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         percpu_ref_get(&ptlrpc_pending);
  89
  90         rc = LNetPut(self, *mdh, ack,
  91                      peer_id, portal, xid, offset, 0);
  92         if (unlikely(rc != 0)) {
  93                 int rc2;
  94                 /* We're going to get an UNLINK event when I unlink below,
  95                  * which will complete just like any other failed send, so
  96                  * I fall through and return success here! */
  97                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  98                        libcfs_id2str(peer_id), portal, xid, rc);
  99                 rc2 = LNetMDUnlink(*mdh);
 100                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
 101         }
 102
 103         RETURN (0);
 104 }
 105
 106 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
 107 {
 108         int i;
 109
 110         for (i = 0; i < count; i++)
 111                 LNetMDUnlink(bd_mds[i]);
 112 }
 113
 114 #ifdef HAVE_SERVER_SUPPORT
 115 /**
 116  * Prepare bulk descriptor for specified incoming request \a req that
 117  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 118  * the bulk to be sent. Used on server-side after request was already
 119  * received.
 120  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 121  * error.
 122  */
 123 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 124                                               unsigned nfrags, unsigned max_brw,
 125                                               unsigned int type,
 126                                               unsigned portal,
 127                                               const struct ptlrpc_bulk_frag_ops
 128                                                 *ops)
 129 {
 130         struct obd_export *exp = req->rq_export;
 131         struct ptlrpc_bulk_desc *desc;
 132
 133         ENTRY;
 134         LASSERT(ptlrpc_is_bulk_op_active(type));
 135
 136         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 137         if (desc == NULL)
 138                 RETURN(NULL);
 139
 140         desc->bd_export = class_export_get(exp);
 141         desc->bd_req = req;
 142
 143         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 144         desc->bd_cbid.cbid_arg = desc;
 145
 146         /* NB we don't assign rq_bulk here; server-side requests are
 147          * re-used, and the handler frees the bulk desc explicitly. */
 148
 149         return desc;
 150 }
 151 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 152
 153 /**
 154  * Starts bulk transfer for descriptor \a desc on the server.
 155  * Returns 0 on success or error code.
 156  */
 157 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 158 {
 159         struct obd_export        *exp = desc->bd_export;
 160         lnet_nid_t                self_nid;
 161         struct lnet_process_id    peer_id;
 162         int                       rc = 0;
 163         __u64                     mbits;
 164         int                       posted_md;
 165         int                       total_md;
 166         struct lnet_md                 md;
 167         ENTRY;
 168
 169         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 170                 RETURN(0);
 171
 172         /* NB no locking required until desc is on the network */
 173         LASSERT(desc->bd_md_count == 0);
 174         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 175
 176         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 177         LASSERT(desc->bd_cbid.cbid_arg == desc);
 178
 179         /*
 180          * Multi-Rail: get the preferred self and peer NIDs from the
 181          * request, so they are based on the route taken by the
 182          * message.
 183          */
 184         self_nid = desc->bd_req->rq_self;
 185         peer_id = desc->bd_req->rq_source;
 186
 187         /* NB total length may be 0 for a read past EOF, so we send 0
 188          * length bulks, since the client expects bulk events.
