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 self4, struct lnet_process_id peer_id4,
  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         struct lnet_nid self;
  56         struct lnet_processid peer_id;
  57         ENTRY;
  58
  59         lnet_nid4_to_nid(self4, &self);
  60         lnet_pid4_to_pid(peer_id4, &peer_id);
  61
  62         LASSERT(portal != 0);
  63         CDEBUG(D_INFO, "peer_id %s\n", libcfs_id2str(peer_id4));
  64         md.start     = base;
  65         md.length    = len;
  66         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  67         md.options   = PTLRPC_MD_OPTIONS;
  68         md.user_ptr  = cbid;
  69         md.handler   = ptlrpc_handler;
  70         LNetInvalidateMDHandle(&md.bulk_handle);
  71
  72         if (bulk_cookie) {
  73                 md.bulk_handle = *bulk_cookie;
  74                 md.options |= LNET_MD_BULK_HANDLE;
  75         }
  76
  77         if (unlikely(ack == LNET_ACK_REQ &&
  78                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  79                 /* don't ask for the ack to simulate failing client */
  80                 ack = LNET_NOACK_REQ;
  81         }
  82
  83         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
  84         if (unlikely(rc != 0)) {
  85                 CERROR ("LNetMDBind failed: %d\n", rc);
  86                 LASSERT (rc == -ENOMEM);
  87                 RETURN (-ENOMEM);
  88         }
  89
  90         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  91                len, portal, xid, offset);
  92
  93         percpu_ref_get(&ptlrpc_pending);
  94
  95         rc = LNetPut(&self, *mdh, ack,
  96                      &peer_id, portal, xid, offset, 0);
  97         if (unlikely(rc != 0)) {
  98                 int rc2;
  99                 /* We're going to get an UNLINK event when I unlink below,
 100                  * which will complete just like any other failed send, so
 101                  * I fall through and return success here! */
 102                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
 103                        libcfs_id2str(peer_id4), portal, xid, rc);
 104                 rc2 = LNetMDUnlink(*mdh);
 105                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
 106         }
 107
 108         RETURN (0);
 109 }
 110
 111 #define mdunlink_iterate_helper(mds, count) \
 112                 __mdunlink_iterate_helper(mds, count, false)
 113 static void __mdunlink_iterate_helper(struct lnet_handle_md *bd_mds,
 114                                       int count, bool discard)
 115 {
 116         int i;
 117
 118         for (i = 0; i < count; i++)
 119                 __LNetMDUnlink(bd_mds[i], discard);
 120 }
 121
 122 #ifdef HAVE_SERVER_SUPPORT
 123 /**
 124  * Prepare bulk descriptor for specified incoming request \a req that
 125  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 126  * the bulk to be sent. Used on server-side after request was already
 127  * received.
 128  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 129  * error.
 130  */
 131 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 132                                               unsigned nfrags, unsigned max_brw,
 133                                               unsigned int type,
 134                                               unsigned portal,
 135                                               const struct ptlrpc_bulk_frag_ops
 136                                                 *ops)
 137 {
 138         struct obd_export *exp = req->rq_export;
 139         struct ptlrpc_bulk_desc *desc;
 140
 141         ENTRY;
 142         LASSERT(ptlrpc_is_bulk_op_active(type));
 143
 144         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 145         if (desc == NULL)
 146                 RETURN(NULL);
 147
 148         desc->bd_export = class_export_get(exp);
 149         desc->bd_req = req;
 150
 151         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 152         desc->bd_cbid.cbid_arg = desc;
 153
 154         /* NB we don't assign rq_bulk here; server-side requests are
 155          * re-used, and the handler frees the bulk desc explicitly. */
 156
 157         return desc;
 158 }
 159 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 160
 161 /**
 162  * Starts bulk transfer for descriptor \a desc on the server.
 163  * Returns 0 on success or error code.
 164  */
 165 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 166 {
 167         struct obd_export       *exp = desc->bd_export;
 168         lnet_nid_t               self_nid4;
 169         struct lnet_nid          self_nid;
 170         struct lnet_process_id   peer_id4;
 171         struct lnet_processid    peer_id;
 172         int                       rc = 0;
 173         __u64                     mbits;
 174         int                       posted_md;
 175         int                      total_md;
 176         struct lnet_md           md;
 177         ENTRY;
 178
 179         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 180                 RETURN(0);
 181
 182         /* NB no locking required until desc is on the network */
 183         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 184
 185         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 186         LASSERT(desc->bd_cbid.cbid_arg == desc);
 187
 188         /*
 189          * Multi-Rail: get the preferred self and peer NIDs from the
 190          * request, so they are based on the route taken by the
 191          * message.
