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, 2016, 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
  41 /**
  42  * Helper function. Sends \a len bytes from \a base at offset \a offset
  43  * over \a conn connection to portal \a portal.
  44  * Returns 0 on success or error code.
  45  */
  46 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
  47                         enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
  48                         lnet_nid_t self, struct lnet_process_id peer_id,
  49                         int portal, __u64 xid, unsigned int offset,
  50                         struct lnet_handle_md *bulk_cookie)
  51 {
  52         int              rc;
  53         struct lnet_md         md;
  54         ENTRY;
  55
  56         LASSERT (portal != 0);
  57         CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
  58         md.start     = base;
  59         md.length    = len;
  60         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  61         md.options   = PTLRPC_MD_OPTIONS;
  62         md.user_ptr  = cbid;
  63         md.eq_handle = ptlrpc_eq_h;
  64         LNetInvalidateMDHandle(&md.bulk_handle);
  65
  66         if (bulk_cookie) {
  67                 md.bulk_handle = *bulk_cookie;
  68                 md.options |= LNET_MD_BULK_HANDLE;
  69         }
  70
  71         if (unlikely(ack == LNET_ACK_REQ &&
  72                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  73                 /* don't ask for the ack to simulate failing client */
  74                 ack = LNET_NOACK_REQ;
  75         }
  76
  77         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  78         if (unlikely(rc != 0)) {
  79                 CERROR ("LNetMDBind failed: %d\n", rc);
  80                 LASSERT (rc == -ENOMEM);
  81                 RETURN (-ENOMEM);
  82         }
  83
  84         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  85                len, portal, xid, offset);
  86
  87         rc = LNetPut(self, *mdh, ack,
  88                      peer_id, portal, xid, offset, 0);
  89         if (unlikely(rc != 0)) {
  90                 int rc2;
  91                 /* We're going to get an UNLINK event when I unlink below,
  92                  * which will complete just like any other failed send, so
  93                  * I fall through and return success here! */
  94                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  95                        libcfs_id2str(peer_id), portal, xid, rc);
  96                 rc2 = LNetMDUnlink(*mdh);
  97                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  98         }
  99
 100         RETURN (0);
 101 }
 102
 103 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
 104 {
 105         int i;
 106
 107         for (i = 0; i < count; i++)
 108                 LNetMDUnlink(bd_mds[i]);
 109 }
 110
 111 #ifdef HAVE_SERVER_SUPPORT
 112 /**
 113  * Prepare bulk descriptor for specified incoming request \a req that
 114  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 115  * the bulk to be sent. Used on server-side after request was already
 116  * received.
 117  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 118  * error.
 119  */
 120 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 121                                               unsigned nfrags, unsigned max_brw,
 122                                               unsigned int type,
 123                                               unsigned portal,
 124                                               const struct ptlrpc_bulk_frag_ops
 125                                                 *ops)
 126 {
 127         struct obd_export *exp = req->rq_export;
 128         struct ptlrpc_bulk_desc *desc;
 129
 130         ENTRY;
 131         LASSERT(ptlrpc_is_bulk_op_active(type));
 132
 133         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 134         if (desc == NULL)
 135                 RETURN(NULL);
 136
 137         desc->bd_export = class_export_get(exp);
 138         desc->bd_req = req;
 139
 140         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 141         desc->bd_cbid.cbid_arg = desc;
 142
 143         /* NB we don't assign rq_bulk here; server-side requests are
 144          * re-used, and the handler frees the bulk desc explicitly. */
 145
 146         return desc;
 147 }
 148 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 149
 150 /**
 151  * Starts bulk transfer for descriptor \a desc on the server.
 152  * Returns 0 on success or error code.
 153  */
 154 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 155 {
 156         struct obd_export        *exp = desc->bd_export;
 157         lnet_nid_t                self_nid;
 158         struct lnet_process_id    peer_id;
 159         int                       rc = 0;
 160         __u64                     mbits;
 161         int                       posted_md;
 162         int                       total_md;
 163         struct lnet_md                 md;
 164         ENTRY;
 165
 166         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 167                 RETURN(0);
 168
 169         /* NB no locking required until desc is on the network */
 170         LASSERT(desc->bd_md_count == 0);
 171         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 172
 173         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 174         LASSERT(desc->bd_cbid.cbid_arg == desc);
 175
 176         /*
 177          * Multi-Rail: get the preferred self and peer NIDs from the
 178          * request, so they are based on the route taken by the
 179          * message.
