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_h;
  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_h;
 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         struct l_wait_info       lwi;
 272         int                      rc;
 273
 274         LASSERT(!in_interrupt());           /* might sleep */
 275
 276         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 277                 return;                         /* never started */
 278
 279         /* We used to poison the pages with 0xab here because we did not want to
 280          * send any meaningful data over the wire for evicted clients (bug 9297)
 281          * However, this is no longer safe now that we use the page cache on the
 282          * OSS (bug 20560) */
 283
 284         /* The unlink ensures the callback happens ASAP and is the last
 285          * one.  If it fails, it must be because completion just happened,
 286          * but we must still l_wait_event() in this case, to give liblustre
 287          * a chance to run server_bulk_callback()*/
 288         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 289
 290         for (;;) {
 291                 /* Network access will complete in finite time but the HUGE
 292                  * timeout lets us CWARN for visibility of sluggish NALs */
 293                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 294                                            cfs_time_seconds(1), NULL, NULL);
 295                 rc = l_wait_event(desc->bd_waitq,
 296                                   !ptlrpc_server_bulk_active(desc), &lwi);
 297                 if (rc == 0)
 298                         return;
 299
 300                 LASSERT(rc == -ETIMEDOUT);
 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 rc2;
 316         int posted_md;
 317         int total_md;
 318         __u64 mbits;
 319         struct lnet_handle_me me_h;
 320         struct lnet_md md;
 321         ENTRY;
 322
 323         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 324                 RETURN(0);
 325
 326         /* NB no locking required until desc is on the network */
 327         LASSERT(desc->bd_nob > 0);
 328         LASSERT(desc->bd_md_count == 0);
 329         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 330         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 331         LASSERT(desc->bd_req != NULL);
 332         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 333
 334         /* cleanup the state of the bulk for it will be reused */
 335         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 336                 desc->bd_nob_transferred = 0;
 337         else if (desc->bd_nob_transferred != 0)
 338                 /* If the network failed after an RPC was sent, this condition
 339                  * could happen.  Rather than assert (was here before), return
 340                  * an EIO error. */
 341                 RETURN(-EIO);
 342
 343         desc->bd_failure = 0;
 344
 345         peer = desc->bd_import->imp_connection->c_peer;
 346
 347         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 348         LASSERT(desc->bd_cbid.cbid_arg == desc);
 349
 350         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 351         /* rq_mbits is matchbits of the final bulk */
 352         mbits = req->rq_mbits - total_md + 1;
 353
 354         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 355                  "first mbits = x%llu, last mbits = x%llu\n",
 356                  mbits, req->rq_mbits);
 357         LASSERTF(!(desc->bd_registered &&
 358                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 359                  mbits != desc->bd_last_mbits,
 360                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 361                  desc->bd_registered, mbits, desc->bd_last_mbits);
 362
 363         desc->bd_registered = 1;
 364         desc->bd_last_mbits = mbits;
 365         desc->bd_md_count = total_md;
 366         md.user_ptr = &desc->bd_cbid;
 367         md.eq_handle = ptlrpc_eq_h;
 368         md.threshold = 1;                       /* PUT or GET */
 369
 370         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 371                 md.options = PTLRPC_MD_OPTIONS |
 372                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 373                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 374                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 375
 376                 if (posted_md > 0 && posted_md + 1 == total_md &&
 377                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 378                         rc = -ENOMEM;
 379                 } else {
 380                         rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 381                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 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_h, 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                         rc2 = LNetMEUnlink(me_h);
 398                         LASSERT(rc2 == 0);
 399                         break;
 400                 }
 401         }
 402
 403         if (rc != 0) {
 404                 LASSERT(rc == -ENOMEM);
 405                 spin_lock(&desc->bd_lock);
 406                 desc->bd_md_count -= total_md - posted_md;
 407                 spin_unlock(&desc->bd_lock);
 408                 LASSERT(desc->bd_md_count >= 0);
 409                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 410                 req->rq_status = -ENOMEM;
 411                 desc->bd_registered = 0;
 412                 RETURN(-ENOMEM);
 413         }
 414
 415         spin_lock(&desc->bd_lock);
 416         /* Holler if peer manages to touch buffers before he knows the mbits */
 417         if (desc->bd_md_count != total_md)
 418                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 419                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 420                       total_md - desc->bd_md_count);
 421         spin_unlock(&desc->bd_lock);
 422
 423         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 424                "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
 425                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 426                desc->bd_iov_count, desc->bd_nob,
 427                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 428
 429         RETURN(0);
 430 }
 431
 432 /**
 433  * Disconnect a bulk desc from the network. Idempotent. Not
 434  * thread-safe (i.e. only interlocks with completion callback).
