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                         struct lnet_nid *self, struct lnet_processid *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
  56         ENTRY;
  57
  58         LASSERT(portal != 0);
  59         CDEBUG(D_INFO, "peer_id %s\n", libcfs_idstr(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 (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, CFS_FAIL_ONCE) &&
  74             ack == LNET_ACK_REQ) {
  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                 /* Will get UNLINK event when unlink below, which will complete
  96                  * like any other failed send, fall through and return success
  97                  */
  98                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  99                        libcfs_idstr(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 int nfrags, unsigned int max_brw,
 129                          unsigned int type, unsigned int portal,
 130                          const struct ptlrpc_bulk_frag_ops *ops)
 131 {
 132         struct obd_export *exp = req->rq_export;
 133         struct ptlrpc_bulk_desc *desc;
 134
 135         ENTRY;
 136         LASSERT(ptlrpc_is_bulk_op_active(type));
 137
 138         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 139         if (desc == NULL)
 140                 RETURN(NULL);
 141
 142         desc->bd_export = class_export_get(exp);
 143         desc->bd_req = req;
 144
 145         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 146         desc->bd_cbid.cbid_arg = desc;
 147
 148         /* NB we don't assign rq_bulk here; server-side requests are
 149          * re-used, and the handler frees the bulk desc explicitly.
 150          */
 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         struct lnet_nid          self_nid;
 164         struct lnet_processid    peer_id;
 165         int                      rc = 0;
 166         __u64                    mbits;
 167         int                      posted_md;
 168         int                      total_md;
 169         struct lnet_md           md;
 170
 171         ENTRY;
 172
 173         if (CFS_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         /* Multi-Rail: get the preferred self and peer NIDs from the
 183          * request, so they are based on the route taken by the message.
 184          */
 185         self_nid = desc->bd_req->rq_self;
 186         peer_id = desc->bd_req->rq_source;
 187
 188         /* NB total length may be 0 for a read past EOF, so we send 0
 189          * length bulks, since the client expects bulk events.
 190          *
 191          * The client may not need all of the bulk mbits for the RPC. The RPC
 192          * used the mbits of the highest bulk mbits needed, and the server masks
 193          * off high bits to get bulk count for this RPC. LU-1431
 194          */
 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                 /* Note. source and sink buf frags are page-aligned. Else send
 208                  * client bulk sizes over and split server buffer accordingly
 209                  */
 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                         CFS_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,
 237                                      mbits, 0, false);
 238
 239                 posted_md++;
 240                 if (rc != 0) {
 241                         CERROR("%s: failed bulk transfer with %s:%u x%llu: rc = %d\n",
 242                                exp->exp_obd->obd_name,
 243                                libcfs_idstr(&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                  */
 254                 spin_lock(&desc->bd_lock);
 255                 desc->bd_refs -= total_md - posted_md;
 256                 spin_unlock(&desc->bd_lock);
 257                 LASSERT(desc->bd_refs >= 0);
 258
 259                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 260                 RETURN(0);
 261         }
 262
 263         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d id %s mbits %#llx-%#llx\n",
 264                desc->bd_iov_count,
 265                desc->bd_nob, desc->bd_portal, libcfs_idstr(&peer_id),
 266                mbits - posted_md, mbits - 1);
 267
 268         RETURN(0);
 269 }
 270
 271 /**
 272  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 273  * serialises with completion callback)
 274  */
 275 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 276 {
 277         LASSERT(!in_interrupt());               /* might sleep */
 278
 279         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 280                 return;                         /* never started */
 281
 282         /* We used to poison the pages with 0xab here because we did not want to
 283          * send any meaningful data over the wire for evicted clients (bug 9297)
 284          * However, this is no longer safe now that we use the page cache on the
 285          * OSS (bug 20560) */
 286
 287         /* The unlink ensures the callback happens ASAP and is the last
 288          * one.  If it fails, it must be because completion just happened,
 289          * but we must still wait_event_idle_timeout() in this case, to give
 290          * us a chance to run server_bulk_callback()
 291          */
 292         __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true);
 293
 294         for (;;) {
 295                 /* Network access will complete in finite time but the HUGE
 296                  * timeout lets us CWARN for visibility of sluggish NALs
 297                  */
 298                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 299
 300                 while (seconds > 0 &&
 301                        wait_event_idle_timeout(desc->bd_waitq,
 302                                                !ptlrpc_server_bulk_active(desc),
 303                                                cfs_time_seconds(1)) == 0)
 304                         seconds -= 1;
 305                 if (seconds > 0)
 306                         return;
 307
 308                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 309         }
 310 }
 311 #endif /* HAVE_SERVER_SUPPORT */
 312
 313 /**
 314  * Register bulk at the sender for later transfer.