 180          */
 181         self_nid = desc->bd_req->rq_self;
 182         peer_id = desc->bd_req->rq_source;
 183
 184         /* NB total length may be 0 for a read past EOF, so we send 0
 185          * length bulks, since the client expects bulk events.
 186          *
 187          * The client may not need all of the bulk mbits for the RPC. The RPC
 188          * used the mbits of the highest bulk mbits needed, and the server masks
 189          * off high bits to get bulk count for this RPC. LU-1431 */
 190         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 191         total_md = desc->bd_req->rq_mbits - mbits + 1;
 192
 193         desc->bd_md_count = total_md;
 194         desc->bd_failure = 0;
 195
 196         md.user_ptr = &desc->bd_cbid;
 197         md.eq_handle = ptlrpc_eq_h;
 198         md.threshold = 2; /* SENT and ACK/REPLY */
 199
 200         for (posted_md = 0; posted_md < total_md; mbits++) {
 201                 md.options = PTLRPC_MD_OPTIONS;
 202
 203                 /* NB it's assumed that source and sink buffer frags are
 204                  * page-aligned. Otherwise we'd have to send client bulk
 205                  * sizes over and split server buffer accordingly */
 206                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 207                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 208                 if (rc != 0) {
 209                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 210                                exp->exp_obd->obd_name, posted_md, rc);
 211                         LASSERT(rc == -ENOMEM);
 212                         if (posted_md == 0) {
 213                                 desc->bd_md_count = 0;
 214                                 RETURN(-ENOMEM);
 215                         }
 216                         break;
 217                 }
 218
 219                 /* LU-6441: last md is not sent and desc->bd_md_count == 1 */
 220                 if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 221                                          CFS_FAIL_ONCE) &&
 222                     total_md > 1 && posted_md == total_md - 1) {
 223                         posted_md++;
 224                         continue;
 225                 }
 226
 227                 /* Network is about to get at the memory */
 228                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 229                         rc = LNetPut(self_nid, desc->bd_mds[posted_md],
 230                                      LNET_ACK_REQ, peer_id,
 231                                      desc->bd_portal, mbits, 0, 0);
 232                 else
 233                         rc = LNetGet(self_nid, desc->bd_mds[posted_md],
 234                                      peer_id, desc->bd_portal, mbits, 0);
 235
 236                 posted_md++;
 237                 if (rc != 0) {
 238                         CERROR("%s: failed bulk transfer with %s:%u x%llu: "
 239                                "rc = %d\n", exp->exp_obd->obd_name,
 240                                libcfs_id2str(peer_id), desc->bd_portal,
 241                                mbits, rc);
 242                         break;
 243                 }
 244         }
 245
 246         if (rc != 0) {
 247                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 248                  * event this creates will signal completion with failure,
 249                  * so we return SUCCESS here! */
 250                 spin_lock(&desc->bd_lock);
 251                 desc->bd_md_count -= total_md - posted_md;
 252                 spin_unlock(&desc->bd_lock);
 253                 LASSERT(desc->bd_md_count >= 0);
 254
 255                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 256                 RETURN(0);
 257         }
 258
 259         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 260                "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
 261                desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
 262                mbits - posted_md, mbits - 1);
 263
 264         RETURN(0);
 265 }
 266
 267 /**
 268  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 269  * serialises with completion callback)
 270  */
 271 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 272 {
 273         struct l_wait_info       lwi;
 274         int                      rc;
 275
 276         LASSERT(!in_interrupt());           /* might sleep */
 277
 278         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 279                 return;                         /* never started */
 280
 281         /* We used to poison the pages with 0xab here because we did not want to
 282          * send any meaningful data over the wire for evicted clients (bug 9297)
 283          * However, this is no longer safe now that we use the page cache on the
 284          * OSS (bug 20560) */
 285
 286         /* The unlink ensures the callback happens ASAP and is the last
 287          * one.  If it fails, it must be because completion just happened,
 288          * but we must still l_wait_event() in this case, to give liblustre
 289          * a chance to run server_bulk_callback()*/
 290         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 291
 292         for (;;) {
 293                 /* Network access will complete in finite time but the HUGE
 294                  * timeout lets us CWARN for visibility of sluggish NALs */
 295                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 296                                            cfs_time_seconds(1), NULL, NULL);
 297                 rc = l_wait_event(desc->bd_waitq,
 298                                   !ptlrpc_server_bulk_active(desc), &lwi);
 299                 if (rc == 0)
 300                         return;
 301
 302                 LASSERT(rc == -ETIMEDOUT);
 303                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 304         }
 305 }
 306 #endif /* HAVE_SERVER_SUPPORT */
 307
 308 /**
 309  * Register bulk at the sender for later transfer.