 435  * Returns 1 on success or 0 if network unregistration failed for whatever
 436  * reason.
 437  */
 438 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 439 {
 440         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 441         struct l_wait_info       lwi;
 442         int                      rc;
 443         ENTRY;
 444
 445         LASSERT(!in_interrupt());     /* might sleep */
 446
 447         /* Let's setup deadline for reply unlink. */
 448         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 449             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 450                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 451
 452         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 453                 RETURN(1);                              /* never registered */
 454
 455         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 456
 457         /* the unlink ensures the callback happens ASAP and is the last
 458          * one.  If it fails, it must be because completion just happened,
 459          * but we must still l_wait_event() in this case to give liblustre
 460          * a chance to run client_bulk_callback() */
 461         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 462
 463         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 464                 RETURN(1);                              /* never registered */
 465
 466         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 467         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 468
 469         /* Do not wait for unlink to finish. */
 470         if (async)
 471                 RETURN(0);
 472
 473         for (;;) {
 474                 /* The wq argument is ignored by user-space wait_event macros */
 475                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 476                                         &req->rq_set->set_waitq :
 477                                         &req->rq_reply_waitq;
 478                 /* Network access will complete in finite time but the HUGE
 479                  * timeout lets us CWARN for visibility of sluggish NALs */
 480                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 481                                            cfs_time_seconds(1), NULL, NULL);
 482                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 483                 if (rc == 0) {
 484                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 485                         RETURN(1);
 486                 }
 487
 488                 LASSERT(rc == -ETIMEDOUT);
 489                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 490                           desc);
 491         }
 492         RETURN(0);
 493 }
 494
 495 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 496 {
 497         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 498         struct ptlrpc_service           *svc = svcpt->scp_service;
 499         int service_time = max_t(int, ktime_get_real_seconds() -
 500                                  req->rq_arrival_time.tv_sec, 1);
 501
 502         if (!(flags & PTLRPC_REPLY_EARLY) &&
 503             (req->rq_type != PTL_RPC_MSG_ERR) &&
 504             (req->rq_reqmsg != NULL) &&
 505             !(lustre_msg_get_flags(req->rq_reqmsg) &
 506               (MSG_RESENT | MSG_REPLAY |
 507                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 508                 /* early replies, errors and recovery requests don't count
 509                  * toward our service time estimate */
 510                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 511
 512                 if (oldse != 0) {
 513                         DEBUG_REQ(D_ADAPTTO, req,
 514                                   "svc %s changed estimate from %d to %d",
 515                                   svc->srv_name, oldse,
 516                                   at_get(&svcpt->scp_at_estimate));
 517                 }
 518         }
 519         /* Report actual service time for client latency calc */
 520         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 521         /* Report service time estimate for future client reqs, but report 0
 522          * (to be ignored by client) if it's an error reply during recovery.
 523          * b=15815
 524          */
 525         if (req->rq_type == PTL_RPC_MSG_ERR &&
 526             (req->rq_export == NULL ||
 527              req->rq_export->exp_obd->obd_recovering)) {
 528                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 529         } else {
 530                 time64_t timeout;
 531
 532                 if (req->rq_export && req->rq_reqmsg != NULL &&
 533                     (flags & PTLRPC_REPLY_EARLY) &&
 534                     lustre_msg_get_flags(req->rq_reqmsg) &
 535                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 536                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 537
 538                         timeout = ktime_get_real_seconds() -
 539                                   req->rq_arrival_time.tv_sec +
 540                                   min_t(time64_t, at_extra,
 541                                         exp_obd->obd_recovery_timeout / 4);
 542                 } else {
 543                         timeout = at_get(&svcpt->scp_at_estimate);
 544                 }
 545                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 546         }
 547
 548         if (req->rq_reqmsg &&
 549             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 550                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 551                        "req_flags=%#x magic=%x/%x len=%d\n",
 552                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 553                        lustre_msg_get_magic(req->rq_reqmsg),
 554                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 555         }
 556 }
 557
 558 /**
 559  * Send request reply from request \a req reply buffer.
