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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_RPC
  34 #include <libcfs/linux/linux-mem.h>
  35 #include <obd_support.h>
  36 #include <lustre_net.h>
  37 #include <lustre_lib.h>
  38 #include <obd.h>
  39 #include <obd_class.h>
  40 #include "ptlrpc_internal.h"
  41 #include <lnet/lib-lnet.h> /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */
  42
  43 /**
  44  * Helper function. Sends \a len bytes from \a base at offset \a offset
  45  * over \a conn connection to portal \a portal.
  46  * Returns 0 on success or error code.
  47  */
  48 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
  49                         enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
  50                         lnet_nid_t self, struct lnet_process_id peer_id,
  51                         int portal, __u64 xid, unsigned int offset,
  52                         struct lnet_handle_md *bulk_cookie)
  53 {
  54         int              rc;
  55         struct lnet_md         md;
  56         ENTRY;
  57
  58         LASSERT (portal != 0);
  59         CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
  60         md.start     = base;
  61         md.length    = len;
  62         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  63         md.options   = PTLRPC_MD_OPTIONS;
  64         md.user_ptr  = cbid;
  65         md.handler   = ptlrpc_handler;
  66         LNetInvalidateMDHandle(&md.bulk_handle);
  67
  68         if (bulk_cookie) {
  69                 md.bulk_handle = *bulk_cookie;
  70                 md.options |= LNET_MD_BULK_HANDLE;
  71         }
  72
  73         if (unlikely(ack == LNET_ACK_REQ &&
  74                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  75                 /* don't ask for the ack to simulate failing client */
  76                 ack = LNET_NOACK_REQ;
  77         }
  78
  79         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
  80         if (unlikely(rc != 0)) {
  81                 CERROR ("LNetMDBind failed: %d\n", rc);
  82                 LASSERT (rc == -ENOMEM);
  83                 RETURN (-ENOMEM);
  84         }
  85
  86         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  87                len, portal, xid, offset);
  88
  89         percpu_ref_get(&ptlrpc_pending);
  90
  91         rc = LNetPut(self, *mdh, ack,
  92                      peer_id, portal, xid, offset, 0);
  93         if (unlikely(rc != 0)) {
  94                 int rc2;
  95                 /* We're going to get an UNLINK event when I unlink below,
  96                  * which will complete just like any other failed send, so
  97                  * I fall through and return success here! */
  98                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  99                        libcfs_id2str(peer_id), portal, xid, rc);
 100                 rc2 = LNetMDUnlink(*mdh);
 101                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
 102         }
 103
 104         RETURN (0);
 105 }
 106
 107 #define mdunlink_iterate_helper(mds, count) \
 108                 __mdunlink_iterate_helper(mds, count, false)
 109 static void __mdunlink_iterate_helper(struct lnet_handle_md *bd_mds,
 110                                       int count, bool discard)
 111 {
 112         int i;
 113
 114         for (i = 0; i < count; i++)
 115                 __LNetMDUnlink(bd_mds[i], discard);
 116 }
 117
 118 #ifdef HAVE_SERVER_SUPPORT
 119 /**
 120  * Prepare bulk descriptor for specified incoming request \a req that
 121  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 122  * the bulk to be sent. Used on server-side after request was already
 123  * received.
 124  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 125  * error.
 126  */
 127 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 128                                               unsigned nfrags, unsigned max_brw,
 129                                               unsigned int type,
 130                                               unsigned portal,
 131                                               const struct ptlrpc_bulk_frag_ops
 132                                                 *ops)
 133 {
 134         struct obd_export *exp = req->rq_export;
 135         struct ptlrpc_bulk_desc *desc;
 136
 137         ENTRY;
 138         LASSERT(ptlrpc_is_bulk_op_active(type));
 139
 140         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 141         if (desc == NULL)
 142                 RETURN(NULL);
 143
 144         desc->bd_export = class_export_get(exp);
 145         desc->bd_req = req;
 146
 147         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 148         desc->bd_cbid.cbid_arg = desc;
 149
 150         /* NB we don't assign rq_bulk here; server-side requests are
 151          * re-used, and the handler frees the bulk desc explicitly. */
 152
 153         return desc;
 154 }
 155 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 156
 157 /**
 158  * Starts bulk transfer for descriptor \a desc on the server.
 159  * Returns 0 on success or error code.
