lustre/ptlrpc/niobuf.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_RPC
  34 #include <obd_support.h>
  35 #include <lustre_net.h>
  36 #include <lustre_lib.h>
  37 #include <obd.h>
  38 #include <obd_class.h>
  39 #include "ptlrpc_internal.h"
  40 #include <lnet/lib-lnet.h> /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */
  41
  42 /**
  43  * Helper function. Sends \a len bytes from \a base at offset \a offset
  44  * over \a conn connection to portal \a portal.
  45  * Returns 0 on success or error code.
  46  */
  47 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
  48                         enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
  49                         lnet_nid_t self, struct lnet_process_id peer_id,
  50                         int portal, __u64 xid, unsigned int offset,
  51                         struct lnet_handle_md *bulk_cookie)
  52 {
  53         int              rc;
  54         struct lnet_md         md;
  55         ENTRY;
  56
  57         LASSERT (portal != 0);
  58         CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
  59         md.start     = base;
  60         md.length    = len;
  61         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  62         md.options   = PTLRPC_MD_OPTIONS;
  63         md.user_ptr  = cbid;
  64         md.eq_handle = ptlrpc_eq;
  65         LNetInvalidateMDHandle(&md.bulk_handle);
  66
  67         if (bulk_cookie) {
  68                 md.bulk_handle = *bulk_cookie;
  69                 md.options |= LNET_MD_BULK_HANDLE;
  70         }
  71
  72         if (unlikely(ack == LNET_ACK_REQ &&
  73                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  74                 /* don't ask for the ack to simulate failing client */
  75                 ack = LNET_NOACK_REQ;
  76         }
  77
  78         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  79         if (unlikely(rc != 0)) {
  80                 CERROR ("LNetMDBind failed: %d\n", rc);
  81                 LASSERT (rc == -ENOMEM);
  82                 RETURN (-ENOMEM);
  83         }
  84
  85         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  86                len, portal, xid, offset);
  87
  88         rc = LNetPut(self, *mdh, ack,
  89                      peer_id, portal, xid, offset, 0);
  90         if (unlikely(rc != 0)) {
  91                 int rc2;
  92                 /* We're going to get an UNLINK event when I unlink below,
  93                  * which will complete just like any other failed send, so
  94                  * I fall through and return success here! */
  95                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  96                        libcfs_id2str(peer_id), portal, xid, rc);
  97                 rc2 = LNetMDUnlink(*mdh);
  98                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  99         }
 100
 101         RETURN (0);
 102 }
 103
 104 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
 105 {
 106         int i;
 107
 108         for (i = 0; i < count; i++)
 109                 LNetMDUnlink(bd_mds[i]);
 110 }
 111
 112 #ifdef HAVE_SERVER_SUPPORT
 113 /**
 114  * Prepare bulk descriptor for specified incoming request \a req that
 115  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 116  * the bulk to be sent. Used on server-side after request was already
 117  * received.
 118  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 119  * error.
 120  */
 121 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 122                                               unsigned nfrags, unsigned max_brw,
 123                                               unsigned int type,
 124                                               unsigned portal,
 125                                               const struct ptlrpc_bulk_frag_ops
 126                                                 *ops)
 127 {
 128         struct obd_export *exp = req->rq_export;
 129         struct ptlrpc_bulk_desc *desc;
 130
 131         ENTRY;
 132         LASSERT(ptlrpc_is_bulk_op_active(type));
 133
 134         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 135         if (desc == NULL)
 136                 RETURN(NULL);
 137
 138         desc->bd_export = class_export_get(exp);
 139         desc->bd_req = req;
 140
 141         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 142         desc->bd_cbid.cbid_arg = desc;
 143
 144         /* NB we don't assign rq_bulk here; server-side requests are
 145          * re-used, and the handler frees the bulk desc explicitly. */
 146
 147         return desc;
 148 }
 149 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 150
 151 /**
 152  * Starts bulk transfer for descriptor \a desc on the server.
 153  * Returns 0 on success or error code.