 560  * \a flags defines reply types
 561  * Returns 0 on success or error code
 562  */
 563 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 564 {
 565         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 566         struct ptlrpc_connection  *conn;
 567         int                        rc;
 568
 569         /* We must already have a reply buffer (only ptlrpc_error() may be
 570          * called without one). The reply generated by sptlrpc layer (e.g.
 571          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 572          * have a request buffer which is either the actual (swabbed) incoming
 573          * request, or a saved copy if this is a req saved in
 574          * target_queue_final_reply().
 575          */
 576         LASSERT (req->rq_no_reply == 0);
 577         LASSERT (req->rq_reqbuf != NULL);
 578         LASSERT (rs != NULL);
 579         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 580         LASSERT (req->rq_repmsg != NULL);
 581         LASSERT (req->rq_repmsg == rs->rs_msg);
 582         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 583         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 584
 585         /* There may be no rq_export during failover */
 586
 587         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 588                      req->rq_export->exp_obd->obd_fail)) {
 589                 /* Failed obd's only send ENODEV */
 590                 req->rq_type = PTL_RPC_MSG_ERR;
 591                 req->rq_status = -ENODEV;
 592                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 593                        req->rq_export->exp_obd->obd_minor);
 594         }
 595
 596         /* In order to keep interoprability with the client (< 2.3) which
 597          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 598          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 599          * reply buffer on client will be overflow.
 600          *
 601          * XXX Remove this whenver we drop the interoprability with such client.
 602          */
 603         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 604                                            sizeof(struct ptlrpc_body_v2), 1);
 605
 606         if (req->rq_type != PTL_RPC_MSG_ERR)
 607                 req->rq_type = PTL_RPC_MSG_REPLY;
 608
 609         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 610         lustre_msg_set_status(req->rq_repmsg,
 611                               ptlrpc_status_hton(req->rq_status));
 612         lustre_msg_set_opc(req->rq_repmsg,
 613                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 614
 615         target_pack_pool_reply(req);
 616
 617         ptlrpc_at_set_reply(req, flags);
 618
 619         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 620                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 621         else
 622                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 623
 624         if (unlikely(conn == NULL)) {
 625                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 626                 return -ENOTCONN;
 627         }
 628         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 629
 630         rc = sptlrpc_svc_wrap_reply(req);
 631         if (unlikely(rc))
 632                 goto out;
 633
 634         req->rq_sent = ktime_get_real_seconds();
 635
 636         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 637                           (rs->rs_difficult && !rs->rs_no_ack) ?
 638                           LNET_ACK_REQ : LNET_NOACK_REQ,
 639                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 640                           ptlrpc_req2svc(req)->srv_rep_portal,
 641                           req->rq_xid, req->rq_reply_off, NULL);
 642 out:
 643         if (unlikely(rc != 0))
 644                 ptlrpc_req_drop_rs(req);
 645         ptlrpc_connection_put(conn);
 646         return rc;
 647 }
 648
 649 int ptlrpc_reply (struct ptlrpc_request *req)
 650 {
 651         if (req->rq_no_reply)
 652                 return 0;
 653         else
 654                 return (ptlrpc_send_reply(req, 0));
 655 }
 656
 657 /**
 658  * For request \a req send an error reply back. Create empty
 659  * reply buffers if necessary.