 315  * Returns 0 on success or error code.
 316  */
 317 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 318 {
 319         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 320         struct lnet_processid peer;
 321         int rc = 0;
 322         int posted_md;
 323         int total_md;
 324         __u64 mbits;
 325         struct lnet_me *me;
 326         struct lnet_md md;
 327
 328         ENTRY;
 329
 330         if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 331                 RETURN(0);
 332
 333         /* NB no locking required until desc is on the network */
 334         LASSERT(desc->bd_nob > 0);
 335         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 336         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 337         LASSERT(desc->bd_req != NULL);
 338         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 339
 340         /* cleanup the state of the bulk for it will be reused */
 341         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 342                 desc->bd_nob_transferred = 0;
 343         else if (desc->bd_nob_transferred != 0)
 344                 /* If network failed after RPC was sent, this condition could
 345                  * happen. Rather than assert (was here before), return EIO err
 346                  */
 347                 RETURN(-EIO);
 348
 349         desc->bd_failure = 0;
 350
 351         peer = desc->bd_import->imp_connection->c_peer;
 352
 353         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 354         LASSERT(desc->bd_cbid.cbid_arg == desc);
 355
 356         total_md = desc->bd_md_count;
 357         /* rq_mbits is matchbits of the final bulk */
 358         mbits = req->rq_mbits - desc->bd_md_count + 1;
 359
 360         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 361                  "first mbits = x%llu, last mbits = x%llu\n",
 362                  mbits, req->rq_mbits);
 363         LASSERTF(!(desc->bd_registered &&
 364                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 365                  mbits != desc->bd_last_mbits,
 366                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 367                  desc->bd_registered, mbits, desc->bd_last_mbits);
 368
 369         desc->bd_registered = 1;
 370         desc->bd_last_mbits = mbits;
 371         desc->bd_refs = total_md;
 372         md.user_ptr = &desc->bd_cbid;
 373         md.handler = ptlrpc_handler;
 374         md.threshold = 1;                       /* PUT or GET */
 375
 376         for (posted_md = 0; posted_md < desc->bd_md_count;
 377              posted_md++, mbits++) {
 378                 md.options = PTLRPC_MD_OPTIONS |
 379                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 380                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 381                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 382
 383                 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
 384                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 385                         rc = -ENOMEM;
 386                 } else {
 387                         me = LNetMEAttach(desc->bd_portal, &peer, mbits, 0,
 388                                   LNET_UNLINK, LNET_INS_AFTER);
 389                         rc = PTR_ERR_OR_ZERO(me);
 390                 }
 391                 if (rc != 0) {
 392                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 393                                desc->bd_import->imp_obd->obd_name, mbits,
 394                                posted_md, rc);
 395                         break;
 396                 }
 397                 percpu_ref_get(&ptlrpc_pending);
 398
 399                 /* About to let the network at it... */
 400                 rc = LNetMDAttach(me, &md, LNET_UNLINK,
 401                                   &desc->bd_mds[posted_md]);
 402                 if (rc != 0) {
 403                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 404                                desc->bd_import->imp_obd->obd_name, mbits,
 405                                posted_md, rc);
 406                         break;
 407                 }
 408         }
 409
 410         if (rc != 0) {
 411                 LASSERT(rc == -ENOMEM);
 412                 spin_lock(&desc->bd_lock);
 413                 desc->bd_refs -= total_md - posted_md;
 414                 spin_unlock(&desc->bd_lock);
 415                 LASSERT(desc->bd_refs >= 0);
 416                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 417                 req->rq_status = -ENOMEM;
 418                 desc->bd_registered = 0;
 419                 RETURN(-ENOMEM);
 420         }
 421
 422         spin_lock(&desc->bd_lock);
 423         /* Holler if peer manages to touch buffers before he knows the mbits */
 424         if (desc->bd_refs != total_md)
 425                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 426                       desc->bd_import->imp_obd->obd_name, libcfs_idstr(&peer),
 427                       total_md - desc->bd_refs);
 428         spin_unlock(&desc->bd_lock);
 429
 430         CDEBUG(D_NET,
 431                "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
 432                desc->bd_refs,
 433                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 434                desc->bd_iov_count, desc->bd_nob,
 435                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 436
 437         RETURN(0);
 438 }
 439
 440 /**
 441  * Disconnect a bulk desc from the network. Idempotent. Not
 442  * thread-safe (i.e. only interlocks with completion callback).