 310  * Returns 0 on success or error code.
 311  */
 312 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 313 {
 314         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 315         struct lnet_process_id peer;
 316         int rc = 0;
 317         int rc2;
 318         int posted_md;
 319         int total_md;
 320         __u64 mbits;
 321         struct lnet_handle_me me_h;
 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_h;
 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                 rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 379                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 380                 if (rc != 0) {
 381                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 382                                desc->bd_import->imp_obd->obd_name, mbits,
 383                                posted_md, rc);
 384                         break;
 385                 }
 386
 387                 /* About to let the network at it... */
 388                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 389                                   &desc->bd_mds[posted_md]);
 390                 if (rc != 0) {
 391                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 392                                desc->bd_import->imp_obd->obd_name, mbits,
 393                                posted_md, rc);
 394                         rc2 = LNetMEUnlink(me_h);
 395                         LASSERT(rc2 == 0);
 396                         break;
 397                 }
 398         }
 399
 400         if (rc != 0) {
 401                 LASSERT(rc == -ENOMEM);
 402                 spin_lock(&desc->bd_lock);
 403                 desc->bd_md_count -= total_md - posted_md;
 404                 spin_unlock(&desc->bd_lock);
 405                 LASSERT(desc->bd_md_count >= 0);
 406                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 407                 req->rq_status = -ENOMEM;
 408                 RETURN(-ENOMEM);
 409         }
 410
 411         spin_lock(&desc->bd_lock);
 412         /* Holler if peer manages to touch buffers before he knows the mbits */
 413         if (desc->bd_md_count != total_md)
 414                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 415                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 416                       total_md - desc->bd_md_count);
 417         spin_unlock(&desc->bd_lock);
 418
 419         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 420                "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
 421                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 422                desc->bd_iov_count, desc->bd_nob,
 423                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 424
 425         RETURN(0);
 426 }
 427
 428 /**
 429  * Disconnect a bulk desc from the network. Idempotent. Not
 430  * thread-safe (i.e. only interlocks with completion callback).
 431  * Returns 1 on success or 0 if network unregistration failed for whatever
 432  * reason.
 433  */
 434 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 435 {
 436         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 437         struct l_wait_info       lwi;
 438         int                      rc;
 439         ENTRY;
 440
 441         LASSERT(!in_interrupt());     /* might sleep */
 442
 443         /* Let's setup deadline for reply unlink. */
 444         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 445             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 446                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 447
 448         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 449                 RETURN(1);                              /* never registered */
 450
 451         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 452
 453         /* the unlink ensures the callback happens ASAP and is the last
 454          * one.  If it fails, it must be because completion just happened,
 455          * but we must still l_wait_event() in this case to give liblustre
 456          * a chance to run client_bulk_callback() */
 457         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 458
 459         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 460                 RETURN(1);                              /* never registered */
 461
 462         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 463         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 464
 465         /* Do not wait for unlink to finish. */
 466         if (async)
 467                 RETURN(0);
 468
 469         for (;;) {
 470                 /* The wq argument is ignored by user-space wait_event macros */
 471                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 472                                         &req->rq_set->set_waitq :
 473                                         &req->rq_reply_waitq;
 474                 /* Network access will complete in finite time but the HUGE
 475                  * timeout lets us CWARN for visibility of sluggish NALs */
 476                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 477                                            cfs_time_seconds(1), NULL, NULL);
 478                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 479                 if (rc == 0) {
 480                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 481                         RETURN(1);
 482                 }
 483
 484                 LASSERT(rc == -ETIMEDOUT);
 485                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 486                           desc);
 487         }
 488         RETURN(0);
 489 }
 490
 491 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 492 {
 493         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 494         struct ptlrpc_service           *svc = svcpt->scp_service;
 495         int service_time = max_t(int, ktime_get_real_seconds() -
 496                                  req->rq_arrival_time.tv_sec, 1);
 497
 498         if (!(flags & PTLRPC_REPLY_EARLY) &&
 499             (req->rq_type != PTL_RPC_MSG_ERR) &&
 500             (req->rq_reqmsg != NULL) &&
 501             !(lustre_msg_get_flags(req->rq_reqmsg) &
 502               (MSG_RESENT | MSG_REPLAY |
 503                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 504                 /* early replies, errors and recovery requests don't count
 505                  * toward our service time estimate */
 506                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 507
 508                 if (oldse != 0) {
 509                         DEBUG_REQ(D_ADAPTTO, req,
 510                                   "svc %s changed estimate from %d to %d",
 511                                   svc->srv_name, oldse,
 512                                   at_get(&svcpt->scp_at_estimate));
 513                 }
 514         }
 515         /* Report actual service time for client latency calc */
 516         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 517         /* Report service time estimate for future client reqs, but report 0
 518          * (to be ignored by client) if it's an error reply during recovery.