 192          */
 193         self_nid4 = desc->bd_req->rq_self;
 194         peer_id4 = desc->bd_req->rq_source;
 195         lnet_nid4_to_nid(self_nid4, &self_nid);
 196         lnet_pid4_to_pid(peer_id4, &peer_id);
 197
 198         /* NB total length may be 0 for a read past EOF, so we send 0
 199          * length bulks, since the client expects bulk events.
 200          *
 201          * The client may not need all of the bulk mbits for the RPC. The RPC
 202          * used the mbits of the highest bulk mbits needed, and the server masks
 203          * off high bits to get bulk count for this RPC. LU-1431 */
 204         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 205         total_md = desc->bd_req->rq_mbits - mbits + 1;
 206         desc->bd_refs = total_md;
 207         desc->bd_failure = 0;
 208
 209         md.user_ptr = &desc->bd_cbid;
 210         md.handler = ptlrpc_handler;
 211         md.threshold = 2; /* SENT and ACK/REPLY */
 212
 213         for (posted_md = 0; posted_md < total_md; mbits++) {
 214                 md.options = PTLRPC_MD_OPTIONS;
 215
 216                 /* NB it's assumed that source and sink buffer frags are
 217                  * page-aligned. Otherwise we'd have to send client bulk
 218                  * sizes over and split server buffer accordingly */
 219                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 220                 rc = LNetMDBind(&md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 221                 if (rc != 0) {
 222                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 223                                exp->exp_obd->obd_name, posted_md, rc);
 224                         LASSERT(rc == -ENOMEM);
 225                         if (posted_md == 0) {
 226                                 desc->bd_md_count = 0;
 227                                 RETURN(-ENOMEM);
 228                         }
 229                         break;
 230                 }
 231                 percpu_ref_get(&ptlrpc_pending);
 232
 233                 /* sanity.sh 224c: lets skip last md */
 234                 if (posted_md == desc->bd_md_max_brw - 1)
 235                         OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 236                                              CFS_FAIL_PTLRPC_OST_BULK_CB2);
 237
 238                 /* Network is about to get at the memory */
 239                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 240                         rc = LNetPut(&self_nid, desc->bd_mds[posted_md],
 241                                      LNET_ACK_REQ, &peer_id,
 242                                      desc->bd_portal, mbits, 0, 0);
 243                 else
 244                         rc = LNetGet(self_nid4, desc->bd_mds[posted_md],
 245                                      peer_id4, desc->bd_portal,
 246                                      mbits, 0, false);
 247
 248                 posted_md++;
 249                 if (rc != 0) {
 250                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 251                                "rc = %d\n", exp->exp_obd->obd_name,
 252                                libcfs_id2str(peer_id4), desc->bd_portal,
 253                                mbits, rc);
 254                         break;
 255                 }
 256         }
 257
 258         if (rc != 0) {
 259                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 260                  * event this creates will signal completion with failure,
 261                  * so we return SUCCESS here! */
 262                 spin_lock(&desc->bd_lock);
 263                 desc->bd_refs -= total_md - posted_md;
 264                 spin_unlock(&desc->bd_lock);
 265                 LASSERT(desc->bd_refs >= 0);
 266
 267                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 268                 RETURN(0);
 269         }
 270
 271         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 272                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 273                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id4),
 274                mbits - posted_md, mbits - 1);
 275
 276         RETURN(0);
 277 }
 278
 279 /**
 280  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 281  * serialises with completion callback)
 282  */
 283 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 284 {
 285         LASSERT(!in_interrupt());           /* might sleep */
 286
 287         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 288                 return;                         /* never started */
 289
 290         /* We used to poison the pages with 0xab here because we did not want to
 291          * send any meaningful data over the wire for evicted clients (bug 9297)
 292          * However, this is no longer safe now that we use the page cache on the
 293          * OSS (bug 20560) */
 294
 295         /* The unlink ensures the callback happens ASAP and is the last
 296          * one.  If it fails, it must be because completion just happened,
 297          * but we must still wait_event_idle_timeout() in this case, to give
 298          * us a chance to run server_bulk_callback()
 299          */
 300         __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true);
 301
 302         for (;;) {
 303                 /* Network access will complete in finite time but the HUGE
 304                  * timeout lets us CWARN for visibility of sluggish NALs */
 305                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 306
 307                 while (seconds > 0 &&
 308                        wait_event_idle_timeout(desc->bd_waitq,
 309                                                !ptlrpc_server_bulk_active(desc),
 310                                                cfs_time_seconds(1)) == 0)
 311                         seconds -= 1;
 312                 if (seconds > 0)
 313                         return;
 314
 315                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 316         }
 317 }
 318 #endif /* HAVE_SERVER_SUPPORT */
 319
 320 /**
 321  * Register bulk at the sender for later transfer.