 443  * Returns 1 on success or 0 if network unregistration failed for whatever
 444  * reason.
 445  */
 446 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 447 {
 448         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 449
 450         ENTRY;
 451
 452         LASSERT(!in_interrupt());     /* might sleep */
 453
 454         if (desc)
 455                 desc->bd_registered = 0;
 456
 457         /* Let's setup deadline for reply unlink. */
 458         if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 459             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 460                 req->rq_bulk_deadline = ktime_get_real_seconds() +
 461                                         PTLRPC_REQ_LONG_UNLINK;
 462
 463         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 464                 RETURN(1);                              /* never registered */
 465
 466         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 467
 468         /* the unlink ensures the callback happens ASAP and is the last
 469          * one.  If it fails, it must be because completion just happened,
 470          * but we must still wait_event_idle_timeout() in this case to give
 471          * us a chance to run client_bulk_callback()
 472          */
 473         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 474
 475         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 476                 RETURN(1);                              /* never registered */
 477
 478         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 479         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 480
 481         /* Do not wait for unlink to finish. */
 482         if (async)
 483                 RETURN(0);
 484
 485         for (;;) {
 486                 /* The wq argument is ignored by user-space wait_event macros */
 487                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 488                                         &req->rq_set->set_waitq :
 489                                         &req->rq_reply_waitq;
 490                 /*
 491                  * Network access will complete in finite time but the HUGE
 492                  * timeout lets us CWARN for visibility of sluggish NALs.
 493                  */
 494                 int seconds = PTLRPC_REQ_LONG_UNLINK;
 495
 496                 while (seconds > 0 &&
 497                        wait_event_idle_timeout(*wq,
 498                                                !ptlrpc_client_bulk_active(req),
 499                                                cfs_time_seconds(1)) == 0)
 500                         seconds -= 1;
 501                 if (seconds > 0) {
 502                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 503                         RETURN(1);
 504                 }
 505
 506                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 507                           desc);
 508         }
 509         RETURN(0);
 510 }
 511
 512 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 513 {
 514         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 515         struct ptlrpc_service           *svc = svcpt->scp_service;
 516         timeout_t service_timeout;
 517         struct obd_device *obd = NULL;
 518
 519         if (req->rq_export)
 520                 obd = req->rq_export->exp_obd;
 521
 522         service_timeout = obd_at_off(obd) ?