 160  */
 161 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 162 {
 163         struct obd_export        *exp = desc->bd_export;
 164         lnet_nid_t                self_nid;
 165         struct lnet_process_id    peer_id;
 166         int                       rc = 0;
 167         __u64                     mbits;
 168         int                       posted_md;
 169         int                       total_md;
 170         struct lnet_md                 md;
 171         ENTRY;
 172
 173         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 174                 RETURN(0);
 175
 176         /* NB no locking required until desc is on the network */
 177         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 178
 179         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 180         LASSERT(desc->bd_cbid.cbid_arg == desc);
 181
 182         /*
 183          * Multi-Rail: get the preferred self and peer NIDs from the
 184          * request, so they are based on the route taken by the
 185          * message.
 186          */
 187         self_nid = desc->bd_req->rq_self;
 188         peer_id = desc->bd_req->rq_source;
 189
 190         /* NB total length may be 0 for a read past EOF, so we send 0
 191          * length bulks, since the client expects bulk events.
 192          *
 193          * The client may not need all of the bulk mbits for the RPC. The RPC
 194          * used the mbits of the highest bulk mbits needed, and the server masks
 195          * off high bits to get bulk count for this RPC. LU-1431 */
 196         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 197         total_md = desc->bd_req->rq_mbits - mbits + 1;
 198         desc->bd_refs = total_md;
 199         desc->bd_failure = 0;
 200
 201         md.user_ptr = &desc->bd_cbid;
 202         md.handler = ptlrpc_handler;
 203         md.threshold = 2; /* SENT and ACK/REPLY */
 204
 205         for (posted_md = 0; posted_md < total_md; mbits++) {
 206                 md.options = PTLRPC_MD_OPTIONS;
 207
 208                 /* NB it's assumed that source and sink buffer frags are
 209                  * page-aligned. Otherwise we'd have to send client bulk
 210                  * sizes over and split server buffer accordingly */
 211                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 212                 rc = LNetMDBind(&md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 213                 if (rc != 0) {
 214                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 215                                exp->exp_obd->obd_name, posted_md, rc);
 216                         LASSERT(rc == -ENOMEM);
 217                         if (posted_md == 0) {
 218                                 desc->bd_md_count = 0;
 219                                 RETURN(-ENOMEM);
 220                         }
 221                         break;
 222                 }
 223                 percpu_ref_get(&ptlrpc_pending);
 224
 225                 /* sanity.sh 224c: lets skip last md */
 226                 if (posted_md == desc->bd_md_max_brw - 1)
 227                         OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 228                                              CFS_FAIL_PTLRPC_OST_BULK_CB2);
 229
 230                 /* Network is about to get at the memory */
 231                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 232                         rc = LNetPut(self_nid, desc->bd_mds[posted_md],
 233                                      LNET_ACK_REQ, peer_id,
 234                                      desc->bd_portal, mbits, 0, 0);
 235                 else
 236                         rc = LNetGet(self_nid, desc->bd_mds[posted_md],
 237                                      peer_id, desc->bd_portal, mbits, 0, false);
 238
 239                 posted_md++;
 240                 if (rc != 0) {
 241                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 242                                "rc = %d\n", exp->exp_obd->obd_name,
 243                                libcfs_id2str(peer_id), desc->bd_portal,
 244                                mbits, rc);
 245                         break;
 246                 }
 247         }
 248
 249         if (rc != 0) {
 250                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 251                  * event this creates will signal completion with failure,
 252                  * so we return SUCCESS here! */
 253                 spin_lock(&desc->bd_lock);
 254                 desc->bd_refs -= total_md - posted_md;
 255                 spin_unlock(&desc->bd_lock);
 256                 LASSERT(desc->bd_refs >= 0);
 257
 258                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 259                 RETURN(0);
 260         }
 261
 262         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 263                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 264                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
 265                mbits - posted_md, mbits - 1);
 266
 267         RETURN(0);
 268 }
 269
 270 /**
 271  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 272  * serialises with completion callback)
 273  */
 274 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 275 {
 276         LASSERT(!in_interrupt());           /* might sleep */
 277
 278         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 279                 return;                         /* never started */
 280
 281         /* We used to poison the pages with 0xab here because we did not want to
 282          * send any meaningful data over the wire for evicted clients (bug 9297)
 283          * However, this is no longer safe now that we use the page cache on the
 284          * OSS (bug 20560) */
 285
 286         /* The unlink ensures the callback happens ASAP and is the last
 287          * one.  If it fails, it must be because completion just happened,
 288          * but we must still wait_event_idle_timeout() in this case, to give
 289          * us a chance to run server_bulk_callback()
 290          */
 291         __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true);
 292
 293         for (;;) {
 294                 /* Network access will complete in finite time but the HUGE
 295                  * timeout lets us CWARN for visibility of sluggish NALs */
 296                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 297
 298                 while (seconds > 0 &&
 299                        wait_event_idle_timeout(desc->bd_waitq,
 300                                                !ptlrpc_server_bulk_active(desc),
 301                                                cfs_time_seconds(1)) == 0)
 302                         seconds -= 1;
 303                 if (seconds > 0)
 304                         return;
 305
 306                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 307         }
 308 }
 309 #endif /* HAVE_SERVER_SUPPORT */
 310
 311 /**
 312  * Register bulk at the sender for later transfer.