 154  */
 155 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 156 {
 157         struct obd_export        *exp = desc->bd_export;
 158         lnet_nid_t                self_nid;
 159         struct lnet_process_id    peer_id;
 160         int                       rc = 0;
 161         __u64                     mbits;
 162         int                       posted_md;
 163         int                       total_md;
 164         struct lnet_md                 md;
 165         ENTRY;
 166
 167         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 168                 RETURN(0);
 169
 170         /* NB no locking required until desc is on the network */
 171         LASSERT(desc->bd_md_count == 0);
 172         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 173
 174         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 175         LASSERT(desc->bd_cbid.cbid_arg == desc);
 176
 177         /*
 178          * Multi-Rail: get the preferred self and peer NIDs from the
 179          * request, so they are based on the route taken by the
 180          * message.
 181          */
 182         self_nid = desc->bd_req->rq_self;
 183         peer_id = desc->bd_req->rq_source;
 184
 185         /* NB total length may be 0 for a read past EOF, so we send 0
 186          * length bulks, since the client expects bulk events.
 187          *
 188          * The client may not need all of the bulk mbits for the RPC. The RPC
 189          * used the mbits of the highest bulk mbits needed, and the server masks
 190          * off high bits to get bulk count for this RPC. LU-1431 */
 191         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 192         total_md = desc->bd_req->rq_mbits - mbits + 1;
 193
 194         desc->bd_md_count = total_md;
 195         desc->bd_failure = 0;
 196
 197         md.user_ptr = &desc->bd_cbid;
 198         md.eq_handle = ptlrpc_eq;
 199         md.threshold = 2; /* SENT and ACK/REPLY */
 200
 201         for (posted_md = 0; posted_md < total_md; mbits++) {
 202                 md.options = PTLRPC_MD_OPTIONS;
 203
 204                 /* NB it's assumed that source and sink buffer frags are
 205                  * page-aligned. Otherwise we'd have to send client bulk
 206                  * sizes over and split server buffer accordingly */
 207                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 208                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 209                 if (rc != 0) {
 210                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 211                                exp->exp_obd->obd_name, posted_md, rc);
 212                         LASSERT(rc == -ENOMEM);
 213                         if (posted_md == 0) {
 214                                 desc->bd_md_count = 0;
 215                                 RETURN(-ENOMEM);
 216                         }
 217                         break;
 218                 }
 219
 220                 /* sanity.sh 224c: lets skip last md */
 221                 if (posted_md == desc->bd_md_max_brw - 1)
 222                         OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 223                                              CFS_FAIL_PTLRPC_OST_BULK_CB2);
 224
 225                 /* Network is about to get at the memory */
 226                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 227                         rc = LNetPut(self_nid, desc->bd_mds[posted_md],
 228                                      LNET_ACK_REQ, peer_id,
 229                                      desc->bd_portal, mbits, 0, 0);
 230                 else
 231                         rc = LNetGet(self_nid, desc->bd_mds[posted_md],
 232                                      peer_id, desc->bd_portal, mbits, 0, false);
 233
 234                 posted_md++;
 235                 if (rc != 0) {
 236                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 237                                "rc = %d\n", exp->exp_obd->obd_name,
 238                                libcfs_id2str(peer_id), desc->bd_portal,
 239                                mbits, rc);
 240                         break;
 241                 }
 242         }
 243
 244         if (rc != 0) {
 245                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 246                  * event this creates will signal completion with failure,
 247                  * so we return SUCCESS here! */
 248                 spin_lock(&desc->bd_lock);
 249                 desc->bd_md_count -= total_md - posted_md;
 250                 spin_unlock(&desc->bd_lock);
 251                 LASSERT(desc->bd_md_count >= 0);
 252
 253                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 254                 RETURN(0);
 255         }
 256
 257         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 258                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 259                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
 260                mbits - posted_md, mbits - 1);
 261
 262         RETURN(0);
 263 }
 264
 265 /**
 266  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 267  * serialises with completion callback)
 268  */
 269 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 270 {
 271         LASSERT(!in_interrupt());           /* might sleep */
 272
 273         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 274                 return;                         /* never started */
 275
 276         /* We used to poison the pages with 0xab here because we did not want to
 277          * send any meaningful data over the wire for evicted clients (bug 9297)
 278          * However, this is no longer safe now that we use the page cache on the
 279          * OSS (bug 20560) */
 280
 281         /* The unlink ensures the callback happens ASAP and is the last
 282          * one.  If it fails, it must be because completion just happened,
 283          * but we must still wait_event_idle_timeout() in this case, to give
 284          * us a chance to run server_bulk_callback()
 285          */
 286         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 287
 288         for (;;) {
 289                 /* Network access will complete in finite time but the HUGE
 290                  * timeout lets us CWARN for visibility of sluggish NALs */
 291                 int seconds = LONG_UNLINK;
 292
 293                 while (seconds > 0 &&
 294                        wait_event_idle_timeout(desc->bd_waitq,
 295                                                !ptlrpc_server_bulk_active(desc),
 296                                                cfs_time_seconds(1)) == 0)
 297                         seconds -= 1;
 298                 if (seconds > 0)
 299                         return;
 300
 301                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 302         }
 303 }
 304 #endif /* HAVE_SERVER_SUPPORT */
 305
 306 /**
 307  * Register bulk at the sender for later transfer.
