL_RPC_MSG_REPLY;
 616
 617         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 618         lustre_msg_set_status(req->rq_repmsg,
 619                               ptlrpc_status_hton(req->rq_status));
 620         lustre_msg_set_opc(req->rq_repmsg,
 621                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 622
 623         target_pack_pool_reply(req);
 624
 625         ptlrpc_at_set_reply(req, flags);
 626
 627         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 628                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 629         else
 630                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 631
 632         if (unlikely(conn == NULL)) {
 633                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 634                 return -ENOTCONN;
 635         }
 636         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 637
 638         rc = sptlrpc_svc_wrap_reply(req);
 639         if (unlikely(rc))
 640                 goto out;
 641
 642         req->rq_sent = ktime_get_real_seconds();
 643
 644         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 645                           (rs->rs_difficult && !rs->rs_no_ack) ?
 646                           LNET_ACK_REQ : LNET_NOACK_REQ,
 647                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 648                           ptlrpc_req2svc(req)->srv_rep_portal,
 649                           req->rq_rep_mbits ? req->rq_rep_mbits : req->rq_xid,
 650                           req->rq_reply_off, NULL);
 651 out:
 652         if (unlikely(rc != 0))
 653                 ptlrpc_req_drop_rs(req);
 654         ptlrpc_connection_put(conn);
 655         return rc;
 656 }
 657
 658 int ptlrpc_reply (struct ptlrpc_request *req)
 659 {
 660         if (req->rq_no_reply)
 661                 return 0;
 662         else
 663                 return (ptlrpc_send_reply(req, 0));
 664 }
 665
 666 /**
 667  * For request \a req send an error reply back. Create empty
 668  * reply buffers if necessary.
 669  */
 670 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 671 {
 672         int rc;
 673         ENTRY;
 674
 675         if (req->rq_no_reply)
 676                 RETURN(0);
 677
 678         if (!req->rq_repmsg) {
 679                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 680                 if (rc)
 681                         RETURN(rc);
 682         }
 683
 684         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 685             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 686             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 687                 req->rq_type = PTL_RPC_MSG_ERR;
 688
 689         rc = ptlrpc_send_reply(req, may_be_difficult);
 690         RETURN(rc);
 691 }
 692
 693 int ptlrpc_error(struct ptlrpc_request *req)
 694 {
 695         return ptlrpc_send_error(req, 0);
 696 }
 697
 698 /**
 699  * Send request \a request.
 700  * if \a noreply is set, don't expect any reply back and don't set up
 701  * reply buffers.
 702  * Returns 0 on success or error code.
 703  */
 704 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 705 {
 706         int rc;
 707         __u32 opc;
 708         int mpflag = 0;
 709         bool rep_mbits = false;
 710         struct lnet_handle_md bulk_cookie;
 711         struct lnet_processid peer;
 712         struct ptlrpc_connection *connection;
 713         struct lnet_me *reply_me = NULL;
 714         struct lnet_md reply_md;
 715         struct obd_import *imp = request->rq_import;
 716         struct obd_device *obd = imp->imp_obd;
 717         ENTRY;
 718
 719         LNetInvalidateMDHandle(&bulk_cookie);
 720
 721         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 722                 RETURN(0);
 723
 724         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) &&
 725                      lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT &&
 726                      strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) {
 727                 RETURN(0);
 728         }
 729
 730         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 731         LASSERT(request->rq_wait_ctx == 0);
 732
 733         /* If this is a re-transmit, we're required to have disengaged
 734          * cleanly from the previous attempt */
 735         LASSERT(!request->rq_receiving_reply);