 322  * Returns 0 on success or error code.
 323  */
 324 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 325 {
 326         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 327         struct lnet_processid peer;
 328         int rc = 0;
 329         int posted_md;
 330         int total_md;
 331         __u64 mbits;
 332         struct lnet_me *me;
 333         struct lnet_md md;
 334         ENTRY;
 335
 336         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 337                 RETURN(0);
 338
 339         /* NB no locking required until desc is on the network */
 340         LASSERT(desc->bd_nob > 0);
 341         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 342         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 343         LASSERT(desc->bd_req != NULL);
 344         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 345
 346         /* cleanup the state of the bulk for it will be reused */
 347         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 348                 desc->bd_nob_transferred = 0;
 349         else if (desc->bd_nob_transferred != 0)
 350                 /* If the network failed after an RPC was sent, this condition
 351                  * could happen.  Rather than assert (was here before), return
 352                  * an EIO error. */
 353                 RETURN(-EIO);
 354
 355         desc->bd_failure = 0;
 356
 357         peer = desc->bd_import->imp_connection->c_peer;
 358
 359         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 360         LASSERT(desc->bd_cbid.cbid_arg == desc);
 361
 362         total_md = desc->bd_md_count;
 363         /* rq_mbits is matchbits of the final bulk */
 364         mbits = req->rq_mbits - desc->bd_md_count + 1;
 365
 366         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 367                  "first mbits = x%llu, last mbits = x%llu\n",
 368                  mbits, req->rq_mbits);
 369         LASSERTF(!(desc->bd_registered &&
 370                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 371                  mbits != desc->bd_last_mbits,
 372                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 373                  desc->bd_registered, mbits, desc->bd_last_mbits);
 374
 375         desc->bd_registered = 1;
 376         desc->bd_last_mbits = mbits;
 377         desc->bd_refs = total_md;
 378         md.user_ptr = &desc->bd_cbid;
 379         md.handler = ptlrpc_handler;
 380         md.threshold = 1;                       /* PUT or GET */
 381
 382         for (posted_md = 0; posted_md < desc->bd_md_count;
 383              posted_md++, mbits++) {
 384                 md.options = PTLRPC_MD_OPTIONS |
 385                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 386                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 387                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 388
 389                 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
 390                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 391                         rc = -ENOMEM;
 392                 } else {
 393                         me = LNetMEAttach(desc->bd_portal, &peer, mbits, 0,
 394                                   LNET_UNLINK, LNET_INS_AFTER);
 395                         rc = PTR_ERR_OR_ZERO(me);
 396                 }
 397                 if (rc != 0) {
 398                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 399                                desc->bd_import->imp_obd->obd_name, mbits,
 400                                posted_md, rc);
 401                         break;
 402                 }
 403                 percpu_ref_get(&ptlrpc_pending);
 404
 405                 /* About to let the network at it... */
 406                 rc = LNetMDAttach(me, &md, LNET_UNLINK,
 407                                   &desc->bd_mds[posted_md]);
 408                 if (rc != 0) {
 409                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 410                                desc->bd_import->imp_obd->obd_name, mbits,
 411                                posted_md, rc);
 412                         break;
 413                 }
 414         }
 415
 416         if (rc != 0) {
 417                 LASSERT(rc == -ENOMEM);
 418                 spin_lock(&desc->bd_lock);
 419                 desc->bd_refs -= total_md - posted_md;
 420                 spin_unlock(&desc->bd_lock);
 421                 LASSERT(desc->bd_refs >= 0);
 422                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 423                 req->rq_status = -ENOMEM;
 424                 desc->bd_registered = 0;
 425                 RETURN(-ENOMEM);
 426         }
 427
 428         spin_lock(&desc->bd_lock);
 429         /* Holler if peer manages to touch buffers before he knows the mbits */
 430         if (desc->bd_refs != total_md)
 431                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 432                       desc->bd_import->imp_obd->obd_name, libcfs_idstr(&peer),
 433                       total_md - desc->bd_refs);
 434         spin_unlock(&desc->bd_lock);
 435
 436         CDEBUG(D_NET,
 437                "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
 438                desc->bd_refs,
 439                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 440                desc->bd_iov_count, desc->bd_nob,
 441                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 442
 443         RETURN(0);
 444 }
 445
 446 /**
 447  * Disconnect a bulk desc from the network. Idempotent. Not
 448  * thread-safe (i.e. only interlocks with completion callback).