 313  * Returns 0 on success or error code.
 314  */
 315 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 316 {
 317         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 318         struct lnet_process_id peer;
 319         int rc = 0;
 320         int posted_md;
 321         int total_md;
 322         __u64 mbits;
 323         struct lnet_me *me;
 324         struct lnet_md md;
 325         ENTRY;
 326
 327         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 328                 RETURN(0);
 329
 330         /* NB no locking required until desc is on the network */
 331         LASSERT(desc->bd_nob > 0);
 332         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 333         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 334         LASSERT(desc->bd_req != NULL);
 335         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 336
 337         /* cleanup the state of the bulk for it will be reused */
 338         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 339                 desc->bd_nob_transferred = 0;
 340         else if (desc->bd_nob_transferred != 0)
 341                 /* If the network failed after an RPC was sent, this condition
 342                  * could happen.  Rather than assert (was here before), return
 343                  * an EIO error. */
 344                 RETURN(-EIO);
 345
 346         desc->bd_failure = 0;
 347
 348         peer = desc->bd_import->imp_connection->c_peer;
 349
 350         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 351         LASSERT(desc->bd_cbid.cbid_arg == desc);
 352
 353         total_md = desc->bd_md_count;
 354         /* rq_mbits is matchbits of the final bulk */
 355         mbits = req->rq_mbits - desc->bd_md_count + 1;
 356
 357         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 358                  "first mbits = x%llu, last mbits = x%llu\n",
 359                  mbits, req->rq_mbits);
 360         LASSERTF(!(desc->bd_registered &&
 361                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 362                  mbits != desc->bd_last_mbits,
 363                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 364                  desc->bd_registered, mbits, desc->bd_last_mbits);
 365
 366         desc->bd_registered = 1;
 367         desc->bd_last_mbits = mbits;
 368         desc->bd_refs = total_md;
 369         md.user_ptr = &desc->bd_cbid;
 370         md.handler = ptlrpc_handler;
 371         md.threshold = 1;                       /* PUT or GET */
 372
 373         for (posted_md = 0; posted_md < desc->bd_md_count;
 374              posted_md++, mbits++) {
 375                 md.options = PTLRPC_MD_OPTIONS |
 376                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 377                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 378                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 379
 380                 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
 381                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 382                         rc = -ENOMEM;
 383                 } else {
 384                         me = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 385                                   LNET_UNLINK, LNET_INS_AFTER);
 386                         rc = PTR_ERR_OR_ZERO(me);
 387                 }
 388                 if (rc != 0) {
 389                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 390                                desc->bd_import->imp_obd->obd_name, mbits,
 391                                posted_md, rc);
 392                         break;
 393                 }
 394                 percpu_ref_get(&ptlrpc_pending);
 395
 396                 /* About to let the network at it... */
 397                 rc = LNetMDAttach(me, &md, LNET_UNLINK,
 398                                   &desc->bd_mds[posted_md]);
 399                 if (rc != 0) {
 400                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 401                                desc->bd_import->imp_obd->obd_name, mbits,
 402                                posted_md, rc);
 403                         break;
 404                 }
 405         }
 406
 407         if (rc != 0) {
 408                 LASSERT(rc == -ENOMEM);
 409                 spin_lock(&desc->bd_lock);
 410                 desc->bd_refs -= total_md - posted_md;
 411                 spin_unlock(&desc->bd_lock);
 412                 LASSERT(desc->bd_refs >= 0);
 413                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 414                 req->rq_status = -ENOMEM;
 415                 desc->bd_registered = 0;
 416                 RETURN(-ENOMEM);
 417         }
 418
 419         spin_lock(&desc->bd_lock);
 420         /* Holler if peer manages to touch buffers before he knows the mbits */
 421         if (desc->bd_refs != total_md)
 422                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 423                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 424                       total_md - desc->bd_refs);
 425         spin_unlock(&desc->bd_lock);
 426
 427         CDEBUG(D_NET,
 428                "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
 429                desc->bd_refs,
 430                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 431                desc->bd_iov_count, desc->bd_nob,
 432                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 433
 434         RETURN(0);
 435 }
 436
 437 /**
 438  * Disconnect a bulk desc from the network. Idempotent. Not
 439  * thread-safe (i.e. only interlocks with completion callback).