 660  */
 661 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 662 {
 663         int rc;
 664         ENTRY;
 665
 666         if (req->rq_no_reply)
 667                 RETURN(0);
 668
 669         if (!req->rq_repmsg) {
 670                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 671                 if (rc)
 672                         RETURN(rc);
 673         }
 674
 675         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 676             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 677             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 678                 req->rq_type = PTL_RPC_MSG_ERR;
 679
 680         rc = ptlrpc_send_reply(req, may_be_difficult);
 681         RETURN(rc);
 682 }
 683
 684 int ptlrpc_error(struct ptlrpc_request *req)
 685 {
 686         return ptlrpc_send_error(req, 0);
 687 }
 688
 689 /**
 690  * Send request \a request.
 691  * if \a noreply is set, don't expect any reply back and don't set up
 692  * reply buffers.
 693  * Returns 0 on success or error code.
 694  */
 695 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 696 {
 697         int rc;
 698         int rc2;
 699         int mpflag = 0;
 700         struct lnet_handle_md bulk_cookie;
 701         struct ptlrpc_connection *connection;
 702         struct lnet_handle_me reply_me_h;
 703         struct lnet_md reply_md;
 704         struct obd_import *imp = request->rq_import;
 705         struct obd_device *obd = imp->imp_obd;
 706         ENTRY;
 707
 708         LNetInvalidateMDHandle(&bulk_cookie);
 709
 710         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 711                 RETURN(0);
 712
 713         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 714         LASSERT(request->rq_wait_ctx == 0);
 715
 716         /* If this is a re-transmit, we're required to have disengaged
 717          * cleanly from the previous attempt */
 718         LASSERT(!request->rq_receiving_reply);
 719         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 720                 (imp->imp_state == LUSTRE_IMP_FULL)));
 721
 722         if (unlikely(obd != NULL && obd->obd_fail)) {
 723                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 724                         obd->obd_name);
 725                 /* this prevents us from waiting in ptlrpc_queue_wait */
 726                 spin_lock(&request->rq_lock);
 727                 request->rq_err = 1;
 728                 spin_unlock(&request->rq_lock);
 729                 request->rq_status = -ENODEV;
 730                 RETURN(-ENODEV);
 731         }
 732
 733         connection = imp->imp_connection;
 734
 735         lustre_msg_set_handle(request->rq_reqmsg,
 736                               &imp->imp_remote_handle);
 737         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 738         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 739                                 imp->imp_conn_cnt);
 740         lustre_msghdr_set_flags(request->rq_reqmsg,
 741                                 imp->imp_msghdr_flags);
 742
 743         /* If it's the first time to resend the request for EINPROGRESS,
 744          * we need to allocate a new XID (see after_reply()), it's different
 745          * from the resend for reply timeout. */
 746         if (request->rq_nr_resend != 0 &&
 747             list_empty(&request->rq_unreplied_list)) {
 748                 __u64 min_xid = 0;
 749                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 750                  * reconstruction */
 751                 spin_lock(&imp->imp_lock);
 752                 ptlrpc_assign_next_xid_nolock(request);
 753                 min_xid = ptlrpc_known_replied_xid(imp);
 754                 spin_unlock(&imp->imp_lock);
 755
 756                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 757                 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
 758                           "resend on EINPROGRESS");
 759         }
 760
 761         if (request->rq_bulk != NULL) {
 762                 ptlrpc_set_bulk_mbits(request);
 763                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 764         }
 765
 766         if (list_empty(&request->rq_unreplied_list) ||
 767             request->rq_xid <= imp->imp_known_replied_xid) {
 768                 DEBUG_REQ(D_ERROR, request, "xid: %llu, replied: %llu, "
 769                           "list_empty:%d\n", request->rq_xid,
 770                           imp->imp_known_replied_xid,
 771                           list_empty(&request->rq_unreplied_list));
 772                 LBUG();
 773         }
 774
 775         /** For enabled AT all request should have AT_SUPPORT in the
 776          * FULL import state when OBD_CONNECT_AT is set */
 777         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 778                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 779                 !(imp->imp_connect_data.ocd_connect_flags &
 780                 OBD_CONNECT_AT));
 781
 782         if (request->rq_resend) {
 783                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 784                 if (request->rq_resend_cb != NULL)
 785                         request->rq_resend_cb(request, &request->rq_async_args);
 786         }
 787         if (request->rq_memalloc)
 788                 mpflag = cfs_memory_pressure_get_and_set();
 789
 790         rc = sptlrpc_cli_wrap_request(request);
 791         if (rc)
 792                 GOTO(out, rc);
 793
 794         /* bulk register should be done after wrap_request() */
 795         if (request->rq_bulk != NULL) {
 796                 rc = ptlrpc_register_bulk (request);
 797                 if (rc != 0)
 798                         GOTO(cleanup_bulk, rc);
 799                 /*
 800                  * All the mds in the request will have the same cpt
 801                  * encoded in the cookie. So we can just get the first
 802                  * one.