 449  * Returns 1 on success or 0 if network unregistration failed for whatever
 450  * reason.
 451  */
 452 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 453 {
 454         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 455         ENTRY;
 456
 457         LASSERT(!in_interrupt());     /* might sleep */
 458
 459         if (desc)
 460                 desc->bd_registered = 0;
 461
 462         /* Let's setup deadline for reply unlink. */
 463         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 464             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 465                 req->rq_bulk_deadline = ktime_get_real_seconds() +
 466                                         PTLRPC_REQ_LONG_UNLINK;
 467
 468         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 469                 RETURN(1);                              /* never registered */
 470
 471         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 472
 473         /* the unlink ensures the callback happens ASAP and is the last
 474          * one.  If it fails, it must be because completion just happened,
 475          * but we must still wait_event_idle_timeout() in this case to give
 476          * us a chance to run client_bulk_callback()
 477          */
 478         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 479
 480         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 481                 RETURN(1);                              /* never registered */
 482
 483         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 484         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 485
 486         /* Do not wait for unlink to finish. */
 487         if (async)
 488                 RETURN(0);
 489
 490         for (;;) {
 491                 /* The wq argument is ignored by user-space wait_event macros */
 492                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 493                                         &req->rq_set->set_waitq :
 494                                         &req->rq_reply_waitq;
 495                 /*
 496                  * Network access will complete in finite time but the HUGE
 497                  * timeout lets us CWARN for visibility of sluggish NALs.
 498                  */
 499                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 500
 501                 while (seconds > 0 &&
 502                        wait_event_idle_timeout(*wq,
 503                                                !ptlrpc_client_bulk_active(req),
 504                                                cfs_time_seconds(1)) == 0)
 505                         seconds -= 1;
 506                 if (seconds > 0) {
 507                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 508                         RETURN(1);
 509                 }
 510
 511                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 512                           desc);
 513         }
 514         RETURN(0);
 515 }
 516
 517 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 518 {
 519         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 520         struct ptlrpc_service           *svc = svcpt->scp_service;
 521         timeout_t service_timeout;
 522
 523         service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
 524                                              req->rq_arrival_time.tv_sec, 1,
 525                                   (AT_OFF ? obd_timeout * 3 / 2 : at_max));
 526         if (!(flags & PTLRPC_REPLY_EARLY) &&
 527             (req->rq_type != PTL_RPC_MSG_ERR) &&
 528             (req->rq_reqmsg != NULL) &&
 529             !(lustre_msg_get_flags(req->rq_reqmsg) &
 530               (MSG_RESENT | MSG_REPLAY |
 531                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 532                 /* early replies, errors and recovery requests don't count
 533                  * toward our service time estimate
 534                  */
 535                 timeout_t oldse = at_measured(&svcpt->scp_at_estimate,
 536                                               service_timeout);
 537
 538                 if (oldse != 0) {
 539                         DEBUG_REQ(D_ADAPTTO, req,
 540                                   "svc %s changed estimate from %d to %d",
 541                                   svc->srv_name, oldse,
 542                                   at_get(&svcpt->scp_at_estimate));
 543                 }
 544         }
 545         /* Report actual service time for client latency calc */
 546         lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
 547         /* Report service time estimate for future client reqs, but report 0
 548          * (to be ignored by client) if it's an error reply during recovery.