 519          * b=15815
 520          */
 521         if (req->rq_type == PTL_RPC_MSG_ERR &&
 522             (req->rq_export == NULL ||
 523              req->rq_export->exp_obd->obd_recovering)) {
 524                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 525         } else {
 526                 time64_t timeout;
 527
 528                 if (req->rq_export && req->rq_reqmsg != NULL &&
 529                     (flags & PTLRPC_REPLY_EARLY) &&
 530                     lustre_msg_get_flags(req->rq_reqmsg) &
 531                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
 532                         struct obd_device *exp_obd = req->rq_export->exp_obd;
 533
 534                         timeout = ktime_get_real_seconds() -
 535                                   req->rq_arrival_time.tv_sec +
 536                                   min_t(time64_t, at_extra,
 537                                         exp_obd->obd_recovery_timeout / 4);
 538                 } else {
 539                         timeout = at_get(&svcpt->scp_at_estimate);
 540                 }
 541                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 542         }
 543
 544         if (req->rq_reqmsg &&
 545             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 546                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 547                        "req_flags=%#x magic=%x/%x len=%d\n",
 548                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 549                        lustre_msg_get_magic(req->rq_reqmsg),
 550                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 551         }
 552 }
 553
 554 /**
 555  * Send request reply from request \a req reply buffer.
 556  * \a flags defines reply types
 557  * Returns 0 on success or error code
 558  */
 559 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 560 {
 561         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 562         struct ptlrpc_connection  *conn;
 563         int                        rc;
 564
 565         /* We must already have a reply buffer (only ptlrpc_error() may be
 566          * called without one). The reply generated by sptlrpc layer (e.g.
 567          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 568          * have a request buffer which is either the actual (swabbed) incoming
 569          * request, or a saved copy if this is a req saved in
 570          * target_queue_final_reply().
 571          */
 572         LASSERT (req->rq_no_reply == 0);
 573         LASSERT (req->rq_reqbuf != NULL);
 574         LASSERT (rs != NULL);
 575         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 576         LASSERT (req->rq_repmsg != NULL);
 577         LASSERT (req->rq_repmsg == rs->rs_msg);
 578         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 579         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 580
 581         /* There may be no rq_export during failover */
 582
 583         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 584                      req->rq_export->exp_obd->obd_fail)) {
 585                 /* Failed obd's only send ENODEV */
 586                 req->rq_type = PTL_RPC_MSG_ERR;
 587                 req->rq_status = -ENODEV;
 588                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 589                        req->rq_export->exp_obd->obd_minor);
 590         }
 591
 592         /* In order to keep interoprability with the client (< 2.3) which
 593          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 594          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 595          * reply buffer on client will be overflow.
 596          *
 597          * XXX Remove this whenver we drop the interoprability with such client.
