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