 440  * Returns 1 on success or 0 if network unregistration failed for whatever
 441  * reason.
 442  */
 443 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 444 {
 445         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 446         ENTRY;
 447
 448         LASSERT(!in_interrupt());     /* might sleep */
 449
 450         if (desc)
 451                 desc->bd_registered = 0;
 452
 453         /* Let's setup deadline for reply unlink. */
 454         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 455             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 456                 req->rq_bulk_deadline = ktime_get_real_seconds() +
 457                                         PTLRPC_REQ_LONG_UNLINK;
 458
 459         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 460                 RETURN(1);                              /* never registered */
 461
 462         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 463
 464         /* the unlink ensures the callback happens ASAP and is the last
 465          * one.  If it fails, it must be because completion just happened,
 466          * but we must still wait_event_idle_timeout() in this case to give
 467          * us a chance to run client_bulk_callback()
 468          */
 469         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 470
 471         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 472                 RETURN(1);                              /* never registered */
 473
 474         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 475         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 476
 477         /* Do not wait for unlink to finish. */
 478         if (async)
 479                 RETURN(0);
 480
 481         for (;;) {
 482                 /* The wq argument is ignored by user-space wait_event macros */
 483                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 484                                         &req->rq_set->set_waitq :
 485                                         &req->rq_reply_waitq;
 486                 /*
 487                  * Network access will complete in finite time but the HUGE
 488                  * timeout lets us CWARN for visibility of sluggish NALs.
 489                  */
 490                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 491
 492                 while (seconds > 0 &&
 493                        wait_event_idle_timeout(*wq,
 494                                                !ptlrpc_client_bulk_active(req),
 495                                                cfs_time_seconds(1)) == 0)
 496                         seconds -= 1;
 497                 if (seconds > 0) {
 498                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 499                         RETURN(1);
 500                 }
 501
 502                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 503                           desc);
 504         }
 505         RETURN(0);
 506 }
 507
 508 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 509 {
 510         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 511         struct ptlrpc_service           *svc = svcpt->scp_service;
 512         timeout_t service_timeout;
 513
 514         service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
 515                                              req->rq_arrival_time.tv_sec, 1,
 516                                   (AT_OFF ? obd_timeout * 3 / 2 : at_max));
 517         if (!(flags & PTLRPC_REPLY_EARLY) &&
 518             (req->rq_type != PTL_RPC_MSG_ERR) &&
 519             (req->rq_reqmsg != NULL) &&
 520             !(lustre_msg_get_flags(req->rq_reqmsg) &
 521               (MSG_RESENT | MSG_REPLAY |
 522                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 523                 /* early replies, errors and recovery requests don't count
 524                  * toward our service time estimate
 525                  */
 526                 timeout_t oldse = at_measured(&svcpt->scp_at_estimate,
 527                                               service_timeout);
 528
 529                 if (oldse != 0) {
 530                         DEBUG_REQ(D_ADAPTTO, req,
 531                                   "svc %s changed estimate from %d to %d",
 532                                   svc->srv_name, oldse,
 533                                   at_get(&svcpt->scp_at_estimate));
 534                 }
 535         }
 536         /* Report actual service time for client latency calc */
 537         lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
 538         /* Report service time estimate for future client reqs, but report 0
 539          * (to be ignored by client) if it's an error reply during recovery.