 736         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 737                   (imp->imp_state == LUSTRE_IMP_FULL)));
 738
 739         if (unlikely(obd != NULL && obd->obd_fail)) {
 740                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 741                        obd->obd_name);
 742                 /* this prevents us from waiting in ptlrpc_queue_wait */
 743                 spin_lock(&request->rq_lock);
 744                 request->rq_err = 1;
 745                 spin_unlock(&request->rq_lock);
 746                 request->rq_status = -ENODEV;
 747                 RETURN(-ENODEV);
 748         }
 749
 750         connection = imp->imp_connection;
 751
 752         lustre_msg_set_handle(request->rq_reqmsg,
 753                               &imp->imp_remote_handle);
 754         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 755         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 756                                 imp->imp_conn_cnt);
 757         lustre_msghdr_set_flags(request->rq_reqmsg,
 758                                 imp->imp_msghdr_flags);
 759
 760         /* If it's the first time to resend the request for EINPROGRESS,
 761          * we need to allocate a new XID (see after_reply()), it's different
 762          * from the resend for reply timeout. */
 763         if (request->rq_nr_resend != 0 &&
 764             list_empty(&request->rq_unreplied_list)) {
 765                 __u64 min_xid = 0;
 766                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 767                  * reconstruction */
 768                 spin_lock(&imp->imp_lock);
 769                 ptlrpc_assign_next_xid_nolock(request);
 770                 min_xid = ptlrpc_known_replied_xid(imp);
 771                 spin_unlock(&imp->imp_lock);
 772
 773                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 774                 DEBUG_REQ(D_RPCTRACE, request,
 775                           "Allocating new XID for resend on EINPROGRESS");
 776         }
 777
 778         opc = lustre_msg_get_opc(request->rq_reqmsg);
 779         if (opc != OST_CONNECT && opc != MDS_CONNECT &&
 780             opc != MGS_CONNECT && OCD_HAS_FLAG(&imp->imp_connect_data, FLAGS2))
 781                 rep_mbits = imp->imp_connect_data.ocd_connect_flags2 &
 782                         OBD_CONNECT2_REP_MBITS;
 783
 784         if ((request->rq_bulk != NULL) || rep_mbits) {
 785                 ptlrpc_set_mbits(request);
 786                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 787         }
 788
 789         if (list_empty(&request->rq_unreplied_list) ||
 790             request->rq_xid <= imp->imp_known_replied_xid) {
 791                 DEBUG_REQ(D_ERROR, request,
 792                           "xid=%llu, replied=%llu, list_empty=%d",
 793                           request->rq_xid, imp->imp_known_replied_xid,
 794                           list_empty(&request->rq_unreplied_list));
 795                 LBUG();
 796         }
 797
 798         /** For enabled AT all request should have AT_SUPPORT in the
 799          * FULL import state when OBD_CONNECT_AT is set */
 800         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 801                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 802                 !(imp->imp_connect_data.ocd_connect_flags &
 803                   OBD_CONNECT_AT));
 804
 805         if (request->rq_resend) {
 806                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 807                 if (request->rq_resend_cb != NULL)
 808                         request->rq_resend_cb(request, &request->rq_async_args);
 809         }
 810         if (request->rq_memalloc)
 811                 mpflag = memalloc_noreclaim_save();
 812
 813         rc = sptlrpc_cli_wrap_request(request);
 814         if (rc)
 815                 GOTO(out, rc);
 816
 817         /* bulk register should be done after wrap_request() */
 818         if (request->rq_bulk != NULL) {
 819                 rc = ptlrpc_register_bulk (request);
 820                 if (rc != 0)
 821                         GOTO(cleanup_bulk, rc);
 822                 /*
 823                  * All the mds in the request will have the same cpt
 824                  * encoded in the cookie. So we can just get the first
 825                  * one.