 308  * Returns 0 on success or error code.
 309  */
 310 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 311 {
 312         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 313         struct lnet_process_id peer;
 314         int rc = 0;
 315         int posted_md;
 316         int total_md;
 317         __u64 mbits;
 318         struct lnet_me *me;
 319         struct lnet_md md;
 320         ENTRY;
 321
 322         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 323                 RETURN(0);
 324
 325         /* NB no locking required until desc is on the network */
 326         LASSERT(desc->bd_nob > 0);
 327         LASSERT(desc->bd_md_count == 0);
 328         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 329         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 330         LASSERT(desc->bd_req != NULL);
 331         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 332
 333         /* cleanup the state of the bulk for it will be reused */
 334         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 335                 desc->bd_nob_transferred = 0;
 336         else if (desc->bd_nob_transferred != 0)
 337                 /* If the network failed after an RPC was sent, this condition
 338                  * could happen.  Rather than assert (was here before), return
 339                  * an EIO error. */
 340                 RETURN(-EIO);
 341
 342         desc->bd_failure = 0;
 343
 344         peer = desc->bd_import->imp_connection->c_peer;
 345
 346         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 347         LASSERT(desc->bd_cbid.cbid_arg == desc);
 348
 349         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 350         /* rq_mbits is matchbits of the final bulk */
 351         mbits = req->rq_mbits - total_md + 1;
 352
 353         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 354                  "first mbits = x%llu, last mbits = x%llu\n",
 355                  mbits, req->rq_mbits);
 356         LASSERTF(!(desc->bd_registered &&
 357                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 358                  mbits != desc->bd_last_mbits,
 359                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 360                  desc->bd_registered, mbits, desc->bd_last_mbits);
 361
 362         desc->bd_registered = 1;
 363         desc->bd_last_mbits = mbits;
 364         desc->bd_md_count = total_md;
 365         md.user_ptr = &desc->bd_cbid;
 366         md.eq_handle = ptlrpc_eq;
 367         md.threshold = 1;                       /* PUT or GET */
 368
 369         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 370                 md.options = PTLRPC_MD_OPTIONS |
 371                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 372                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 373                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 374
 375                 if (posted_md > 0 && posted_md + 1 == total_md &&
 376                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 377                         rc = -ENOMEM;
 378                 } else {
 379                         me = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 380                                   LNET_UNLINK, LNET_INS_AFTER);
 381                         rc = PTR_ERR_OR_ZERO(me);
 382                 }
 383                 if (rc != 0) {
 384                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 385                                desc->bd_import->imp_obd->obd_name, mbits,
 386                                posted_md, rc);
 387                         break;
 388                 }
 389
 390                 /* About to let the network at it... */
 391                 rc = LNetMDAttach(me, md, LNET_UNLINK,
 392                                   &desc->bd_mds[posted_md]);
 393                 if (rc != 0) {
 394                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 395                                desc->bd_import->imp_obd->obd_name, mbits,
 396                                posted_md, rc);
 397                         LNetMEUnlink(me);
 398                         break;
 399                 }
 400         }
 401
 402         if (rc != 0) {
 403                 LASSERT(rc == -ENOMEM);
 404                 spin_lock(&desc->bd_lock);
 405                 desc->bd_md_count -= total_md - posted_md;
 406                 spin_unlock(&desc->bd_lock);
 407                 LASSERT(desc->bd_md_count >= 0);
 408                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 409                 req->rq_status = -ENOMEM;
 410                 desc->bd_registered = 0;
 411                 RETURN(-ENOMEM);
 412         }
 413
 414         spin_lock(&desc->bd_lock);
 415         /* Holler if peer manages to touch buffers before he knows the mbits */
 416         if (desc->bd_md_count != total_md)
 417                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 418                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 419                       total_md - desc->bd_md_count);
 420         spin_unlock(&desc->bd_lock);
 421
 422         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 423                "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
 424                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 425                desc->bd_iov_count, desc->bd_nob,
 426                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 427
 428         RETURN(0);
 429 }
 430
 431 /**
 432  * Disconnect a bulk desc from the network. Idempotent. Not
 433  * thread-safe (i.e. only interlocks with completion callback).