 803                  */
 804                 bulk_cookie = request->rq_bulk->bd_mds[0];
 805         }
 806
 807         if (!noreply) {
 808                 LASSERT (request->rq_replen != 0);
 809                 if (request->rq_repbuf == NULL) {
 810                         LASSERT(request->rq_repdata == NULL);
 811                         LASSERT(request->rq_repmsg == NULL);
 812                         rc = sptlrpc_cli_alloc_repbuf(request,
 813                                                       request->rq_replen);
 814                         if (rc) {
 815                                 /* this prevents us from looping in
 816                                  * ptlrpc_queue_wait */
 817                                 spin_lock(&request->rq_lock);
 818                                 request->rq_err = 1;
 819                                 spin_unlock(&request->rq_lock);
 820                                 request->rq_status = rc;
 821                                 GOTO(cleanup_bulk, rc);
 822                         }
 823                 } else {
 824                         request->rq_repdata = NULL;
 825                         request->rq_repmsg = NULL;
 826                 }
 827
 828                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 829                                   connection->c_peer, request->rq_xid, 0,
 830                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 831                 if (rc != 0) {
 832                         CERROR("LNetMEAttach failed: %d\n", rc);
 833                         LASSERT (rc == -ENOMEM);
 834                         GOTO(cleanup_bulk, rc = -ENOMEM);
 835                 }
 836         }
 837
 838         spin_lock(&request->rq_lock);
 839         /* We are responsible for unlinking the reply buffer */
 840         request->rq_reply_unlinked = noreply;
 841         request->rq_receiving_reply = !noreply;
 842         /* Clear any flags that may be present from previous sends. */
 843         request->rq_req_unlinked = 0;
 844         request->rq_replied = 0;
 845         request->rq_err = 0;
 846         request->rq_timedout = 0;
 847         request->rq_net_err = 0;
 848         request->rq_resend = 0;
 849         request->rq_restart = 0;
 850         request->rq_reply_truncated = 0;
 851         spin_unlock(&request->rq_lock);
 852
 853         if (!noreply) {
 854                 reply_md.start     = request->rq_repbuf;
 855                 reply_md.length    = request->rq_repbuf_len;
 856                 /* Allow multiple early replies */
 857                 reply_md.threshold = LNET_MD_THRESH_INF;
 858                 /* Manage remote for early replies */
 859                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 860                         LNET_MD_MANAGE_REMOTE |
 861                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 862                 reply_md.user_ptr  = &request->rq_reply_cbid;
 863                 reply_md.eq_handle = ptlrpc_eq_h;
 864
 865                 /* We must see the unlink callback to set rq_reply_unlinked,
 866                  * so we can't auto-unlink */
 867                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 868                                   &request->rq_reply_md_h);
 869                 if (rc != 0) {
 870                         CERROR("LNetMDAttach failed: %d\n", rc);
 871                         LASSERT (rc == -ENOMEM);
 872                         spin_lock(&request->rq_lock);
 873                         /* ...but the MD attach didn't succeed... */
 874                         request->rq_receiving_reply = 0;
 875                         spin_unlock(&request->rq_lock);
 876                         GOTO(cleanup_me, rc = -ENOMEM);
 877                 }
 878
 879                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
 880                        ", portal %u\n",
 881                        request->rq_repbuf_len, request->rq_xid,
 882                        request->rq_reply_portal);
 883         }
 884
 885         /* add references on request for request_out_callback */
 886         ptlrpc_request_addref(request);
 887         if (obd != NULL && obd->obd_svc_stats != NULL)
 888                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 889                         atomic_read(&imp->imp_inflight));
 890
 891         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 892