 523                           obd_timeout * 3 / 2 : obd_get_at_max(obd);
 524         service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
 525                                   req->rq_arrival_time.tv_sec, 1,
 526                                   service_timeout);
 527         if (!(flags & PTLRPC_REPLY_EARLY) &&
 528             (req->rq_type != PTL_RPC_MSG_ERR) &&
 529             (req->rq_reqmsg != NULL) &&
 530             !(lustre_msg_get_flags(req->rq_reqmsg) &
 531               (MSG_RESENT | MSG_REPLAY |
 532                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 533                 /* early replies, errors and recovery requests don't count
 534                  * toward our service time estimate
 535                  */
 536                 timeout_t oldse = obd_at_measure(obd, &svcpt->scp_at_estimate,
 537                                                  service_timeout);
 538
 539                 if (oldse != 0) {
 540                         DEBUG_REQ(D_ADAPTTO, req,
 541                                   "svc %s changed estimate from %d to %d",
 542                                   svc->srv_name, oldse,
 543                                   obd_at_get(obd, &svcpt->scp_at_estimate));
 544                 }
 545         }
 546         /* Report actual service time for client latency calc */
 547         lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
 548         /* Report service time estimate for future client reqs, but report 0
 549          * (to be ignored by client) if it's an error reply during recovery.
 550          * b=15815
 551          */
 552         if (req->rq_type == PTL_RPC_MSG_ERR &&
 553             (req->rq_export == NULL || 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                         timeout = ktime_get_real_seconds() -
 563                                   req->rq_arrival_time.tv_sec +
 564                                   min_t(timeout_t, at_extra,
 565                                         obd->obd_recovery_timeout / 4);
 566                 } else {
 567                         timeout = obd_at_get(obd, &svcpt->scp_at_estimate);
 568                 }
 569                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 570         }
 571
 572         if (req->rq_reqmsg &&
 573             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 574                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x req_flags=%#x magic=%x/%x len=%d\n",
 575                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 576                        lustre_msg_get_magic(req->rq_reqmsg),
 577                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 578         }
 579 }
 580
 581 /**
 582  * Send request reply from request \a req reply buffer.
 583  * \a flags defines reply types
 584  * Returns 0 on success or error code
 585  */
 586 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 587 {
 588         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 589         struct ptlrpc_connection  *conn;
 590         int                        rc;
 591
 592         /* We must already have a reply buffer (only ptlrpc_error() may be
 593          * called without one). The reply generated by sptlrpc layer (e.g.
 594          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 595          * have a request buffer which is either the actual (swabbed) incoming
 596          * request, or a saved copy if this is a req saved in
 597          * target_queue_final_reply().
 598          */
 599         LASSERT(req->rq_no_reply == 0);
 600         LASSERT(req->rq_reqbuf != NULL);
 601         LASSERT(rs != NULL);
 602         LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 603         LASSERT(req->rq_repmsg != NULL);
 604         LASSERT(req->rq_repmsg == rs->rs_msg);
 605         LASSERT(rs->rs_cb_id.cbid_fn == reply_out_callback);
 606         LASSERT(rs->rs_cb_id.cbid_arg == rs);
 607
 608         /* There may be no rq_export during failover */
 609
 610         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 611                      req->rq_export->exp_obd->obd_fail)) {
 612                 /* Failed obd's only send ENODEV */
 613                 req->rq_type = PTL_RPC_MSG_ERR;
 614                 req->rq_status = -ENODEV;
 615                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 616                        req->rq_export->exp_obd->obd_minor);
 617         }
 618
 619         if (req->rq_type != PTL_RPC_MSG_ERR)
 620                 req->rq_type = PTL_RPC_MSG_REPLY;
 621
 622         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 623         lustre_msg_set_status(req->rq_repmsg,
 624                               ptlrpc_status_hton(req->rq_status));
 625         lustre_msg_set_opc(req->rq_repmsg,
 626                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 627
 628         target_pack_pool_reply(req);
 629
 630         ptlrpc_at_set_reply(req, flags);
 631
 632         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 633                 conn = ptlrpc_connection_get(&req->rq_peer, &req->rq_self,
 634                                              NULL);
 635         else
 636                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 637
 638         if (unlikely(conn == NULL)) {
 639                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 640                 return -ENOTCONN;
 641         }
 642         ptlrpc_rs_addref(rs);  /* +1 ref for the network */
 643
 644         rc = sptlrpc_svc_wrap_reply(req);
 645         if (unlikely(rc))
 646                 goto out;
 647
 648         req->rq_sent = ktime_get_real_seconds();
 649
 650         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 651                           (rs->rs_difficult && !rs->rs_no_ack) ?