 598          */
 599         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 600                                            sizeof(struct ptlrpc_body_v2), 1);
 601
 602         if (req->rq_type != PTL_RPC_MSG_ERR)
 603                 req->rq_type = PTL_RPC_MSG_REPLY;
 604
 605         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 606         lustre_msg_set_status(req->rq_repmsg,
 607                               ptlrpc_status_hton(req->rq_status));
 608         lustre_msg_set_opc(req->rq_repmsg,
 609                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 610
 611         target_pack_pool_reply(req);
 612
 613         ptlrpc_at_set_reply(req, flags);
 614
 615         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 616                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 617         else
 618                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 619
 620         if (unlikely(conn == NULL)) {
 621                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 622                 return -ENOTCONN;
 623         }
 624         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 625
 626         rc = sptlrpc_svc_wrap_reply(req);
 627         if (unlikely(rc))
 628                 goto out;
 629
 630         req->rq_sent = ktime_get_real_seconds();
 631
 632         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 633                           (rs->rs_difficult && !rs->rs_no_ack) ?
 634                           LNET_ACK_REQ : LNET_NOACK_REQ,
 635                           &rs->rs_cb_id, req->rq_self, req->rq_source,
 636                           ptlrpc_req2svc(req)->srv_rep_portal,
 637                           req->rq_xid, req->rq_reply_off, NULL);
 638 out:
 639         if (unlikely(rc != 0))
 640                 ptlrpc_req_drop_rs(req);
 641         ptlrpc_connection_put(conn);
 642         return rc;
 643 }
 644
 645 int ptlrpc_reply (struct ptlrpc_request *req)
 646 {
 647         if (req->rq_no_reply)
 648                 return 0;
 649         else
 650                 return (ptlrpc_send_reply(req, 0));
 651 }
 652
 653 /**
 654  * For request \a req send an error reply back. Create empty
 655  * reply buffers if necessary.
 656  */
 657 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 658 {
 659         int rc;
 660         ENTRY;
 661
 662         if (req->rq_no_reply)
 663                 RETURN(0);
 664
 665         if (!req->rq_repmsg) {
 666                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 667                 if (rc)
 668                         RETURN(rc);
 669         }
 670
 671         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 672             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 673             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 674                 req->rq_type = PTL_RPC_MSG_ERR;
 675
 676         rc = ptlrpc_send_reply(req, may_be_difficult);
 677         RETURN(rc);
 678 }
 679
 680 int ptlrpc_error(struct ptlrpc_request *req)
 681 {
 682         return ptlrpc_send_error(req, 0);
 683 }
 684
 685 /**
 686  * Send request \a request.
 687  * if \a noreply is set, don't expect any reply back and don't set up
 688  * reply buffers.
 689  * Returns 0 on success or error code.
 690  */
 691 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 692 {
 693         int rc;
 694         int rc2;
 695         int mpflag = 0;
 696         struct lnet_handle_md bulk_cookie;
 697         struct ptlrpc_connection *connection;
 698         struct lnet_handle_me reply_me_h;
 699         struct lnet_md reply_md;
 700         struct obd_import *imp = request->rq_import;
 701         struct obd_device *obd = imp->imp_obd;
 702         ENTRY;
 703
 704         LNetInvalidateMDHandle(&bulk_cookie);
 705
 706         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 707                 RETURN(0);
 708
 709         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 710         LASSERT(request->rq_wait_ctx == 0);
 711
 712         /* If this is a re-transmit, we're required to have disengaged
 713          * cleanly from the previous attempt */
 714         LASSERT(!request->rq_receiving_reply);
 715         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 716                 (imp->imp_state == LUSTRE_IMP_FULL)));
 717
 718         if (unlikely(obd != NULL && obd->obd_fail)) {
 719                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 720                         obd->obd_name);
 721                 /* this prevents us from waiting in ptlrpc_queue_wait */
 722                 spin_lock(&request->rq_lock);
 723                 request->rq_err = 1;
 724                 spin_unlock(&request->rq_lock);
 725                 request->rq_status = -ENODEV;
 726                 RETURN(-ENODEV);
 727         }
 728
 729         connection = imp->imp_connection;
 730
 731         lustre_msg_set_handle(request->rq_reqmsg,
 732                               &imp->imp_remote_handle);
 733         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 734         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 735                                 imp->imp_conn_cnt);