 826                  */
 827                 bulk_cookie = request->rq_bulk->bd_mds[0];
 828         }
 829
 830         if (!noreply) {
 831                 LASSERT (request->rq_replen != 0);
 832                 if (request->rq_repbuf == NULL) {
 833                         LASSERT(request->rq_repdata == NULL);
 834                         LASSERT(request->rq_repmsg == NULL);
 835                         rc = sptlrpc_cli_alloc_repbuf(request,
 836                                                       request->rq_replen);
 837                         if (rc) {
 838                                 /* this prevents us from looping in
 839                                  * ptlrpc_queue_wait */
 840                                 spin_lock(&request->rq_lock);
 841                                 request->rq_err = 1;
 842                                 spin_unlock(&request->rq_lock);
 843                                 request->rq_status = rc;
 844                                 GOTO(cleanup_bulk, rc);
 845                         }
 846                 } else {
 847                         request->rq_repdata = NULL;
 848                         request->rq_repmsg = NULL;
 849                 }
 850
 851                 peer.pid = connection->c_peer.pid;
 852                 lnet_nid4_to_nid(connection->c_peer.nid, &peer.nid);
 853                 if (request->rq_bulk &&
 854                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
 855                         reply_me = ERR_PTR(-ENOMEM);
 856                 } else {
 857                         reply_me = LNetMEAttach(request->rq_reply_portal,
 858                                                 &peer,
 859                                                 rep_mbits ? request->rq_mbits :
 860                                                 request->rq_xid,
 861                                                 0, LNET_UNLINK, LNET_INS_AFTER);
 862                 }
 863
 864                 if (IS_ERR(reply_me)) {
 865                         rc = PTR_ERR(reply_me);
 866                         CERROR("LNetMEAttach failed: %d\n", rc);
 867                         LASSERT(rc == -ENOMEM);
 868                         GOTO(cleanup_bulk, rc = -ENOMEM);
 869                 }
 870         }
 871
 872         spin_lock(&request->rq_lock);
 873         /* We are responsible for unlinking the reply buffer */
 874         request->rq_reply_unlinked = noreply;
 875         request->rq_receiving_reply = !noreply;
 876         /* Clear any flags that may be present from previous sends. */
 877         request->rq_req_unlinked = 0;
 878         request->rq_replied = 0;
 879         request->rq_err = 0;
 880         request->rq_timedout = 0;
 881         request->rq_net_err = 0;
 882         request->rq_resend = 0;
 883         request->rq_restart = 0;
 884         request->rq_reply_truncated = 0;
 885         spin_unlock(&request->rq_lock);
 886
 887         if (!noreply) {
 888                 reply_md.start     = request->rq_repbuf;
 889                 reply_md.length    = request->rq_repbuf_len;
 890                 /* Allow multiple early replies */
 891                 reply_md.threshold = LNET_MD_THRESH_INF;
 892                 /* Manage remote for early replies */
 893                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 894                         LNET_MD_MANAGE_REMOTE |
 895                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 896                 reply_md.user_ptr  = &request->rq_reply_cbid;
 897                 reply_md.handler = ptlrpc_handler;
 898
 899                 /* We must see the unlink callback to set rq_reply_unlinked,
 900                  * so we can't auto-unlink */
 901                 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
 902                                   &request->rq_reply_md_h);
 903                 if (rc != 0) {
 904                         CERROR("LNetMDAttach failed: %d\n", rc);
 905                         LASSERT(rc == -ENOMEM);
 906                         spin_lock(&request->rq_lock);
 907                         /* ...but the MD attach didn't succeed... */
 908                         request->rq_receiving_reply = 0;
 909                         spin_unlock(&request->rq_lock);
 910                         GOTO(cleanup_bulk, rc = -ENOMEM);
 911                 }
 912                 percpu_ref_get(&ptlrpc_pending);
 913
 914                 CDEBUG(D_NET,
 915                        "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 916                        request->rq_repbuf_len, request->rq_xid,
 917                        request->rq_reply_portal);
 918         }
 919
 920         /* add references on request for request_out_callback */
 921         ptlrpc_request_addref(request);
 922         if (obd != NULL && obd->obd_svc_stats != NULL)
 923                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 924                                     atomic_read(&imp->imp_inflight));
 925
 926         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 927
 928         request->rq_sent_ns = ktime_get_real();
 929         request->rq_sent = ktime_get_real_seconds();
 930         /* We give the server rq_timeout secs to process the req, and
 931          * add the network latency for our local timeout.