 652                           LNET_ACK_REQ : LNET_NOACK_REQ,
 653                           &rs->rs_cb_id, &req->rq_self,
 654                           &req->rq_source,
 655                           ptlrpc_req2svc(req)->srv_rep_portal,
 656                           req->rq_rep_mbits ? req->rq_rep_mbits : req->rq_xid,
 657                           req->rq_reply_off, NULL);
 658 out:
 659         if (unlikely(rc != 0))
 660                 ptlrpc_req_drop_rs(req);
 661         ptlrpc_connection_put(conn);
 662         return rc;
 663 }
 664
 665 int ptlrpc_reply(struct ptlrpc_request *req)
 666 {
 667         if (req->rq_no_reply)
 668                 return 0;
 669         else
 670                 return (ptlrpc_send_reply(req, 0));
 671 }
 672
 673 /**
 674  * For request \a req send an error reply back. Create empty
 675  * reply buffers if necessary.
 676  */
 677 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 678 {
 679         int rc;
 680
 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_status != -EROFS)
 696                 req->rq_type = PTL_RPC_MSG_ERR;
 697
 698         rc = ptlrpc_send_reply(req, may_be_difficult);
 699         RETURN(rc);
 700 }
 701
 702 int ptlrpc_error(struct ptlrpc_request *req)
 703 {
 704         return ptlrpc_send_error(req, 0);
 705 }
 706
 707 /**
 708  * Send request \a request.
 709  * if \a noreply is set, don't expect any reply back and don't set up
 710  * reply buffers.
 711  * Returns 0 on success or error code.
 712  */
 713 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 714 {
 715         int rc;
 716         __u32 opc;
 717         int mpflag = 0;
 718         bool rep_mbits = false;
 719         struct lnet_handle_md bulk_cookie;
 720         struct lnet_processid peer;
 721         struct ptlrpc_connection *connection;
 722         struct lnet_me *reply_me = NULL;
 723         struct lnet_md reply_md;
 724         struct obd_import *imp = request->rq_import;
 725         struct obd_device *obd = imp->imp_obd;
 726
 727         ENTRY;
 728
 729         LNetInvalidateMDHandle(&bulk_cookie);
 730
 731         if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 732                 RETURN(0);
 733
 734         if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) &&
 735                      lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT &&
 736                      strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) {
 737                 RETURN(0);
 738         }
 739
 740         if (unlikely(CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DROP_MGS, cfs_fail_val) &&
 741                      lustre_msg_get_opc(request->rq_reqmsg) == MGS_CONNECT)) {
 742                 DEBUG_REQ(D_INFO, request, "Simulate MGS connect failure");
 743                 RETURN(0);
 744         }
 745
 746
 747         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 748         LASSERT(request->rq_wait_ctx == 0);
 749
 750         /* If this is re-transmit, disengaged cleanly from previous attempt */
 751         LASSERT(!request->rq_receiving_reply);
 752         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 753                   (imp->imp_state == LUSTRE_IMP_FULL)));
 754
 755         if (unlikely(obd != NULL && obd->obd_fail)) {
 756                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 757                        obd->obd_name);
 758                 /* this prevents us from waiting in ptlrpc_queue_wait */
 759                 spin_lock(&request->rq_lock);
 760                 request->rq_err = 1;
 761                 spin_unlock(&request->rq_lock);
 762                 request->rq_status = -ENODEV;
 763                 RETURN(-ENODEV);
 764         }
 765
 766         connection = imp->imp_connection;
 767
 768         lustre_msg_set_handle(request->rq_reqmsg,
 769                               &imp->imp_remote_handle);
 770         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 771         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 772                                 imp->imp_conn_cnt);