 189          *
 190          * The client may not need all of the bulk mbits for the RPC. The RPC
 191          * used the mbits of the highest bulk mbits needed, and the server masks
 192          * off high bits to get bulk count for this RPC. LU-1431 */
 193         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 194         total_md = desc->bd_req->rq_mbits - mbits + 1;
 195
 196         desc->bd_md_count = total_md;
 197         desc->bd_failure = 0;
 198
 199         md.user_ptr = &desc->bd_cbid;
 200         md.eq_handle = ptlrpc_eq;
 201         md.threshold = 2; /* SENT and ACK/REPLY */
 202
 203         for (posted_md = 0; posted_md < total_md; mbits++) {
 204                 md.options = PTLRPC_MD_OPTIONS;
 205
 206                 /* NB it's assumed that source and sink buffer frags are
 207                  * page-aligned. Otherwise we'd have to send client bulk
 208                  * sizes over and split server buffer accordingly */
 209                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 210                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 211                 if (rc != 0) {
 212                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 213                                exp->exp_obd->obd_name, posted_md, rc);
 214                         LASSERT(rc == -ENOMEM);
 215                         if (posted_md == 0) {
 216                                 desc->bd_md_count = 0;
 217                                 RETURN(-ENOMEM);
 218                         }
 219                         break;
 220                 }
 221                 percpu_ref_get(&ptlrpc_pending);
 222
 223                 /* sanity.sh 224c: lets skip last md */
 224                 if (posted_md == desc->bd_md_max_brw - 1)
 225                         OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 226                                              CFS_FAIL_PTLRPC_OST_BULK_CB2);
 227
 228                 /* Network is about to get at the memory */
 229                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 230                         rc = LNetPut(self_nid, desc->bd_mds[posted_md],
 231                                      LNET_ACK_REQ, peer_id,
 232                                      desc->bd_portal, mbits, 0, 0);
 233                 else
 234                         rc = LNetGet(self_nid, desc->bd_mds[posted_md],
 235                                      peer_id, desc->bd_portal, mbits, 0, false);
 236
 237                 posted_md++;
 238                 if (rc != 0) {
 239                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 240                                "rc = %d\n", exp->exp_obd->obd_name,
 241                                libcfs_id2str(peer_id), desc->bd_portal,
 242                                mbits, rc);
 243                         break;
 244                 }
 245         }
 246
 247         if (rc != 0) {
 248                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 249                  * event this creates will signal completion with failure,
 250                  * so we return SUCCESS here! */
 251                 spin_lock(&desc->bd_lock);
 252                 desc->bd_md_count -= total_md - posted_md;
 253                 spin_unlock(&desc->bd_lock);
 254                 LASSERT(desc->bd_md_count >= 0);
 255
 256                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 257                 RETURN(0);
 258         }
 259
 260         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 261                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 262                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
 263                mbits - posted_md, mbits - 1);
 264
 265         RETURN(0);
 266 }
 267
 268 /**
 269  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 270  * serialises with completion callback)
 271  */
 272 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 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 wait_event_idle_timeout() in this case, to give
 287          * us a chance to run server_bulk_callback()
 288          */
 289         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 290
 291         for (;;) {
 292                 /* Network access will complete in finite time but the HUGE
 293                  * timeout lets us CWARN for visibility of sluggish NALs */
 294                 int seconds = LONG_UNLINK;
 295
 296                 while (seconds > 0 &&
 297                        wait_event_idle_timeout(desc->bd_waitq,
 298                                                !ptlrpc_server_bulk_active(desc),
 299                                                cfs_time_seconds(1)) == 0)
 300                         seconds -= 1;
 301                 if (seconds > 0)
 302                         return;
 303
 304                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 305         }
 306 }
 307 #endif /* HAVE_SERVER_SUPPORT */
 308
 309 /**
 310  * Register bulk at the sender for later transfer.
 311  * Returns 0 on success or error code.
 312  */
 313 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 314 {
 315         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 316         struct lnet_process_id peer;
 317         int rc = 0;
 318         int posted_md;
 319         int total_md;
 320         __u64 mbits;
 321         struct lnet_me *me;
 322         struct lnet_md md;
 323         ENTRY;
 324
 325         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 326                 RETURN(0);
 327
 328         /* NB no locking required until desc is on the network */
 329         LASSERT(desc->bd_nob > 0);
 330         LASSERT(desc->bd_md_count == 0);
 331         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 332         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 333         LASSERT(desc->bd_req != NULL);
 334         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 335
 336         /* cleanup the state of the bulk for it will be reused */
 337         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 338                 desc->bd_nob_transferred = 0;
 339         else if (desc->bd_nob_transferred != 0)
 340                 /* If the network failed after an RPC was sent, this condition
 341                  * could happen.  Rather than assert (was here before), return
 342                  * an EIO error. */
 343                 RETURN(-EIO);
 344
 345         desc->bd_failure = 0;
 346
 347         peer = desc->bd_import->imp_connection->c_peer;
 348
 349         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 350         LASSERT(desc->bd_cbid.cbid_arg == desc);
 351
 352         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 353         /* rq_mbits is matchbits of the final bulk */
 354         mbits = req->rq_mbits - total_md + 1;
 355
 356         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 357                  "first mbits = x%llu, last mbits = x%llu\n",
 358                  mbits, req->rq_mbits);
 359         LASSERTF(!(desc->bd_registered &&
 360                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 361                  mbits != desc->bd_last_mbits,
 362                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 363                  desc->bd_registered, mbits, desc->bd_last_mbits);
 364
 365         desc->bd_registered = 1;
 366         desc->bd_last_mbits = mbits;
 367         desc->bd_md_count = total_md;
 368         md.user_ptr = &desc->bd_cbid;
 369         md.eq_handle = ptlrpc_eq;
 370         md.threshold = 1;                       /* PUT or GET */
 371