 540          * b=15815
 541          */
 542         if (req->rq_type == PTL_RPC_MSG_ERR &&
 543             (req->rq_export == NULL ||
 544              req->rq_export->exp_obd->obd_recovering)) {
 545                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 546         } else {
 547                 timeout_t timeout;
 548
 549                 if (req->rq_export && req->rq_reqmsg != NULL &&
 550                     (flags & PTLRPC_REPLY_EARLY) &&
 551                     lustre_msg_get_flags(req->rq_reqmsg) &
 552                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 553                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 554
 555                         timeout = ktime_get_real_seconds() -
 556                                   req->rq_arrival_time.tv_sec +
 557                                   min_t(timeout_t, at_extra,
 558                                         exp_obd->obd_recovery_timeout / 4);
 559                 } else {
 560                         timeout = at_get(&svcpt->scp_at_estimate);
 561                 }
 562                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 563         }
 564
 565         if (req->rq_reqmsg &&
 566             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 567                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 568                        "req_flags=%#x magic=%x/%x len=%d\n",
 569                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 570                        lustre_msg_get_magic(req->rq_reqmsg),
 571                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 572         }
 573 }
 574
 575 /**
 576  * Send request reply from request \a req reply buffer.
 577  * \a flags defines reply types
 578  * Returns 0 on success or error code
 579  */
 580 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 581 {
 582         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 583         struct ptlrpc_connection  *conn;
 584         int                        rc;
 585
 586         /* We must already have a reply buffer (only ptlrpc_error() may be
 587          * called without one). The reply generated by sptlrpc layer (e.g.
 588          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 589          * have a request buffer which is either the actual (swabbed) incoming
 590          * request, or a saved copy if this is a req saved in
 591          * target_queue_final_reply().
 592          */
 593         LASSERT (req->rq_no_reply == 0);
 594         LASSERT (req->rq_reqbuf != NULL);
 595         LASSERT (rs != NULL);
 596         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 597         LASSERT (req->rq_repmsg != NULL);
 598         LASSERT (req->rq_repmsg == rs->rs_msg);
 599         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 600         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 601
 602         /* There may be no rq_export during failover */
 603
 604         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 605                      req->rq_export->exp_obd->obd_fail)) {
 606                 /* Failed obd's only send ENODEV */
 607                 req->rq_type = PTL_RPC_MSG_ERR;
 608                 req->rq_status = -ENODEV;
 609                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 610                        req->rq_export->exp_obd->obd_minor);
 611         }
 612
 613         if (req->rq_type != PTL_RPC_MSG_ERR)
 614                 req->rq_type = PTL_RPC_MSG_REPLY;
 615
 616         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 617         lustre_msg_set_status(req->rq_repmsg,
 618                               ptlrpc_status_hton(req->rq_status));
 619         lustre_msg_set_opc(req->rq_repmsg,
 620                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 621
 622         target_pack_pool_reply(req);
 623
 624         ptlrpc_at_set_reply(req, flags);
 625
 626         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 627                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 628         else
 629                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 630
 631         if (unlikely(conn == NULL)) {
 632                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 633                 return -ENOTCONN;
 634         }
 635         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 636
 637         rc = sptlrpc_svc_wrap_reply(req);
 638         if (unlikely(rc))
 639                 goto out;
 640
 641         req->rq_sent = ktime_get_real_seconds();
 642
 643         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 644                           (rs->rs_difficult && !rs->rs_no_ack) ?
 645                           LNET_ACK_REQ : LNET_NOACK_REQ,
 646                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 647                           ptlrpc_req2svc(req)->srv_rep_portal,
 648                           req->rq_xid, req->rq_reply_off, NULL);
 649 out:
 650         if (unlikely(rc != 0))
 651                 ptlrpc_req_drop_rs(req);
 652         ptlrpc_connection_put(conn);
 653         return rc;
 654 }
 655
 656 int ptlrpc_reply (struct ptlrpc_request *req)
 657 {
 658         if (req->rq_no_reply)
 659                 return 0;
 660         else
 661                 return (ptlrpc_send_reply(req, 0));
 662 }
 663
 664 /**
 665  * For request \a req send an error reply back. Create empty
 666  * reply buffers if necessary.