 434  * Returns 1 on success or 0 if network unregistration failed for whatever
 435  * reason.
 436  */
 437 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 438 {
 439         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 440         ENTRY;
 441
 442         LASSERT(!in_interrupt());     /* might sleep */
 443
 444         /* Let's setup deadline for reply unlink. */
 445         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 446             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 447                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 448
 449         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 450                 RETURN(1);                              /* never registered */
 451
 452         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 453
 454         /* the unlink ensures the callback happens ASAP and is the last
 455          * one.  If it fails, it must be because completion just happened,
 456          * but we must still wait_event_idle_timeout() in this case to give
 457          * us a chance to run client_bulk_callback()
 458          */
 459         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 460
 461         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 462                 RETURN(1);                              /* never registered */
 463
 464         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 465         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 466
 467         /* Do not wait for unlink to finish. */
 468         if (async)
 469                 RETURN(0);
 470
 471         for (;;) {
 472                 /* The wq argument is ignored by user-space wait_event macros */
 473                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 474                                         &req->rq_set->set_waitq :
 475                                         &req->rq_reply_waitq;
 476                 /*
 477                  * Network access will complete in finite time but the HUGE
 478                  * timeout lets us CWARN for visibility of sluggish NALs.
 479                  */
 480                 int seconds = LONG_UNLINK;
 481
 482                 while (seconds > 0 &&
 483                        wait_event_idle_timeout(*wq,
 484                                                !ptlrpc_client_bulk_active(req),
 485                                                cfs_time_seconds(1)) == 0)
 486                         seconds -= 1;
 487                 if (seconds > 0) {
 488                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 489                         RETURN(1);
 490                 }
 491
 492                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 493                           desc);
 494         }
 495         RETURN(0);
 496 }
 497
 498 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 499 {
 500         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 501         struct ptlrpc_service           *svc = svcpt->scp_service;
 502         int service_time = max_t(int, ktime_get_real_seconds() -
 503                                  req->rq_arrival_time.tv_sec, 1);
 504
 505         if (!(flags & PTLRPC_REPLY_EARLY) &&
 506             (req->rq_type != PTL_RPC_MSG_ERR) &&
 507             (req->rq_reqmsg != NULL) &&
 508             !(lustre_msg_get_flags(req->rq_reqmsg) &
 509               (MSG_RESENT | MSG_REPLAY |
 510                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 511                 /* early replies, errors and recovery requests don't count
 512                  * toward our service time estimate */
 513                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 514
 515                 if (oldse != 0) {
 516                         DEBUG_REQ(D_ADAPTTO, req,
 517                                   "svc %s changed estimate from %d to %d",
 518                                   svc->srv_name, oldse,
 519                                   at_get(&svcpt->scp_at_estimate));
 520                 }
 521         }
 522         /* Report actual service time for client latency calc */
 523         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 524         /* Report service time estimate for future client reqs, but report 0
 525          * (to be ignored by client) if it's an error reply during recovery.