 893         request->rq_sent_ns = ktime_get_real();
 894         request->rq_sent = ktime_get_real_seconds();
 895         /* We give the server rq_timeout secs to process the req, and
 896            add the network latency for our local timeout. */
 897         request->rq_deadline = request->rq_sent + request->rq_timeout +
 898                 ptlrpc_at_get_net_latency(request);
 899
 900         ptlrpc_pinger_sending_on_import(imp);
 901
 902         DEBUG_REQ(D_INFO, request, "send flg=%x",
 903                   lustre_msg_get_flags(request->rq_reqmsg));
 904         rc = ptl_send_buf(&request->rq_req_md_h,
 905                           request->rq_reqbuf, request->rq_reqdata_len,
 906                           LNET_NOACK_REQ, &request->rq_req_cbid,
 907                           LNET_NID_ANY, connection->c_peer,
 908                           request->rq_request_portal,
 909                           request->rq_xid, 0, &bulk_cookie);
 910         if (likely(rc == 0))
 911                 GOTO(out, rc);
 912
 913         request->rq_req_unlinked = 1;
 914         ptlrpc_req_finished(request);
 915         if (noreply)
 916                 GOTO(out, rc);
 917
 918  cleanup_me:
 919         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 920          * nobody apart from the PUT's target has the right nid+XID to
 921          * access the reply buffer. */
 922         rc2 = LNetMEUnlink(reply_me_h);
 923         LASSERT (rc2 == 0);
 924         /* UNLINKED callback called synchronously */
 925         LASSERT(!request->rq_receiving_reply);
 926
 927  cleanup_bulk:
 928         /* We do sync unlink here as there was no real transfer here so
 929          * the chance to have long unlink to sluggish net is smaller here. */
 930         ptlrpc_unregister_bulk(request, 0);
 931  out:
 932         if (rc == -ENOMEM) {
 933                 /* set rq_sent so that this request is treated
 934                  * as a delayed send in the upper layers */
 935                 request->rq_sent = ktime_get_real_seconds();
 936         }
 937
 938         if (request->rq_memalloc)
 939                 cfs_memory_pressure_restore(mpflag);
 940
 941         return rc;
 942 }
 943 EXPORT_SYMBOL(ptl_send_rpc);
 944
 945 /**
 946  * Register request buffer descriptor for request receiving.
 947  */
 948 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 949 {
 950         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 951         static struct lnet_process_id match_id = {
 952                 .nid = LNET_NID_ANY,
 953                 .pid = LNET_PID_ANY
 954         };
 955         int rc;
 956         struct lnet_md md;
 957         struct lnet_handle_me me_h;
 958
 959         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 960                service->srv_req_portal);
 961
 962         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 963                 return (-ENOMEM);
 964
 965         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 966          * which means buffer can only be attached on local CPT, and LND
 967          * threads can find it by grabbing a local lock */
 968         rc = LNetMEAttach(service->srv_req_portal,
 969                           match_id, 0, ~0, LNET_UNLINK,
 970                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 971                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 972         if (rc != 0) {
 973                 CERROR("LNetMEAttach failed: %d\n", rc);
 974                 return (-ENOMEM);
 975         }
 976
 977         LASSERT(rqbd->rqbd_refcount == 0);
 978         rqbd->rqbd_refcount = 1;
 979
 980         md.start     = rqbd->rqbd_buffer;
 981         md.length    = service->srv_buf_size;
 982         md.max_size  = service->srv_max_req_size;
 983         md.threshold = LNET_MD_THRESH_INF;
 984         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 985         md.user_ptr  = &rqbd->rqbd_cbid;
 986         md.eq_handle = ptlrpc_eq_h;
 987
 988         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 989         if (rc == 0)
 990                 return (0);
 991
 992         CERROR("LNetMDAttach failed: %d; \n", rc);
 993         LASSERT (rc == -ENOMEM);
 994         rc = LNetMEUnlink (me_h);
 995         LASSERT (rc == 0);
 996         rqbd->rqbd_refcount = 0;
 997
 998         return (-ENOMEM);
 999 }