 773         lustre_msghdr_set_flags(request->rq_reqmsg,
 774                                 imp->imp_msghdr_flags);
 775
 776         /* First time to resend request for EINPROGRESS, need to allocate new
 777          * XID(see after_reply()), it's different from resend for reply timeout
 778          */
 779         if (request->rq_nr_resend != 0 &&
 780             list_empty(&request->rq_unreplied_list)) {
 781                 __u64 min_xid = 0;
 782                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 783                  * reconstruction
 784                  */
 785                 spin_lock(&imp->imp_lock);
 786                 ptlrpc_assign_next_xid_nolock(request);
 787                 min_xid = ptlrpc_known_replied_xid(imp);
 788                 spin_unlock(&imp->imp_lock);
 789
 790                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 791                 DEBUG_REQ(D_RPCTRACE, request,
 792                           "Allocating new XID for resend on EINPROGRESS");
 793         }
 794
 795         opc = lustre_msg_get_opc(request->rq_reqmsg);
 796         if (opc != OST_CONNECT && opc != MDS_CONNECT &&
 797             opc != MGS_CONNECT && OCD_HAS_FLAG(&imp->imp_connect_data, FLAGS2))
 798                 rep_mbits = imp->imp_connect_data.ocd_connect_flags2 &
 799                         OBD_CONNECT2_REP_MBITS;
 800
 801         if ((request->rq_bulk != NULL) || rep_mbits) {
 802                 ptlrpc_set_mbits(request);
 803                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 804         }
 805
 806         if (list_empty(&request->rq_unreplied_list) ||
 807             request->rq_xid <= imp->imp_known_replied_xid) {
 808                 DEBUG_REQ(D_ERROR, request,
 809                           "xid=%llu, replied=%llu, list_empty=%d",
 810                           request->rq_xid, imp->imp_known_replied_xid,
 811                           list_empty(&request->rq_unreplied_list));
 812                 LBUG();
 813         }
 814
 815         /**
 816          * For enabled AT all request should have AT_SUPPORT in the
 817          * FULL import state when OBD_CONNECT_AT is set.
 818          * This check has a race with ptlrpc_connect_import_locked()
 819          * with low chance, don't panic, only report.
 820          */
 821         if (!(obd_at_off(obd) || imp->imp_state != LUSTRE_IMP_FULL ||
 822             (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 823             !(imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_AT))) {
 824                 DEBUG_REQ(D_HA, request, "Wrong state of import detected, AT=%d, imp=%d, msghdr=%d, conn=%d\n",
 825                           obd_at_off(obd), imp->imp_state != LUSTRE_IMP_FULL,
 826                           (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT),
 827                           !(imp->imp_connect_data.ocd_connect_flags &
 828                             OBD_CONNECT_AT));
 829         }
 830         if (request->rq_resend) {
 831                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 832                 if (request->rq_resend_cb != NULL)
 833                         request->rq_resend_cb(request, &request->rq_async_args);
 834         }
 835         if (request->rq_memalloc)
 836                 mpflag = memalloc_noreclaim_save();
 837
 838         rc = sptlrpc_cli_wrap_request(request);
 839         if (rc)
 840                 GOTO(out, rc);
 841
 842         /* bulk register should be done after wrap_request() */
 843         if (request->rq_bulk != NULL) {
 844                 rc = ptlrpc_register_bulk(request);
 845                 if (rc != 0)
 846                         GOTO(cleanup_bulk, rc);
 847                 /*
 848                  * All the mds in the request will have the same cpt
 849                  * encoded in the cookie. So we can just get the first
 850                  * one.