 526          * b=15815
 527          */
 528         if (req->rq_type == PTL_RPC_MSG_ERR &&
 529             (req->rq_export == NULL ||
 530              req->rq_export->exp_obd->obd_recovering)) {
 531                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 532         } else {
 533                 time64_t timeout;
 534
 535                 if (req->rq_export && req->rq_reqmsg != NULL &&
 536                     (flags & PTLRPC_REPLY_EARLY) &&
 537                     lustre_msg_get_flags(req->rq_reqmsg) &
 538                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 539                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 540
 541                         timeout = ktime_get_real_seconds() -
 542                                   req->rq_arrival_time.tv_sec +
 543                                   min_t(time64_t, at_extra,
 544                                         exp_obd->obd_recovery_timeout / 4);
 545                 } else {
 546                         timeout = at_get(&svcpt->scp_at_estimate);
 547                 }
 548                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 549         }
 550
 551         if (req->rq_reqmsg &&
 552             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 553                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 554                        "req_flags=%#x magic=%x/%x len=%d\n",
 555                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 556                        lustre_msg_get_magic(req->rq_reqmsg),
 557                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 558         }
 559 }
 560
 561 /**
 562  * Send request reply from request \a req reply buffer.
 563  * \a flags defines reply types
 564  * Returns 0 on success or error code
 565  */
 566 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 567 {
 568         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 569         struct ptlrpc_connection  *conn;
 570         int                        rc;
 571
 572         /* We must already have a reply buffer (only ptlrpc_error() may be
 573          * called without one). The reply generated by sptlrpc layer (e.g.
 574          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 575          * have a request buffer which is either the actual (swabbed) incoming
 576          * request, or a saved copy if this is a req saved in
 577          * target_queue_final_reply().
 578          */
 579         LASSERT (req->rq_no_reply == 0);
 580         LASSERT (req->rq_reqbuf != NULL);
 581         LASSERT (rs != NULL);
 582         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 583         LASSERT (req->rq_repmsg != NULL);
 584         LASSERT (req->rq_repmsg == rs->rs_msg);
 585         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 586         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 587
 588         /* There may be no rq_export during failover */
 589
 590         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 591                      req->rq_export->exp_obd->obd_fail)) {
 592                 /* Failed obd's only send ENODEV */
 593                 req->rq_type = PTL_RPC_MSG_ERR;
 594                 req->rq_status = -ENODEV;
 595                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 596                        req->rq_export->exp_obd->obd_minor);
 597         }
 598
 599         if (req->rq_type != PTL_RPC_MSG_ERR)
 600                 req->rq_type = PTL_RPC_MSG_REPLY;
 601
 602         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 603         lustre_msg_set_status(req->rq_repmsg,
 604                               ptlrpc_status_hton(req->rq_status));
 605         lustre_msg_set_opc(req->rq_repmsg,
 606                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 607
 608         target_pack_pool_reply(req);
 609
 610         ptlrpc_at_set_reply(req, flags);
 611
 612         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 613                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 614         else
 615                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 616
 617         if (unlikely(conn == NULL)) {
 618                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 619                 return -ENOTCONN;
 620         }
 621         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 622
 623         rc = sptlrpc_svc_wrap_reply(req);
 624         if (unlikely(rc))
 625                 goto out;
 626
 627         req->rq_sent = ktime_get_real_seconds();
 628
 629         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 630                           (rs->rs_difficult && !rs->rs_no_ack) ?
 631                           LNET_ACK_REQ : LNET_NOACK_REQ,
 632                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 633                           ptlrpc_req2svc(req)->srv_rep_portal,
 634                           req->rq_xid, req->rq_reply_off, NULL);
 635 out:
 636         if (unlikely(rc != 0))
 637                 ptlrpc_req_drop_rs(req);
 638         ptlrpc_connection_put(conn);
 639         return rc;
 640 }
 641
 642 int ptlrpc_reply (struct ptlrpc_request *req)
 643 {
 644         if (req->rq_no_reply)
 645                 return 0;
 646         else
 647                 return (ptlrpc_send_reply(req, 0));
 648 }
 649
 650 /**
 651  * For request \a req send an error reply back. Create empty
 652  * reply buffers if necessary.
 653  */
 654 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 655 {
 656         int rc;
 657         ENTRY;
 658
 659         if (req->rq_no_reply)
 660                 RETURN(0);
 661
 662         if (!req->rq_repmsg) {
 663                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 664                 if (rc)
 665                         RETURN(rc);
 666         }
 667
 668         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 669             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 670             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 671                 req->rq_type = PTL_RPC_MSG_ERR;
 672
 673         rc = ptlrpc_send_reply(req, may_be_difficult);
 674         RETURN(rc);
 675 }
 676
 677 int ptlrpc_error(struct ptlrpc_request *req)
 678 {
 679         return ptlrpc_send_error(req, 0);
 680 }
 681
 682 /**
 683  * Send request \a request.