 372         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 373                 md.options = PTLRPC_MD_OPTIONS |
 374                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 375                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 376                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 377
 378                 if (posted_md > 0 && posted_md + 1 == total_md &&
 379                     OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
 380                         rc = -ENOMEM;
 381                 } else {
 382                         me = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 383                                   LNET_UNLINK, LNET_INS_AFTER);
 384                         rc = PTR_ERR_OR_ZERO(me);
 385                 }
 386                 if (rc != 0) {
 387                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 388                                desc->bd_import->imp_obd->obd_name, mbits,
 389                                posted_md, rc);
 390                         break;
 391                 }
 392                 percpu_ref_get(&ptlrpc_pending);
 393
 394                 /* About to let the network at it... */
 395                 rc = LNetMDAttach(me, md, LNET_UNLINK,
 396                                   &desc->bd_mds[posted_md]);
 397                 if (rc != 0) {
 398                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 399                                desc->bd_import->imp_obd->obd_name, mbits,
 400                                posted_md, rc);
 401                         LNetMEUnlink(me);
 402                         break;
 403                 }
 404         }
 405
 406         if (rc != 0) {
 407                 LASSERT(rc == -ENOMEM);
 408                 spin_lock(&desc->bd_lock);
 409                 desc->bd_md_count -= total_md - posted_md;
 410                 spin_unlock(&desc->bd_lock);
 411                 LASSERT(desc->bd_md_count >= 0);
 412                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 413                 req->rq_status = -ENOMEM;
 414                 desc->bd_registered = 0;
 415                 RETURN(-ENOMEM);
 416         }
 417
 418         spin_lock(&desc->bd_lock);
 419         /* Holler if peer manages to touch buffers before he knows the mbits */
 420         if (desc->bd_md_count != total_md)
 421                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 422                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 423                       total_md - desc->bd_md_count);
 424         spin_unlock(&desc->bd_lock);
 425
 426         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 427                "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
 428                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 429                desc->bd_iov_count, desc->bd_nob,
 430                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 431
 432         RETURN(0);
 433 }
 434
 435 /**
 436  * Disconnect a bulk desc from the network. Idempotent. Not
 437  * thread-safe (i.e. only interlocks with completion callback).
 438  * Returns 1 on success or 0 if network unregistration failed for whatever
 439  * reason.
 440  */
 441 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 442 {
 443         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 444         ENTRY;
 445
 446         LASSERT(!in_interrupt());     /* might sleep */
 447
 448         /* Let's setup deadline for reply unlink. */
 449         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 450             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 451                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 452
 453         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 454                 RETURN(1);                              /* never registered */
 455
 456         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 457
 458         /* the unlink ensures the callback happens ASAP and is the last
 459          * one.  If it fails, it must be because completion just happened,
 460          * but we must still wait_event_idle_timeout() in this case to give
 461          * us a chance to run client_bulk_callback()
 462          */
 463         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 464
 465         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 466                 RETURN(1);                              /* never registered */
 467
 468         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 469         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 470
 471         /* Do not wait for unlink to finish. */
 472         if (async)
 473                 RETURN(0);
 474
 475         for (;;) {
 476                 /* The wq argument is ignored by user-space wait_event macros */
 477                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 478                                         &req->rq_set->set_waitq :
 479                                         &req->rq_reply_waitq;
 480                 /*
 481                  * Network access will complete in finite time but the HUGE
 482                  * timeout lets us CWARN for visibility of sluggish NALs.
 483                  */
 484                 int seconds = LONG_UNLINK;
 485
 486                 while (seconds > 0 &&
 487                        wait_event_idle_timeout(*wq,
 488                                                !ptlrpc_client_bulk_active(req),
 489                                                cfs_time_seconds(1)) == 0)
 490                         seconds -= 1;
 491                 if (seconds > 0) {
 492                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 493                         RETURN(1);
 494                 }
 495
 496                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 497                           desc);
 498         }
 499         RETURN(0);
 500 }
 501
 502 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 503 {
 504         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 505         struct ptlrpc_service           *svc = svcpt->scp_service;
 506         timeout_t service_timeout;
 507
 508         service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
 509                                              req->rq_arrival_time.tv_sec, 1,
 510                                   (AT_OFF ? obd_timeout * 3 / 2 : at_max));
 511         if (!(flags & PTLRPC_REPLY_EARLY) &&
 512             (req->rq_type != PTL_RPC_MSG_ERR) &&
 513             (req->rq_reqmsg != NULL) &&
 514             !(lustre_msg_get_flags(req->rq_reqmsg) &
 515               (MSG_RESENT | MSG_REPLAY |
 516                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 517                 /* early replies, errors and recovery requests don't count
 518                  * toward our service time estimate */
 519                 int oldse = at_measured(&svcpt->scp_at_estimate,
 520                                         service_timeout);
 521
 522                 if (oldse != 0) {
 523                         DEBUG_REQ(D_ADAPTTO, req,
 524                                   "svc %s changed estimate from %d to %d",
 525                                   svc->srv_name, oldse,
 526                                   at_get(&svcpt->scp_at_estimate));
 527                 }
 528         }
 529         /* Report actual service time for client latency calc */
 530         lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
 531         /* Report service time estimate for future client reqs, but report 0
 532          * (to be ignored by client) if it's an error reply during recovery.