 851                  */
 852                 bulk_cookie = request->rq_bulk->bd_mds[0];
 853         }
 854
 855         if (!noreply) {
 856                 LASSERT(request->rq_replen != 0);
 857                 if (request->rq_repbuf == NULL) {
 858                         LASSERT(request->rq_repdata == NULL);
 859                         LASSERT(request->rq_repmsg == NULL);
 860                         rc = sptlrpc_cli_alloc_repbuf(request,
 861                                                       request->rq_replen);
 862                         if (rc) {
 863                                 /* prevent from looping in ptlrpc_queue_wait */
 864                                 spin_lock(&request->rq_lock);
 865                                 request->rq_err = 1;
 866                                 spin_unlock(&request->rq_lock);
 867                                 request->rq_status = rc;
 868                                 GOTO(cleanup_bulk, rc);
 869                         }
 870                 } else {
 871                         request->rq_repdata = NULL;
 872                         request->rq_repmsg = NULL;
 873                 }
 874
 875                 peer = connection->c_peer;
 876                 if (request->rq_bulk &&
 877                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
 878                         reply_me = ERR_PTR(-ENOMEM);
 879                 } else {
 880                         reply_me = LNetMEAttach(request->rq_reply_portal,
 881                                                 &peer,
 882                                                 rep_mbits ? request->rq_mbits :
 883                                                 request->rq_xid,
 884                                                 0, LNET_UNLINK, LNET_INS_AFTER);
 885                 }
 886
 887                 if (IS_ERR(reply_me)) {
 888                         rc = PTR_ERR(reply_me);
 889                         CERROR("LNetMEAttach failed: %d\n", rc);
 890                         LASSERT(rc == -ENOMEM);
 891                         GOTO(cleanup_bulk, rc = -ENOMEM);
 892                 }
 893         }
 894
 895         spin_lock(&request->rq_lock);
 896         /* We are responsible for unlinking the reply buffer */
 897         request->rq_reply_unlinked = noreply;
 898         request->rq_receiving_reply = !noreply;
 899         /* Clear any flags that may be present from previous sends. */
 900         request->rq_req_unlinked = 0;
 901         request->rq_replied = 0;
 902         request->rq_err = 0;
 903         request->rq_timedout = 0;
 904         request->rq_net_err = 0;
 905         request->rq_resend = 0;
 906         request->rq_restart = 0;
 907         request->rq_reply_truncated = 0;
 908         spin_unlock(&request->rq_lock);
 909
 910         if (!noreply) {
 911                 reply_md.start     = request->rq_repbuf;
 912                 reply_md.length    = request->rq_repbuf_len;
 913                 /* Allow multiple early replies */
 914                 reply_md.threshold = LNET_MD_THRESH_INF;
 915                 /* Manage remote for early replies */
 916                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 917                         LNET_MD_MANAGE_REMOTE |
 918                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 919                 reply_md.user_ptr  = &request->rq_reply_cbid;
 920                 reply_md.handler = ptlrpc_handler;
 921
 922                 /* We must see the unlink callback to set rq_reply_unlinked,
 923                  * so we can't auto-unlink
 924                  */
 925                 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
 926                                   &request->rq_reply_md_h);
 927                 if (rc != 0) {
 928                         CERROR("LNetMDAttach failed: %d\n", rc);
 929                         LASSERT(rc == -ENOMEM);
 930                         spin_lock(&request->rq_lock);
 931                         /* ...but the MD attach didn't succeed... */
 932                         request->rq_receiving_reply = 0;
 933                         spin_unlock(&request->rq_lock);
 934                         GOTO(cleanup_bulk, rc = -ENOMEM);
 935                 }
 936                 percpu_ref_get(&ptlrpc_pending);
 937
 938                 CDEBUG(D_NET,
 939                        "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 940                        request->rq_repbuf_len, request->rq_xid,
 941                        request->rq_reply_portal);
 942         }
 943
 944         /* add references on request for request_out_callback */
 945         ptlrpc_request_addref(request);
 946         if (obd != NULL && obd->obd_svc_stats != NULL)
 947                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 948                                     atomic_read(&imp->imp_inflight));
 949
 950         CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 951
 952         request->rq_sent_ns = ktime_get_real();
 953         request->rq_sent = ktime_get_real_seconds();
 954         /* We give the server rq_timeout secs to process the req, and
 955          * add the network latency for our local timeout.