 549          * b=15815
 550          */
 551         if (req->rq_type == PTL_RPC_MSG_ERR &&
 552             (req->rq_export == NULL ||
 553              req->rq_export->exp_obd->obd_recovering)) {
 554                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 555         } else {
 556                 timeout_t timeout;
 557
 558                 if (req->rq_export && req->rq_reqmsg != NULL &&
 559                     (flags & PTLRPC_REPLY_EARLY) &&
 560                     lustre_msg_get_flags(req->rq_reqmsg) &
 561                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 562                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 563
 564                         timeout = ktime_get_real_seconds() -
 565                                   req->rq_arrival_time.tv_sec +
 566                                   min_t(timeout_t, at_extra,
 567                                         exp_obd->obd_recovery_timeout / 4);
 568                 } else {
 569                         timeout = at_get(&svcpt->scp_at_estimate);
 570                 }
 571                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 572         }
 573
 574         if (req->rq_reqmsg &&
 575             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 576                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 577                        "req_flags=%#x magic=%x/%x len=%d\n",
 578                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 579                        lustre_msg_get_magic(req->rq_reqmsg),
 580                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 581         }
 582 }
 583
 584 /**
 585  * Send request reply from request \a req reply buffer.
 586  * \a flags defines reply types
 587  * Returns 0 on success or error code
 588  */
 589 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 590 {
 591         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 592         struct ptlrpc_connection  *conn;
 593         int                        rc;
 594
 595         /* We must already have a reply buffer (only ptlrpc_error() may be
 596          * called without one). The reply generated by sptlrpc layer (e.g.
 597          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 598          * have a request buffer which is either the actual (swabbed) incoming
 599          * request, or a saved copy if this is a req saved in
 600          * target_queue_final_reply().
 601          */
 602         LASSERT (req->rq_no_reply == 0);
 603         LASSERT (req->rq_reqbuf != NULL);
 604         LASSERT (rs != NULL);
 605         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 606         LASSERT (req->rq_repmsg != NULL);
 607         LASSERT (req->rq_repmsg == rs->rs_msg);
 608         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 609         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 610
 611         /* There may be no rq_export during failover */
 612
 613         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 614                      req->rq_export->exp_obd->obd_fail)) {
 615                 /* Failed obd's only send ENODEV */
 616                 req->rq_type = PTL_RPC_MSG_ERR;
 617                 req->rq_status = -ENODEV;
 618                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 619                        req->rq_export->exp_obd->obd_minor);
 620         }
 621
 622         if (req->rq_type != PTL_RPC_MSG_ERR)
 623                 req->rq_type = PTL_RPC_MSG_REPLY;
 624
 625         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 626         lustre_msg_set_status(req->rq_repmsg,
 627                               ptlrpc_status_hton(req->rq_status));
 628         lustre_msg_set_opc(req->rq_repmsg,
 629                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 630
 631         target_pack_pool_reply(req);
 632
 633         ptlrpc_at_set_reply(req, flags);
 634
 635         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 636                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 637         else
 638                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 639
 640         if (unlikely(conn == NULL)) {
 641                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 642                 return -ENOTCONN;
 643         }
 644         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 645
 646         rc = sptlrpc_svc_wrap_reply(req);
 647         if (unlikely(rc))
 648                 goto out;
 649
 650         req->rq_sent = ktime_get_real_seconds();
 651
 652         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 653                           (rs->rs_difficult && !rs->rs_no_ack) ?
 654                           LNET_ACK_REQ : LNET_NOACK_REQ,
 655                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 656                           ptlrpc_req2svc(req)->srv_rep_portal,
 657                           req->rq_rep_mbits ? req->rq_rep_mbits : req->rq_xid,
 658                           req->rq_reply_off, NULL);
 659 out:
 660         if (unlikely(rc != 0))
 661                 ptlrpc_req_drop_rs(req);
 662         ptlrpc_connection_put(conn);
 663         return rc;
 664 }
 665
 666 int ptlrpc_reply (struct ptlrpc_request *req)
 667 {
 668         if (req->rq_no_reply)
 669                 return 0;
 670         else
 671                 return (ptlrpc_send_reply(req, 0));
 672 }
 673
 674 /**
 675  * For request \a req send an error reply back. Create empty
 676  * reply buffers if necessary.
 677  */
 678 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 679 {
 680         int rc;
 681         ENTRY;
 682
 683         if (req->rq_no_reply)
 684                 RETURN(0);
 685
 686         if (!req->rq_repmsg) {
 687                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 688                 if (rc)
 689                         RETURN(rc);
 690         }
 691
 692         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 693             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 694             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 695                 req->rq_type = PTL_RPC_MSG_ERR;
 696
 697         rc = ptlrpc_send_reply(req, may_be_difficult);
 698         RETURN(rc);
 699 }
 700
 701 int ptlrpc_error(struct ptlrpc_request *req)
 702 {
 703         return ptlrpc_send_error(req, 0);
 704 }
 705
 706 /**
 707  * Send request \a request.