 533          * b=15815
 534          */
 535         if (req->rq_type == PTL_RPC_MSG_ERR &&
 536             (req->rq_export == NULL ||
 537              req->rq_export->exp_obd->obd_recovering)) {
 538                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 539         } else {
 540                 time64_t timeout;
 541
 542                 if (req->rq_export && req->rq_reqmsg != NULL &&
 543                     (flags & PTLRPC_REPLY_EARLY) &&
 544                     lustre_msg_get_flags(req->rq_reqmsg) &
 545                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 546                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 547
 548                         timeout = ktime_get_real_seconds() -
 549                                   req->rq_arrival_time.tv_sec +
 550                                   min_t(time64_t, at_extra,
 551                                         exp_obd->obd_recovery_timeout / 4);
 552                 } else {
 553                         timeout = at_get(&svcpt->scp_at_estimate);
 554                 }
 555                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 556         }
 557
 558         if (req->rq_reqmsg &&
 559             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 560                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 561                        "req_flags=%#x magic=%x/%x len=%d\n",
 562                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 563                        lustre_msg_get_magic(req->rq_reqmsg),
 564                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 565         }
 566 }
 567
 568 /**
 569  * Send request reply from request \a req reply buffer.
 570  * \a flags defines reply types
 571  * Returns 0 on success or error code
 572  */
 573 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 574 {
 575         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 576         struct ptlrpc_connection  *conn;
 577         int                        rc;
 578
 579         /* We must already have a reply buffer (only ptlrpc_error() may be
 580          * called without one). The reply generated by sptlrpc layer (e.g.
 581          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 582          * have a request buffer which is either the actual (swabbed) incoming
 583          * request, or a saved copy if this is a req saved in
 584          * target_queue_final_reply().
 585          */
 586         LASSERT (req->rq_no_reply == 0);
 587         LASSERT (req->rq_reqbuf != NULL);
 588         LASSERT (rs != NULL);
 589         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 590         LASSERT (req->rq_repmsg != NULL);
 591         LASSERT (req->rq_repmsg == rs->rs_msg);
 592         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 593         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 594
 595         /* There may be no rq_export during failover */
 596
 597         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 598                      req->rq_export->exp_obd->obd_fail)) {
 599                 /* Failed obd's only send ENODEV */
 600                 req->rq_type = PTL_RPC_MSG_ERR;
 601                 req->rq_status = -ENODEV;
 602                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 603                        req->rq_export->exp_obd->obd_minor);
 604         }
 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 mpflag = 0;
 699         struct lnet_handle_md bulk_cookie;
 700         struct ptlrpc_connection *connection;
 701         struct lnet_me *reply_me = NULL;
 702         struct lnet_md reply_md;
 703         struct obd_import *imp = request->rq_import;
 704         struct obd_device *obd = imp->imp_obd;
 705         ENTRY;
 706
 707         LNetInvalidateMDHandle(&bulk_cookie);
 708
 709         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 710                 RETURN(0);
 711
 712         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 713         LASSERT(request->rq_wait_ctx == 0);
 714
 715         /* If this is a re-transmit, we're required to have disengaged
 716          * cleanly from the previous attempt */
 717         LASSERT(!request->rq_receiving_reply);
 718         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 719                 (imp->imp_state == LUSTRE_IMP_FULL)));
 720
 721         if (unlikely(obd != NULL && obd->obd_fail)) {
 722                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 723                         obd->obd_name);
 724                 /* this prevents us from waiting in ptlrpc_queue_wait */
 725                 spin_lock(&request->rq_lock);
 726                 request->rq_err = 1;
 727                 spin_unlock(&request->rq_lock);
 728                 request->rq_status = -ENODEV;
 729                 RETURN(-ENODEV);
 730         }
 731
 732         connection = imp->imp_connection;
 733
 734         lustre_msg_set_handle(request->rq_reqmsg,
 735                               &imp->imp_remote_handle);