 956          */
 957         request->rq_deadline = request->rq_sent + request->rq_timeout +
 958                 ptlrpc_at_get_net_latency(request);
 959
 960         DEBUG_REQ(D_INFO, request, "send flags=%x",
 961                   lustre_msg_get_flags(request->rq_reqmsg));
 962
 963         if (unlikely(opc == OBD_PING &&
 964             CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND_FAIL, cfs_fail_val))) {
 965                 DEBUG_REQ(D_INFO, request, "Simulate delay send failure");
 966                 GOTO(skip_send, rc);
 967         }
 968
 969         rc = ptl_send_buf(&request->rq_req_md_h,
 970                           request->rq_reqbuf, request->rq_reqdata_len,
 971                           LNET_NOACK_REQ, &request->rq_req_cbid,
 972                           NULL,
 973                           &connection->c_peer,
 974                           request->rq_request_portal,
 975                           request->rq_xid, 0, &bulk_cookie);
 976         if (likely(rc == 0))
 977                 GOTO(out, rc);
 978
 979 skip_send:
 980         request->rq_req_unlinked = 1;
 981         ptlrpc_req_finished(request);
 982         if (noreply)
 983                 GOTO(out, rc);
 984
 985         LNetMDUnlink(request->rq_reply_md_h);
 986
 987         /* UNLINKED callback called synchronously */
 988         LASSERT(!request->rq_receiving_reply);
 989
 990  cleanup_bulk:
 991         /* We do sync unlink here as there was no real transfer here so
 992          * the chance to have long unlink to sluggish net is smaller here.
 993          */
 994         ptlrpc_unregister_bulk(request, 0);
 995  out:
 996         if (rc == -ENOMEM) {
 997                 /* set rq_sent so that this request is treated
 998                  * as a delayed send in the upper layers
 999                  */
1000                 request->rq_sent = ktime_get_real_seconds();
1001         }
1002
1003         if (request->rq_memalloc)
1004                 memalloc_noreclaim_restore(mpflag);
1005
1006         return rc;
1007 }
1008 EXPORT_SYMBOL(ptl_send_rpc);
1009
1010 /* Register request buffer descriptor for request receiving. */
1011 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
1012 {
1013         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
1014         static struct lnet_processid match_id = {
1015                 .nid = LNET_ANY_NID,
1016                 .pid = LNET_PID_ANY
1017         };
1018         int rc;
1019         struct lnet_md md;
1020         struct lnet_me *me;
1021
1022         CDEBUG(D_NET, "%s: registering portal %d\n", service->srv_name,
1023                service->srv_req_portal);
1024
1025         if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
1026                 return -ENOMEM;
1027
1028         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
1029          * which means buffer can only be attached on local CPT, and LND
1030          * threads can find it by grabbing a local lock
1031          */
1032         me = LNetMEAttach(service->srv_req_portal,
1033                           &match_id, 0, ~0, LNET_UNLINK,
1034                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
1035                           LNET_INS_LOCAL : LNET_INS_AFTER);
1036         if (IS_ERR(me)) {
1037                 CERROR("%s: LNetMEAttach failed: rc = %ld\n",
1038                        service->srv_name, PTR_ERR(me));
1039                 return PTR_ERR(me);
1040         }
1041
1042         LASSERT(rqbd->rqbd_refcount == 0);
1043         rqbd->rqbd_refcount = 1;
1044
1045         md.start     = rqbd->rqbd_buffer;
1046         md.length    = service->srv_buf_size;
1047         md.max_size  = service->srv_max_req_size;
1048         md.threshold = LNET_MD_THRESH_INF;
1049         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
1050         md.user_ptr  = &rqbd->rqbd_cbid;
1051         md.handler   = ptlrpc_handler;
1052
1053         rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
1054         if (rc == 0) {
1055                 percpu_ref_get(&ptlrpc_pending);
1056                 return 0;
1057         }
1058
1059         CERROR("%s: LNetMDAttach failed: rc = %d\n", service->srv_name, rc);
1060         LASSERT(rc == -ENOMEM);
1061         rqbd->rqbd_refcount = 0;
1062
1063         return rc;
1064 }