 708  * if \a noreply is set, don't expect any reply back and don't set up
 709  * reply buffers.
 710  * Returns 0 on success or error code.
 711  */
 712 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 713 {
 714         int rc;
 715         __u32 opc;
 716         int mpflag = 0;
 717         bool rep_mbits = false;
 718         struct lnet_handle_md bulk_cookie;
 719         struct lnet_processid peer;
 720         struct ptlrpc_connection *connection;
 721         struct lnet_me *reply_me = NULL;
 722         struct lnet_md reply_md;
 723         struct obd_import *imp = request->rq_import;
 724         struct obd_device *obd = imp->imp_obd;
 725         ENTRY;
 726
 727         LNetInvalidateMDHandle(&bulk_cookie);
 728
 729         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 730                 RETURN(0);
 731
 732         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) &&
 733                      lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT &&
 734                      strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) {
 735                 RETURN(0);
 736         }
 737
 738         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 739         LASSERT(request->rq_wait_ctx == 0);
 740
 741         /* If this is a re-transmit, we're required to have disengaged
 742          * cleanly from the previous attempt */
 743         LASSERT(!request->rq_receiving_reply);
 744         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 745                   (imp->imp_state == LUSTRE_IMP_FULL)));
 746
 747         if (unlikely(obd != NULL && obd->obd_fail)) {
 748                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 749                        obd->obd_name);
 750                 /* this prevents us from waiting in ptlrpc_queue_wait */
 751                 spin_lock(&request->rq_lock);
 752                 request->rq_err = 1;
 753                 spin_unlock(&request->rq_lock);
 754                 request->rq_status = -ENODEV;
 755                 RETURN(-ENODEV);
 756         }
 757
 758         connection = imp->imp_connection;
 759
 760         lustre_msg_set_handle(request->rq_reqmsg,
 761                               &imp->imp_remote_handle);
 762         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 763         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 764                                 imp->imp_conn_cnt);
 765         lustre_msghdr_set_flags(request->rq_reqmsg,
 766                                 imp->imp_msghdr_flags);
 767
 768         /* If it's the first time to resend the request for EINPROGRESS,
 769          * we need to allocate a new XID (see after_reply()), it's different
 770          * from the resend for reply timeout. */
 771         if (request->rq_nr_resend != 0 &&
 772             list_empty(&request->rq_unreplied_list)) {
 773                 __u64 min_xid = 0;
 774                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 775                  * reconstruction */
 776                 spin_lock(&imp->imp_lock);
 777                 ptlrpc_assign_next_xid_nolock(request);
 778                 min_xid = ptlrpc_known_replied_xid(imp);
 779                 spin_unlock(&imp->imp_lock);
 780
 781                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 782                 DEBUG_REQ(D_RPCTRACE, request,
 783                           "Allocating new XID for resend on EINPROGRESS");
 784         }
 785
 786         opc = lustre_msg_get_opc(request->rq_reqmsg);
 787         if (opc != OST_CONNECT && opc != MDS_CONNECT &&
 788             opc != MGS_CONNECT && OCD_HAS_FLAG(&imp->imp_connect_data, FLAGS2))
 789                 rep_mbits = imp->imp_connect_data.ocd_connect_flags2 &
 790                         OBD_CONNECT2_REP_MBITS;
 791
 792         if ((request->rq_bulk != NULL) || rep_mbits) {
 793                 ptlrpc_set_mbits(request);
 794                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 795         }
 796
 797         if (list_empty(&request->rq_unreplied_list) ||
 798             request->rq_xid <= imp->imp_known_replied_xid) {
 799                 DEBUG_REQ(D_ERROR, request,
 800                           "xid=%llu, replied=%llu, list_empty=%d",
 801                           request->rq_xid, imp->imp_known_replied_xid,
 802                           list_empty(&request->rq_unreplied_list));
 803                 LBUG();
 804         }
 805
 806         /** For enabled AT all request should have AT_SUPPORT in the
 807          * FULL import state when OBD_CONNECT_AT is set */
 808         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 809                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 810                 !(imp->imp_connect_data.ocd_connect_flags &
 811                   OBD_CONNECT_AT));
 812
 813         if (request->rq_resend) {
 814                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 815                 if (request->rq_resend_cb != NULL)
 816                         request->rq_resend_cb(request, &request->rq_async_args);
 817         }
 818         if (request->rq_memalloc)
 819                 mpflag = memalloc_noreclaim_save();
 820
 821         rc = sptlrpc_cli_wrap_request(request);
 822         if (rc)
 823                 GOTO(out, rc);
 824
 825         /* bulk register should be done after wrap_request() */
 826         if (request->rq_bulk != NULL) {
 827                 rc = ptlrpc_register_bulk (request);
 828                 if (rc != 0)
 829                         GOTO(cleanup_bulk, rc);
 830                 /*
 831                  * All the mds in the request will have the same cpt
 832                  * encoded in the cookie. So we can just get the first
 833                  * one.