 736         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 737         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 738                                 imp->imp_conn_cnt);
 739         lustre_msghdr_set_flags(request->rq_reqmsg,
 740                                 imp->imp_msghdr_flags);
 741
 742         /* If it's the first time to resend the request for EINPROGRESS,
 743          * we need to allocate a new XID (see after_reply()), it's different
 744          * from the resend for reply timeout. */
 745         if (request->rq_nr_resend != 0 &&
 746             list_empty(&request->rq_unreplied_list)) {
 747                 __u64 min_xid = 0;
 748                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 749                  * reconstruction */
 750                 spin_lock(&imp->imp_lock);
 751                 ptlrpc_assign_next_xid_nolock(request);
 752                 min_xid = ptlrpc_known_replied_xid(imp);
 753                 spin_unlock(&imp->imp_lock);
 754
 755                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 756                 DEBUG_REQ(D_RPCTRACE, request,
 757                           "Allocating new XID for resend on EINPROGRESS");
 758         }
 759
 760         if (request->rq_bulk != NULL) {
 761                 ptlrpc_set_bulk_mbits(request);
 762                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 763         }
 764
 765         if (list_empty(&request->rq_unreplied_list) ||
 766             request->rq_xid <= imp->imp_known_replied_xid) {
 767                 DEBUG_REQ(D_ERROR, request,
 768                           "xid=%llu, replied=%llu, list_empty=%d",
 769                           request->rq_xid, imp->imp_known_replied_xid,
 770                           list_empty(&request->rq_unreplied_list));
 771                 LBUG();
 772         }
 773
 774         /** For enabled AT all request should have AT_SUPPORT in the
 775          * FULL import state when OBD_CONNECT_AT is set */
 776         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 777                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 778                 !(imp->imp_connect_data.ocd_connect_flags &
 779                 OBD_CONNECT_AT));
 780
 781         if (request->rq_resend) {
 782                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 783                 if (request->rq_resend_cb != NULL)
 784                         request->rq_resend_cb(request, &request->rq_async_args);
 785         }
 786         if (request->rq_memalloc)
 787                 mpflag = cfs_memory_pressure_get_and_set();
 788
 789         rc = sptlrpc_cli_wrap_request(request);
 790         if (rc)
 791                 GOTO(out, rc);
 792
 793         /* bulk register should be done after wrap_request() */
 794         if (request->rq_bulk != NULL) {
 795                 rc = ptlrpc_register_bulk (request);
 796                 if (rc != 0)
 797                         GOTO(cleanup_bulk, rc);
 798                 /*
 799                  * All the mds in the request will have the same cpt
 800                  * encoded in the cookie. So we can just get the first
 801                  * one.
 802                  */
 803                 bulk_cookie = request->rq_bulk->bd_mds[0];
 804         }
 805
 806         if (!noreply) {
 807                 LASSERT (request->rq_replen != 0);
 808                 if (request->rq_repbuf == NULL) {
 809                         LASSERT(request->rq_repdata == NULL);
 810                         LASSERT(request->rq_repmsg == NULL);
 811                         rc = sptlrpc_cli_alloc_repbuf(request,
 812                                                       request->rq_replen);
 813                         if (rc) {
 814                                 /* this prevents us from looping in
 815                                  * ptlrpc_queue_wait */
 816                                 spin_lock(&request->rq_lock);
 817                                 request->rq_err = 1;
 818                                 spin_unlock(&request->rq_lock);
 819                                 request->rq_status = rc;
 820                                 GOTO(cleanup_bulk, rc);
 821                         }
 822                 } else {
 823                         request->rq_repdata = NULL;
 824                         request->rq_repmsg = NULL;
 825                 }
 826
 827                 reply_me = LNetMEAttach(request->rq_reply_portal,
 828                                         connection->c_peer, request->rq_xid, 0,
 829                                         LNET_UNLINK, LNET_INS_AFTER);
 830                 if (IS_ERR(reply_me)) {
 831                         rc = PTR_ERR(reply_me);
 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;
 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, 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                 percpu_ref_get(&ptlrpc_pending);
 879
 880                 CDEBUG(D_NET,
 881                        "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 882                        request->rq_repbuf_len, request->rq_xid,