 736         lustre_msghdr_set_flags(request->rq_reqmsg,
 737                                 imp->imp_msghdr_flags);
 738
 739         /* If it's the first time to resend the request for EINPROGRESS,
 740          * we need to allocate a new XID (see after_reply()), it's different
 741          * from the resend for reply timeout. */
 742         if (request->rq_nr_resend != 0 &&
 743             list_empty(&request->rq_unreplied_list)) {
 744                 __u64 min_xid = 0;
 745                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 746                  * reconstruction */
 747                 spin_lock(&imp->imp_lock);
 748                 ptlrpc_assign_next_xid_nolock(request);
 749                 min_xid = ptlrpc_known_replied_xid(imp);
 750                 spin_unlock(&imp->imp_lock);
 751
 752                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 753                 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
 754                           "resend on EINPROGRESS");
 755         }
 756
 757         if (request->rq_bulk != NULL) {
 758                 ptlrpc_set_bulk_mbits(request);
 759                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 760         }
 761
 762         if (list_empty(&request->rq_unreplied_list) ||
 763             request->rq_xid <= imp->imp_known_replied_xid) {
 764                 DEBUG_REQ(D_ERROR, request, "xid: %llu, replied: %llu, "
 765                           "list_empty:%d\n", request->rq_xid,
 766                           imp->imp_known_replied_xid,
 767                           list_empty(&request->rq_unreplied_list));
 768                 LBUG();
 769         }
 770
 771         /** For enabled AT all request should have AT_SUPPORT in the
 772          * FULL import state when OBD_CONNECT_AT is set */
 773         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 774                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 775                 !(imp->imp_connect_data.ocd_connect_flags &
 776                 OBD_CONNECT_AT));
 777
 778         if (request->rq_resend) {
 779                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 780                 if (request->rq_resend_cb != NULL)
 781                         request->rq_resend_cb(request, &request->rq_async_args);
 782         }
 783         if (request->rq_memalloc)
 784                 mpflag = cfs_memory_pressure_get_and_set();
 785
 786         rc = sptlrpc_cli_wrap_request(request);
 787         if (rc)
 788                 GOTO(out, rc);
 789
 790         /* bulk register should be done after wrap_request() */
 791         if (request->rq_bulk != NULL) {
 792                 rc = ptlrpc_register_bulk (request);
 793                 if (rc != 0)
 794                         GOTO(out, rc);
 795                 /*
 796                  * All the mds in the request will have the same cpt
 797                  * encoded in the cookie. So we can just get the first
 798                  * one.
 799                  */
 800                 bulk_cookie = request->rq_bulk->bd_mds[0];
 801         }
 802
 803         if (!noreply) {
 804                 LASSERT (request->rq_replen != 0);
 805                 if (request->rq_repbuf == NULL) {
 806                         LASSERT(request->rq_repdata == NULL);
 807                         LASSERT(request->rq_repmsg == NULL);
 808                         rc = sptlrpc_cli_alloc_repbuf(request,
 809                                                       request->rq_replen);
 810                         if (rc) {
 811                                 /* this prevents us from looping in
 812                                  * ptlrpc_queue_wait */
 813                                 spin_lock(&request->rq_lock);
 814                                 request->rq_err = 1;
 815                                 spin_unlock(&request->rq_lock);
 816                                 request->rq_status = rc;
 817                                 GOTO(cleanup_bulk, rc);
 818                         }
 819                 } else {
 820                         request->rq_repdata = NULL;
 821                         request->rq_repmsg = NULL;
 822                 }
 823
 824                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 825                                   connection->c_peer, request->rq_xid, 0,
 826                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 827                 if (rc != 0) {
 828                         CERROR("LNetMEAttach failed: %d\n", rc);
 829                         LASSERT (rc == -ENOMEM);
 830                         GOTO(cleanup_bulk, rc = -ENOMEM);
 831                 }
 832         }
 833
 834         spin_lock(&request->rq_lock);
 835         /* We are responsible for unlinking the reply buffer */
 836         request->rq_reply_unlinked = noreply;
 837         request->rq_receiving_reply = !noreply;
 838         /* Clear any flags that may be present from previous sends. */