 932          */
 933         request->rq_deadline = request->rq_sent + request->rq_timeout +
 934                 ptlrpc_at_get_net_latency(request);
 935
 936         DEBUG_REQ(D_INFO, request, "send flags=%x",
 937                   lustre_msg_get_flags(request->rq_reqmsg));
 938         rc = ptl_send_buf(&request->rq_req_md_h,
 939                           request->rq_reqbuf, request->rq_reqdata_len,
 940                           LNET_NOACK_REQ, &request->rq_req_cbid,
 941                           LNET_NID_ANY, connection->c_peer,
 942                           request->rq_request_portal,
 943                           request->rq_xid, 0, &bulk_cookie);
 944         if (likely(rc == 0))
 945                 GOTO(out, rc);
 946
 947         request->rq_req_unlinked = 1;
 948         ptlrpc_req_finished(request);
 949         if (noreply)
 950                 GOTO(out, rc);
 951
 952         LNetMDUnlink(request->rq_reply_md_h);
 953
 954         /* UNLINKED callback called synchronously */
 955         LASSERT(!request->rq_receiving_reply);
 956
 957  cleanup_bulk:
 958         /* We do sync unlink here as there was no real transfer here so
 959          * the chance to have long unlink to sluggish net is smaller here. */
 960         ptlrpc_unregister_bulk(request, 0);
 961  out:
 962         if (rc == -ENOMEM) {
 963                 /* set rq_sent so that this request is treated
 964                  * as a delayed send in the upper layers */
 965                 request->rq_sent = ktime_get_real_seconds();
 966         }
 967
 968         if (request->rq_memalloc)
 969                 memalloc_noreclaim_restore(mpflag);
 970
 971         return rc;
 972 }
 973 EXPORT_SYMBOL(ptl_send_rpc);
 974
 975 /**
 976  * Register request buffer descriptor for request receiving.
 977  */
 978 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 979 {
 980         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 981         static struct lnet_processid match_id = {
 982                 .nid = LNET_ANY_NID,
 983                 .pid = LNET_PID_ANY
 984         };
 985         int rc;
 986         struct lnet_md md;
 987         struct lnet_me *me;
 988
 989         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 990                service->srv_req_portal);
 991
 992         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 993                 return (-ENOMEM);
 994
 995         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 996          * which means buffer can only be attached on local CPT, and LND
 997          * threads can find it by grabbing a local lock */
 998         me = LNetMEAttach(service->srv_req_portal,
 999                           &match_id, 0, ~0, LNET_UNLINK,
1000                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
1001                           LNET_INS_LOCAL : LNET_INS_AFTER);
1002         if (IS_ERR(me)) {
1003                 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
1004                 return -ENOMEM;
1005         }
1006
1007         LASSERT(rqbd->rqbd_refcount == 0);
1008         rqbd->rqbd_refcount = 1;
1009
1010         md.start     = rqbd->rqbd_buffer;
1011         md.length    = service->srv_buf_size;
1012         md.max_size  = service->srv_max_req_size;
1013         md.threshold = LNET_MD_THRESH_INF;
1014         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
1015         md.user_ptr  = &rqbd->rqbd_cbid;
1016         md.handler   = ptlrpc_handler;
1017
1018         rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
1019         if (rc == 0) {
1020                 percpu_ref_get(&ptlrpc_pending);
1021                 return 0;
1022         }
1023
1024         CERROR("ptlrpc: LNetMDAttach failed: rc = %d\n", rc);
1025         LASSERT(rc == -ENOMEM);
1026         LASSERT(rc == 0);
1027         rqbd->rqbd_refcount = 0;
1028
1029         return -ENOMEM;
1030 }