 883                        request->rq_reply_portal);
 884         }
 885
 886         /* add references on request for request_out_callback */
 887         ptlrpc_request_addref(request);
 888         if (obd != NULL && obd->obd_svc_stats != NULL)
 889                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 890                         atomic_read(&imp->imp_inflight));
 891
 892         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 893
 894         request->rq_sent_ns = ktime_get_real();
 895         request->rq_sent = ktime_get_real_seconds();
 896         /* We give the server rq_timeout secs to process the req, and
 897            add the network latency for our local timeout. */
 898         request->rq_deadline = request->rq_sent + request->rq_timeout +
 899                 ptlrpc_at_get_net_latency(request);
 900
 901         ptlrpc_pinger_sending_on_import(imp);
 902
 903         DEBUG_REQ(D_INFO, request, "send flags=%x",
 904                   lustre_msg_get_flags(request->rq_reqmsg));
 905         rc = ptl_send_buf(&request->rq_req_md_h,
 906                           request->rq_reqbuf, request->rq_reqdata_len,
 907                           LNET_NOACK_REQ, &request->rq_req_cbid,
 908                           LNET_NID_ANY, connection->c_peer,
 909                           request->rq_request_portal,
 910                           request->rq_xid, 0, &bulk_cookie);
 911         if (likely(rc == 0))
 912                 GOTO(out, rc);
 913
 914         request->rq_req_unlinked = 1;
 915         ptlrpc_req_finished(request);
 916         if (noreply)
 917                 GOTO(out, rc);
 918
 919  cleanup_me:
 920         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 921          * nobody apart from the PUT's target has the right nid+XID to
 922          * access the reply buffer.
 923          */
 924         LNetMEUnlink(reply_me);
 925         /* UNLINKED callback called synchronously */
 926         LASSERT(!request->rq_receiving_reply);
 927
 928  cleanup_bulk:
 929         /* We do sync unlink here as there was no real transfer here so
 930          * the chance to have long unlink to sluggish net is smaller here. */
 931         ptlrpc_unregister_bulk(request, 0);
 932         if (request->rq_bulk != NULL)
 933                 request->rq_bulk->bd_registered = 0;
 934  out:
 935         if (rc == -ENOMEM) {
 936                 /* set rq_sent so that this request is treated
 937                  * as a delayed send in the upper layers */
 938                 request->rq_sent = ktime_get_real_seconds();
 939         }
 940
 941         if (request->rq_memalloc)
 942                 cfs_memory_pressure_restore(mpflag);
 943
 944         return rc;
 945 }
 946 EXPORT_SYMBOL(ptl_send_rpc);
 947
 948 /**
 949  * Register request buffer descriptor for request receiving.
 950  */
 951 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 952 {
 953         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 954         static struct lnet_process_id match_id = {
 955                 .nid = LNET_NID_ANY,
 956                 .pid = LNET_PID_ANY
 957         };
 958         int rc;
 959         struct lnet_md md;
 960         struct lnet_me *me;
 961
 962         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 963                service->srv_req_portal);
 964
 965         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 966                 return (-ENOMEM);
 967
 968         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 969          * which means buffer can only be attached on local CPT, and LND
 970          * threads can find it by grabbing a local lock */
 971         me = LNetMEAttach(service->srv_req_portal,
 972                           match_id, 0, ~0, LNET_UNLINK,
 973                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 974                           LNET_INS_LOCAL : LNET_INS_AFTER);
 975         if (IS_ERR(me)) {
 976                 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
 977                 return -ENOMEM;
 978         }
 979
 980         LASSERT(rqbd->rqbd_refcount == 0);
 981         rqbd->rqbd_refcount = 1;
 982
 983         md.start     = rqbd->rqbd_buffer;
 984         md.length    = service->srv_buf_size;
 985         md.max_size  = service->srv_max_req_size;
 986         md.threshold = LNET_MD_THRESH_INF;
 987         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 988         md.user_ptr  = &rqbd->rqbd_cbid;
 989         md.eq_handle = ptlrpc_eq;
 990
 991         rc = LNetMDAttach(me, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 992         if (rc == 0) {
 993                 percpu_ref_get(&ptlrpc_pending);
 994                 return 0;
 995         }
 996
 997         CERROR("ptlrpc: LNetMDAttach failed: rc = %d\n", rc);
 998         LASSERT(rc == -ENOMEM);
 999         LNetMEUnlink(me);
1000         LASSERT(rc == 0);
1001         rqbd->rqbd_refcount = 0;
1002
1003         return -ENOMEM;
1004 }