 834                  */
 835                 bulk_cookie = request->rq_bulk->bd_mds[0];
 836         }
 837
 838         if (!noreply) {
 839                 LASSERT (request->rq_replen != 0);
 840                 if (request->rq_repbuf == NULL) {
 841                         LASSERT(request->rq_repdata == NULL);
 842                         LASSERT(request->rq_repmsg == NULL);
 843                         rc = sptlrpc_cli_alloc_repbuf(request,
 844                                                       request->rq_replen);
 845                         if (rc) {
 846                                 /* this prevents us from looping in
 847                                  * ptlrpc_queue_wait */
 848                                 spin_lock(&request->rq_lock);
 849                                 request->rq_err = 1;
 850                                 spin_unlock(&request->rq_lock);
 851                                 request->rq_status = rc;
 852                                 GOTO(cleanup_bulk, rc);
 853                         }
 854                 } else {
 855                         request->rq_repdata = NULL;
 856                         request->rq_repmsg = NULL;
 857                 }
 858
 859                 peer = connection->c_peer;
 860                 if (request->rq_bulk &&
 861                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
 862                         reply_me = ERR_PTR(-ENOMEM);
 863                 } else {
 864                         reply_me = LNetMEAttach(request->rq_reply_portal,
 865                                                 &peer,
 866                                                 rep_mbits ? request->rq_mbits :
 867                                                 request->rq_xid,
 868                                                 0, LNET_UNLINK, LNET_INS_AFTER);
 869                 }
 870
 871                 if (IS_ERR(reply_me)) {
 872                         rc = PTR_ERR(reply_me);
 873                         CERROR("LNetMEAttach failed: %d\n", rc);
 874                         LASSERT(rc == -ENOMEM);
 875                         GOTO(cleanup_bulk, rc = -ENOMEM);
 876                 }
 877         }
 878
 879         spin_lock(&request->rq_lock);
 880         /* We are responsible for unlinking the reply buffer */
 881         request->rq_reply_unlinked = noreply;
 882         request->rq_receiving_reply = !noreply;
 883         /* Clear any flags that may be present from previous sends. */
 884         request->rq_req_unlinked = 0;
 885         request->rq_replied = 0;
 886         request->rq_err = 0;
 887         request->rq_timedout = 0;
 888         request->rq_net_err = 0;
 889         request->rq_resend = 0;
 890         request->rq_restart = 0;
 891         request->rq_reply_truncated = 0;
 892         spin_unlock(&request->rq_lock);
 893
 894         if (!noreply) {
 895                 reply_md.start     = request->rq_repbuf;
 896                 reply_md.length    = request->rq_repbuf_len;
 897                 /* Allow multiple early replies */
 898                 reply_md.threshold = LNET_MD_THRESH_INF;
 899                 /* Manage remote for early replies */
 900                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 901                         LNET_MD_MANAGE_REMOTE |
 902                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 903                 reply_md.user_ptr  = &request->rq_reply_cbid;
 904                 reply_md.handler = ptlrpc_handler;
 905
 906                 /* We must see the unlink callback to set rq_reply_unlinked,
 907                  * so we can't auto-unlink */
 908                 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
 909                                   &request->rq_reply_md_h);
 910                 if (rc != 0) {
 911                         CERROR("LNetMDAttach failed: %d\n", rc);
 912                         LASSERT(rc == -ENOMEM);
 913                         spin_lock(&request->rq_lock);
 914                         /* ...but the MD attach didn't succeed... */
 915                         request->rq_receiving_reply = 0;
 916                         spin_unlock(&request->rq_lock);
 917                         GOTO(cleanup_bulk, rc = -ENOMEM);
 918                 }
 919                 percpu_ref_get(&ptlrpc_pending);
 920
 921                 CDEBUG(D_NET,
 922                        "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 923                        request->rq_repbuf_len, request->rq_xid,
 924                        request->rq_reply_portal);
 925         }
 926
 927         /* add references on request for request_out_callback */
 928         ptlrpc_request_addref(request);
 929         if (obd != NULL && obd->obd_svc_stats != NULL)
 930                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 931                                     atomic_read(&imp->imp_inflight));
 932
 933         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 934
 935         request->rq_sent_ns = ktime_get_real();
 936         request->rq_sent = ktime_get_real_seconds();
 937         /* We give the server rq_timeout secs to process the req, and
 938          * add the network latency for our local timeout.