 684  * if \a noreply is set, don't expect any reply back and don't set up
 685  * reply buffers.
 686  * Returns 0 on success or error code.
 687  */
 688 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 689 {
 690         int rc;
 691         int mpflag = 0;
 692         struct lnet_handle_md bulk_cookie;
 693         struct ptlrpc_connection *connection;
 694         struct lnet_me *reply_me = NULL;
 695         struct lnet_md reply_md;
 696         struct obd_import *imp = request->rq_import;
 697         struct obd_device *obd = imp->imp_obd;
 698         ENTRY;
 699
 700         LNetInvalidateMDHandle(&bulk_cookie);
 701
 702         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 703                 RETURN(0);
 704
 705         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 706         LASSERT(request->rq_wait_ctx == 0);
 707
 708         /* If this is a re-transmit, we're required to have disengaged
 709          * cleanly from the previous attempt */
 710         LASSERT(!request->rq_receiving_reply);
 711         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 712                 (imp->imp_state == LUSTRE_IMP_FULL)));
 713
 714         if (unlikely(obd != NULL && obd->obd_fail)) {
 715                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 716                         obd->obd_name);
 717                 /* this prevents us from waiting in ptlrpc_queue_wait */
 718                 spin_lock(&request->rq_lock);
 719                 request->rq_err = 1;
 720                 spin_unlock(&request->rq_lock);
 721                 request->rq_status = -ENODEV;
 722                 RETURN(-ENODEV);
 723         }
 724
 725         connection = imp->imp_connection;
 726
 727         lustre_msg_set_handle(request->rq_reqmsg,
 728                               &imp->imp_remote_handle);
 729         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 730         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 731                                 imp->imp_conn_cnt);
 732         lustre_msghdr_set_flags(request->rq_reqmsg,
 733                                 imp->imp_msghdr_flags);
 734
 735         /* If it's the first time to resend the request for EINPROGRESS,
 736          * we need to allocate a new XID (see after_reply()), it's different
 737          * from the resend for reply timeout. */
 738         if (request->rq_nr_resend != 0 &&
 739             list_empty(&request->rq_unreplied_list)) {
 740                 __u64 min_xid = 0;
 741                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 742                  * reconstruction */
 743                 spin_lock(&imp->imp_lock);
 744                 ptlrpc_assign_next_xid_nolock(request);
 745                 min_xid = ptlrpc_known_replied_xid(imp);
 746                 spin_unlock(&imp->imp_lock);
 747
 748                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 749                 DEBUG_REQ(D_RPCTRACE, request,
 750                           "Allocating new XID for resend on EINPROGRESS");
 751         }
 752
 753         if (request->rq_bulk != NULL) {
 754                 ptlrpc_set_bulk_mbits(request);
 755                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 756         }
 757
 758         if (list_empty(&request->rq_unreplied_list) ||
 759             request->rq_xid <= imp->imp_known_replied_xid) {
 760                 DEBUG_REQ(D_ERROR, request,
 761                           "xid=%llu, replied=%llu, list_empty=%d",
 762                           request->rq_xid, imp->imp_known_replied_xid,
 763                           list_empty(&request->rq_unreplied_list));
 764                 LBUG();
 765         }
 766
 767         /** For enabled AT all request should have AT_SUPPORT in the
 768          * FULL import state when OBD_CONNECT_AT is set */
 769         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 770                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 771                 !(imp->imp_connect_data.ocd_connect_flags &
 772                 OBD_CONNECT_AT));
 773
 774         if (request->rq_resend) {
 775                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 776                 if (request->rq_resend_cb != NULL)
 777                         request->rq_resend_cb(request, &request->rq_async_args);
 778         }
 779         if (request->rq_memalloc)
 780                 mpflag = cfs_memory_pressure_get_and_set();
 781
 782         rc = sptlrpc_cli_wrap_request(request);
 783         if (rc)
 784                 GOTO(out, rc);
 785
 786         /* bulk register should be done after wrap_request() */
 787         if (request->rq_bulk != NULL) {
 788                 rc = ptlrpc_register_bulk (request);
 789                 if (rc != 0)
 790                         GOTO(cleanup_bulk, rc);
 791                 /*
 792                  * All the mds in the request will have the same cpt
 793                  * encoded in the cookie. So we can just get the first
 794                  * one.