 667  */
 668 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 669 {
 670         int rc;
 671         ENTRY;
 672
 673         if (req->rq_no_reply)
 674                 RETURN(0);
 675
 676         if (!req->rq_repmsg) {
 677                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 678                 if (rc)
 679                         RETURN(rc);
 680         }
 681
 682         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 683             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 684             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 685                 req->rq_type = PTL_RPC_MSG_ERR;
 686
 687         rc = ptlrpc_send_reply(req, may_be_difficult);
 688         RETURN(rc);
 689 }
 690
 691 int ptlrpc_error(struct ptlrpc_request *req)
 692 {
 693         return ptlrpc_send_error(req, 0);
 694 }
 695
 696 /**
 697  * Send request \a request.
 698  * if \a noreply is set, don't expect any reply back and don't set up
 699  * reply buffers.
 700  * Returns 0 on success or error code.
 701  */
 702 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 703 {
 704         int rc;
 705         int mpflag = 0;
 706         struct lnet_handle_md bulk_cookie;
 707         struct ptlrpc_connection *connection;
 708         struct lnet_me *reply_me = NULL;
 709         struct lnet_md reply_md;
 710         struct obd_import *imp = request->rq_import;
 711         struct obd_device *obd = imp->imp_obd;
 712         ENTRY;
 713
 714         LNetInvalidateMDHandle(&bulk_cookie);
 715
 716         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 717                 RETURN(0);
 718
 719         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) &&
 720                      lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT &&
 721                      strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) {
 722                 RETURN(0);
 723         }
 724
 725         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 726         LASSERT(request->rq_wait_ctx == 0);
 727
 728         /* If this is a re-transmit, we're required to have disengaged
 729          * cleanly from the previous attempt */
 730         LASSERT(!request->rq_receiving_reply);
 731         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 732                   (imp->imp_state == LUSTRE_IMP_FULL)));
 733
 734         if (unlikely(obd != NULL && obd->obd_fail)) {
 735                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 736                        obd->obd_name);
 737                 /* this prevents us from waiting in ptlrpc_queue_wait */
 738                 spin_lock(&request->rq_lock);
 739                 request->rq_err = 1;
 740                 spin_unlock(&request->rq_lock);
 741                 request->rq_status = -ENODEV;
 742                 RETURN(-ENODEV);
 743         }
 744
 745         connection = imp->imp_connection;
 746
 747         lustre_msg_set_handle(request->rq_reqmsg,
 748                               &imp->imp_remote_handle);
 749         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 750         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 751                                 imp->imp_conn_cnt);
 752         lustre_msghdr_set_flags(request->rq_reqmsg,
 753                                 imp->imp_msghdr_flags);
 754
 755         /* If it's the first time to resend the request for EINPROGRESS,
 756          * we need to allocate a new XID (see after_reply()), it's different
 757          * from the resend for reply timeout. */
 758         if (request->rq_nr_resend != 0 &&
 759             list_empty(&request->rq_unreplied_list)) {
 760                 __u64 min_xid = 0;
 761                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 762                  * reconstruction */
 763                 spin_lock(&imp->imp_lock);
 764                 ptlrpc_assign_next_xid_nolock(request);
 765                 min_xid = ptlrpc_known_replied_xid(imp);
 766                 spin_unlock(&imp->imp_lock);
 767
 768                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 769                 DEBUG_REQ(D_RPCTRACE, request,
 770                           "Allocating new XID for resend on EINPROGRESS");
 771         }
 772
 773         if (request->rq_bulk != NULL) {
 774                 ptlrpc_set_bulk_mbits(request);
 775                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 776         }
 777
 778         if (list_empty(&request->rq_unreplied_list) ||
 779             request->rq_xid <= imp->imp_known_replied_xid) {
 780                 DEBUG_REQ(D_ERROR, request,
 781                           "xid=%llu, replied=%llu, list_empty=%d",
 782                           request->rq_xid, imp->imp_known_replied_xid,
 783                           list_empty(&request->rq_unreplied_list));
 784                 LBUG();
 785         }
 786
 787         /** For enabled AT all request should have AT_SUPPORT in the
 788          * FULL import state when OBD_CONNECT_AT is set */
 789         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 790                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 791                 !(imp->imp_connect_data.ocd_connect_flags &
 792                   OBD_CONNECT_AT));
 793
 794         if (request->rq_resend) {
 795                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 796                 if (request->rq_resend_cb != NULL)
 797                         request->rq_resend_cb(request, &request->rq_async_args);
 798         }
 799         if (request->rq_memalloc)
 800                 mpflag = memalloc_noreclaim_save();
 801
 802         rc = sptlrpc_cli_wrap_request(request);
 803         if (rc)
 804                 GOTO(out, rc);
 805
 806         /* bulk register should be done after wrap_request() */
 807         if (request->rq_bulk != NULL) {
 808                 rc = ptlrpc_register_bulk (request);
 809                 if (rc != 0)
 810                         GOTO(cleanup_bulk, rc);
 811                 /*
 812                  * All the mds in the request will have the same cpt
 813                  * encoded in the cookie. So we can just get the first
 814                  * one.