 939          */
 940         request->rq_deadline = request->rq_sent + request->rq_timeout +
 941                 ptlrpc_at_get_net_latency(request);
 942
 943         DEBUG_REQ(D_INFO, request, "send flags=%x",
 944                   lustre_msg_get_flags(request->rq_reqmsg));
 945         rc = ptl_send_buf(&request->rq_req_md_h,
 946                           request->rq_reqbuf, request->rq_reqdata_len,
 947                           LNET_NOACK_REQ, &request->rq_req_cbid,
 948                           LNET_NID_ANY,
 949                           lnet_pid_to_pid4(&connection->c_peer),
 950                           request->rq_request_portal,
 951                           request->rq_xid, 0, &bulk_cookie);
 952         if (likely(rc == 0))
 953                 GOTO(out, rc);
 954
 955         request->rq_req_unlinked = 1;
 956         ptlrpc_req_finished(request);
 957         if (noreply)
 958                 GOTO(out, rc);
 959
 960         LNetMDUnlink(request->rq_reply_md_h);
 961
 962         /* UNLINKED callback called synchronously */
 963         LASSERT(!request->rq_receiving_reply);
 964
 965  cleanup_bulk:
 966         /* We do sync unlink here as there was no real transfer here so
 967          * the chance to have long unlink to sluggish net is smaller here. */
 968         ptlrpc_unregister_bulk(request, 0);
 969  out:
 970         if (rc == -ENOMEM) {
 971                 /* set rq_sent so that this request is treated
 972                  * as a delayed send in the upper layers */
 973                 request->rq_sent = ktime_get_real_seconds();
 974         }
 975
 976         if (request->rq_memalloc)
 977                 memalloc_noreclaim_restore(mpflag);
 978
 979         return rc;
 980 }
 981 EXPORT_SYMBOL(ptl_send_rpc);
 982
 983 /**
 984  * Register request buffer descriptor for request receiving.
 985  */
 986 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 987 {
 988         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 989         static struct lnet_processid match_id = {
 990                 .nid = LNET_ANY_NID,
 991                 .pid = LNET_PID_ANY
 992         };
 993         int rc;
 994         struct lnet_md md;
 995         struct lnet_me *me;
 996
 997         CDEBUG(D_NET, "%s: registering portal %d\n", service->srv_name,
 998                service->srv_req_portal);
 999
1000         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
1001                 return -ENOMEM;
1002
1003         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
1004          * which means buffer can only be attached on local CPT, and LND
1005          * threads can find it by grabbing a local lock */
1006         me = LNetMEAttach(service->srv_req_portal,
1007                           &match_id, 0, ~0, LNET_UNLINK,
1008                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
1009                           LNET_INS_LOCAL : LNET_INS_AFTER);
1010         if (IS_ERR(me)) {
1011                 CERROR("%s: LNetMEAttach failed: rc = %ld\n",
1012                        service->srv_name, PTR_ERR(me));
1013                 return PTR_ERR(me);
1014         }
1015
1016         LASSERT(rqbd->rqbd_refcount == 0);
1017         rqbd->rqbd_refcount = 1;
1018
1019         md.start     = rqbd->rqbd_buffer;
1020         md.length    = service->srv_buf_size;
1021         md.max_size  = service->srv_max_req_size;
1022         md.threshold = LNET_MD_THRESH_INF;
1023         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
1024         md.user_ptr  = &rqbd->rqbd_cbid;
1025         md.handler   = ptlrpc_handler;
1026
1027         rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
1028         if (rc == 0) {
1029                 percpu_ref_get(&ptlrpc_pending);
1030                 return 0;
1031         }
1032
1033         CERROR("%s: LNetMDAttach failed: rc = %d\n", service->srv_name, rc);
1034         LASSERT(rc == -ENOMEM);
1035         rqbd->rqbd_refcount = 0;
1036
1037         return rc;
1038 }