 795                  */
 796                 bulk_cookie = request->rq_bulk->bd_mds[0];
 797         }
 798
 799         if (!noreply) {
 800                 LASSERT (request->rq_replen != 0);
 801                 if (request->rq_repbuf == NULL) {
 802                         LASSERT(request->rq_repdata == NULL);
 803                         LASSERT(request->rq_repmsg == NULL);
 804                         rc = sptlrpc_cli_alloc_repbuf(request,
 805                                                       request->rq_replen);
 806                         if (rc) {
 807                                 /* this prevents us from looping in
 808                                  * ptlrpc_queue_wait */
 809                                 spin_lock(&request->rq_lock);
 810                                 request->rq_err = 1;
 811                                 spin_unlock(&request->rq_lock);
 812                                 request->rq_status = rc;
 813                                 GOTO(cleanup_bulk, rc);
 814                         }
 815                 } else {
 816                         request->rq_repdata = NULL;
 817                         request->rq_repmsg = NULL;
 818                 }
 819
 820                 reply_me = LNetMEAttach(request->rq_reply_portal,
 821                                         connection->c_peer, request->rq_xid, 0,
 822                                         LNET_UNLINK, LNET_INS_AFTER);
 823                 if (IS_ERR(reply_me)) {
 824                         rc = PTR_ERR(reply_me);
 825                         CERROR("LNetMEAttach failed: %d\n", rc);
 826                         LASSERT(rc == -ENOMEM);
 827                         GOTO(cleanup_bulk, rc = -ENOMEM);
 828                 }
 829         }
 830
 831         spin_lock(&request->rq_lock);
 832         /* We are responsible for unlinking the reply buffer */
 833         request->rq_reply_unlinked = noreply;
 834         request->rq_receiving_reply = !noreply;
 835         /* Clear any flags that may be present from previous sends. */
 836         request->rq_req_unlinked = 0;
 837         request->rq_replied = 0;
 838         request->rq_err = 0;
 839         request->rq_timedout = 0;
 840         request->rq_net_err = 0;
 841         request->rq_resend = 0;
 842         request->rq_restart = 0;
 843         request->rq_reply_truncated = 0;
 844         spin_unlock(&request->rq_lock);
 845
 846         if (!noreply) {
 847                 reply_md.start     = request->rq_repbuf;
 848                 reply_md.length    = request->rq_repbuf_len;
 849                 /* Allow multiple early replies */
 850                 reply_md.threshold = LNET_MD_THRESH_INF;
 851                 /* Manage remote for early replies */
 852                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 853                         LNET_MD_MANAGE_REMOTE |
 854                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 855                 reply_md.user_ptr  = &request->rq_reply_cbid;
 856                 reply_md.eq_handle = ptlrpc_eq;
 857
 858                 /* We must see the unlink callback to set rq_reply_unlinked,
 859                  * so we can't auto-unlink */
 860                 rc = LNetMDAttach(reply_me, reply_md, LNET_RETAIN,
 861                                   &request->rq_reply_md_h);
 862                 if (rc != 0) {
 863                         CERROR("LNetMDAttach failed: %d\n", rc);
 864                         LASSERT(rc == -ENOMEM);
 865                         spin_lock(&request->rq_lock);
 866                         /* ...but the MD attach didn't succeed... */
 867                         request->rq_receiving_reply = 0;
 868                         spin_unlock(&request->rq_lock);
 869                         GOTO(cleanup_me, rc = -ENOMEM);
 870                 }
 871
 872                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
 873                        ", portal %u\n",
 874                        request->rq_repbuf_len, request->rq_xid,
 875                        request->rq_reply_portal);
 876         }
 877
 878         /* add references on request for request_out_callback */
 879         ptlrpc_request_addref(request);
 880         if (obd != NULL && obd->obd_svc_stats != NULL)
 881                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 882                         atomic_read(&imp->imp_inflight));
 883
 884         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 885