 815                  */
 816                 bulk_cookie = request->rq_bulk->bd_mds[0];
 817         }
 818
 819         if (!noreply) {
 820                 LASSERT (request->rq_replen != 0);
 821                 if (request->rq_repbuf == NULL) {
 822                         LASSERT(request->rq_repdata == NULL);
 823                         LASSERT(request->rq_repmsg == NULL);
 824                         rc = sptlrpc_cli_alloc_repbuf(request,
 825                                                       request->rq_replen);
 826                         if (rc) {
 827                                 /* this prevents us from looping in
 828                                  * ptlrpc_queue_wait */
 829                                 spin_lock(&request->rq_lock);
 830                                 request->rq_err = 1;
 831                                 spin_unlock(&request->rq_lock);
 832                                 request->rq_status = rc;
 833                                 GOTO(cleanup_bulk, rc);
 834                         }
 835                 } else {
 836                         request->rq_repdata = NULL;
 837                         request->rq_repmsg = NULL;
 838                 }
 839
 840                 if (request->rq_bulk &&
 841                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
 842                         reply_me = ERR_PTR(-ENOMEM);
 843                 } else {
 844                         reply_me = LNetMEAttach(request->rq_reply_portal,
 845                                                 connection->c_peer,
 846                                                 request->rq_xid, 0,
 847                                                 LNET_UNLINK, LNET_INS_AFTER);
 848                 }
 849
 850                 if (IS_ERR(reply_me)) {
 851                         rc = PTR_ERR(reply_me);
 852                         CERROR("LNetMEAttach failed: %d\n", rc);
 853                         LASSERT(rc == -ENOMEM);
 854                         GOTO(cleanup_bulk, rc = -ENOMEM);
 855                 }
 856         }
 857
 858         spin_lock(&request->rq_lock);
 859         /* We are responsible for unlinking the reply buffer */
 860         request->rq_reply_unlinked = noreply;
 861         request->rq_receiving_reply = !noreply;
 862         /* Clear any flags that may be present from previous sends. */
 863         request->rq_req_unlinked = 0;
 864         request->rq_replied = 0;
 865         request->rq_err = 0;
 866         request->rq_timedout = 0;
 867         request->rq_net_err = 0;
 868         request->rq_resend = 0;
 869         request->rq_restart = 0;
 870         request->rq_reply_truncated = 0;
 871         spin_unlock(&request->rq_lock);
 872
 873         if (!noreply) {
 874                 reply_md.start     = request->rq_repbuf;
 875                 reply_md.length    = request->rq_repbuf_len;
 876                 /* Allow multiple early replies */
 877                 reply_md.threshold = LNET_MD_THRESH_INF;
 878                 /* Manage remote for early replies */
 879                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 880                         LNET_MD_MANAGE_REMOTE |
 881                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 882                 reply_md.user_ptr  = &request->rq_reply_cbid;
 883                 reply_md.handler = ptlrpc_handler;
 884
 885                 /* We must see the unlink callback to set rq_reply_unlinked,
 886                  * so we can't auto-unlink */
 887                 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
 888                                   &request->rq_reply_md_h);
 889                 if (rc != 0) {
 890                         CERROR("LNetMDAttach failed: %d\n", rc);
 891                         LASSERT(rc == -ENOMEM);
 892                         spin_lock(&request->rq_lock);
 893                         /* ...but the MD attach didn't succeed... */
 894                         request->rq_receiving_reply = 0;
 895                         spin_unlock(&request->rq_lock);
 896                         GOTO(cleanup_bulk, rc = -ENOMEM);
 897                 }
 898                 percpu_ref_get(&ptlrpc_pending);
 899
 900                 CDEBUG(D_NET,
 901                        "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 902                        request->rq_repbuf_len, request->rq_xid,
 903                        request->rq_reply_portal);
 904         }
 905
 906         /* add references on request for request_out_callback */
 907         ptlrpc_request_addref(request);
 908         if (obd != NULL && obd->obd_svc_stats != NULL)
 909                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 910                                     atomic_read(&imp->imp_inflight));
 911
 912         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 913
 914         request->rq_sent_ns = ktime_get_real();
 915         request->rq_sent = ktime_get_real_seconds();
 916         /* We give the server rq_timeout secs to process the req, and
 917          * add the network latency for our local timeout.
 918          */
 919         request->rq_deadline = request->rq_sent + request->rq_timeout +
 920                 ptlrpc_at_get_net_latency(request);
 921
 922         DEBUG_REQ(D_INFO, request, "send flags=%x",
 923                   lustre_msg_get_flags(request->rq_reqmsg));
 924         rc = ptl_send_buf(&request->rq_req_md_h,
 925                           request->rq_reqbuf, request->rq_reqdata_len,
 926                           LNET_NOACK_REQ, &request->rq_req_cbid,
 927                           LNET_NID_ANY, connection->c_peer,
 928                           request->rq_request_portal,
 929                           request->rq_xid, 0, &bulk_cookie);
 930         if (likely(rc == 0))
 931                 GOTO(out, rc);
 932
 933         request->rq_req_unlinked = 1;
 934         ptlrpc_req_finished(request);
 935         if (noreply)
 936                 GOTO(out, rc);
 937
 938         LNetMDUnlink(request->rq_reply_md_h);
 939
 940         /* UNLINKED callback called synchronously */
 941         LASSERT(!request->rq_receiving_reply);
 942
 943  cleanup_bulk:
 944         /* We do sync unlink here as there was no real transfer here so
 945          * the chance to have long unlink to sluggish net is smaller here. */
 946         ptlrpc_unregister_bulk(request, 0);
 947  out:
 948         if (rc == -ENOMEM) {
 949                 /* set rq_sent so that this request is treated
 950                  * as a delayed send in the upper layers */
 951                 request->rq_sent = ktime_get_real_seconds();
 952         }
 953
 954         if (request->rq_memalloc)
 955                 memalloc_noreclaim_restore(mpflag);
 956
 957         return rc;
 958 }
 959 EXPORT_SYMBOL(ptl_send_rpc);
 960
 961 /**
 962  * Register request buffer descriptor for request receiving.
 963  */
 964 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 965 {
 966         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 967         static struct lnet_process_id match_id = {
 968                 .nid = LNET_NID_ANY,
 969                 .pid = LNET_PID_ANY
 970         };
 971         int rc;
 972         struct lnet_md md;
 973         struct lnet_me *me;
 974
 975         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 976                service->srv_req_portal);
 977
 978         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 979                 return (-ENOMEM);
 980
 981         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 982          * which means buffer can only be attached on local CPT, and LND
 983          * threads can find it by grabbing a local lock */
 984         me = LNetMEAttach(service->srv_req_portal,
 985                           match_id, 0, ~0, LNET_UNLINK,
 986                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 987                           LNET_INS_LOCAL : LNET_INS_AFTER);
 988         if (IS_ERR(me)) {
 989                 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
 990                 return -ENOMEM;
 991         }
 992
 993         LASSERT(rqbd->rqbd_refcount == 0);
 994         rqbd->rqbd_refcount = 1;
 995
 996         md.start     = rqbd->rqbd_buffer;
 997         md.length    = service->srv_buf_size;
 998         md.max_size  = service->srv_max_req_size;
 999         md.threshold = LNET_MD_THRESH_INF;
1000         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
1001         md.user_ptr  = &rqbd->rqbd_cbid;
1002         md.handler   = ptlrpc_handler;
1003
1004         rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
1005         if (rc == 0) {
1006                 percpu_ref_get(&ptlrpc_pending);
1007                 return 0;
1008         }
1009
1010         CERROR("ptlrpc: LNetMDAttach failed: rc = %d\n", rc);
1011         LASSERT(rc == -ENOMEM);
1012         LASSERT(rc == 0);
1013         rqbd->rqbd_refcount = 0;
1014
1015         return -ENOMEM;
1016 }