 886         request->rq_sent_ns = ktime_get_real();
 887         request->rq_sent = ktime_get_real_seconds();
 888         /* We give the server rq_timeout secs to process the req, and
 889            add the network latency for our local timeout. */
 890         request->rq_deadline = request->rq_sent + request->rq_timeout +
 891                 ptlrpc_at_get_net_latency(request);
 892
 893         ptlrpc_pinger_sending_on_import(imp);
 894
 895         DEBUG_REQ(D_INFO, request, "send flags=%x",
 896                   lustre_msg_get_flags(request->rq_reqmsg));
 897         rc = ptl_send_buf(&request->rq_req_md_h,
 898                           request->rq_reqbuf, request->rq_reqdata_len,
 899                           LNET_NOACK_REQ, &request->rq_req_cbid,
 900                           LNET_NID_ANY, connection->c_peer,
 901                           request->rq_request_portal,
 902                           request->rq_xid, 0, &bulk_cookie);
 903         if (likely(rc == 0))
 904                 GOTO(out, rc);
 905
 906         request->rq_req_unlinked = 1;
 907         ptlrpc_req_finished(request);
 908         if (noreply)
 909                 GOTO(out, rc);
 910
 911  cleanup_me:
 912         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 913          * nobody apart from the PUT's target has the right nid+XID to
 914          * access the reply buffer.
 915          */
 916         LNetMEUnlink(reply_me);
 917         /* UNLINKED callback called synchronously */
 918         LASSERT(!request->rq_receiving_reply);
 919
 920  cleanup_bulk:
 921         /* We do sync unlink here as there was no real transfer here so
 922          * the chance to have long unlink to sluggish net is smaller here. */
 923         ptlrpc_unregister_bulk(request, 0);
 924         if (request->rq_bulk != NULL)
 925                 request->rq_bulk->bd_registered = 0;
 926  out:
 927         if (rc == -ENOMEM) {
 928                 /* set rq_sent so that this request is treated
 929                  * as a delayed send in the upper layers */
 930                 request->rq_sent = ktime_get_real_seconds();
 931         }
 932
 933         if (request->rq_memalloc)
 934                 cfs_memory_pressure_restore(mpflag);
 935
 936         return rc;
 937 }
 938 EXPORT_SYMBOL(ptl_send_rpc);
 939
 940 /**
 941  * Register request buffer descriptor for request receiving.
 942  */
 943 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 944 {
 945         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 946         static struct lnet_process_id match_id = {
 947                 .nid = LNET_NID_ANY,
 948                 .pid = LNET_PID_ANY
 949         };
 950         int rc;
 951         struct lnet_md md;
 952         struct lnet_me *me;
 953
 954         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 955                service->srv_req_portal);
 956
 957         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 958                 return (-ENOMEM);
 959
 960         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 961          * which means buffer can only be attached on local CPT, and LND
 962          * threads can find it by grabbing a local lock */
 963         me = LNetMEAttach(service->srv_req_portal,
 964                           match_id, 0, ~0, LNET_UNLINK,
 965                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 966                           LNET_INS_LOCAL : LNET_INS_AFTER);
 967         if (IS_ERR(me)) {
 968                 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
 969                 return -ENOMEM;
 970         }
 971
 972         LASSERT(rqbd->rqbd_refcount == 0);
 973         rqbd->rqbd_refcount = 1;
 974
 975         md.start     = rqbd->rqbd_buffer;
 976         md.length    = service->srv_buf_size;
 977         md.max_size  = service->srv_max_req_size;
 978         md.threshold = LNET_MD_THRESH_INF;
 979         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 980         md.user_ptr  = &rqbd->rqbd_cbid;
 981         md.eq_handle = ptlrpc_eq;
 982
 983         rc = LNetMDAttach(me, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 984         if (rc == 0)
 985                 return 0;
 986
 987         CERROR("LNetMDAttach failed: %d;\n", rc);
 988         LASSERT(rc == -ENOMEM);
 989         LNetMEUnlink(me);
 990         LASSERT(rc == 0);
 991         rqbd->rqbd_refcount = 0;
 992
 993         return -ENOMEM;
 994 }