 839         request->rq_req_unlinked = 0;
 840         request->rq_replied = 0;
 841         request->rq_err = 0;
 842         request->rq_timedout = 0;
 843         request->rq_net_err = 0;
 844         request->rq_resend = 0;
 845         request->rq_restart = 0;
 846         request->rq_reply_truncated = 0;
 847         spin_unlock(&request->rq_lock);
 848
 849         if (!noreply) {
 850                 reply_md.start     = request->rq_repbuf;
 851                 reply_md.length    = request->rq_repbuf_len;
 852                 /* Allow multiple early replies */
 853                 reply_md.threshold = LNET_MD_THRESH_INF;
 854                 /* Manage remote for early replies */
 855                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 856                         LNET_MD_MANAGE_REMOTE |
 857                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 858                 reply_md.user_ptr  = &request->rq_reply_cbid;
 859                 reply_md.eq_handle = ptlrpc_eq_h;
 860
 861                 /* We must see the unlink callback to set rq_reply_unlinked,
 862                  * so we can't auto-unlink */
 863                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 864                                   &request->rq_reply_md_h);
 865                 if (rc != 0) {
 866                         CERROR("LNetMDAttach failed: %d\n", rc);
 867                         LASSERT (rc == -ENOMEM);
 868                         spin_lock(&request->rq_lock);
 869                         /* ...but the MD attach didn't succeed... */
 870                         request->rq_receiving_reply = 0;
 871                         spin_unlock(&request->rq_lock);
 872                         GOTO(cleanup_me, rc = -ENOMEM);
 873                 }
 874
 875                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
 876                        ", portal %u\n",
 877                        request->rq_repbuf_len, request->rq_xid,
 878                        request->rq_reply_portal);
 879         }
 880
 881         /* add references on request for request_out_callback */
 882         ptlrpc_request_addref(request);
 883         if (obd != NULL && obd->obd_svc_stats != NULL)
 884                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 885                         atomic_read(&imp->imp_inflight));
 886
 887         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 888
 889         request->rq_sent_ns = ktime_get_real();
 890         request->rq_sent = ktime_get_real_seconds();
 891         /* We give the server rq_timeout secs to process the req, and
 892            add the network latency for our local timeout. */
 893         request->rq_deadline = request->rq_sent + request->rq_timeout +
 894                 ptlrpc_at_get_net_latency(request);
 895
 896         ptlrpc_pinger_sending_on_import(imp);
 897
 898         DEBUG_REQ(D_INFO, request, "send flg=%x",
 899                   lustre_msg_get_flags(request->rq_reqmsg));
 900         rc = ptl_send_buf(&request->rq_req_md_h,
 901                           request->rq_reqbuf, request->rq_reqdata_len,
 902                           LNET_NOACK_REQ, &request->rq_req_cbid,
 903                           LNET_NID_ANY, connection->c_peer,
 904                           request->rq_request_portal,
 905                           request->rq_xid, 0, &bulk_cookie);
 906         if (likely(rc == 0))
 907                 GOTO(out, rc);
 908
 909         request->rq_req_unlinked = 1;
 910         ptlrpc_req_finished(request);
 911         if (noreply)
 912                 GOTO(out, rc);
 913
 914  cleanup_me:
 915         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 916          * nobody apart from the PUT's target has the right nid+XID to
 917          * access the reply buffer. */
 918         rc2 = LNetMEUnlink(reply_me_h);
 919         LASSERT (rc2 == 0);
 920         /* UNLINKED callback called synchronously */
 921         LASSERT(!request->rq_receiving_reply);
 922
 923  cleanup_bulk:
 924         /* We do sync unlink here as there was no real transfer here so
 925          * the chance to have long unlink to sluggish net is smaller here. */
 926         ptlrpc_unregister_bulk(request, 0);
 927  out:
 928         if (rc == -ENOMEM) {
 929                 /* set rq_sent so that this request is treated
 930                  * as a delayed send in the upper layers */
 931                 request->rq_sent = ktime_get_real_seconds();
 932         }
 933
 934         if (request->rq_memalloc)
 935                 cfs_memory_pressure_restore(mpflag);
 936
 937         return rc;
 938 }
 939 EXPORT_SYMBOL(ptl_send_rpc);
 940
 941 /**
 942  * Register request buffer descriptor for request receiving.
 943  */
 944 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 945 {
 946         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 947         static struct lnet_process_id match_id = {
 948                 .nid = LNET_NID_ANY,
 949                 .pid = LNET_PID_ANY
 950         };
 951         int rc;
 952         struct lnet_md md;
 953         struct lnet_handle_me me_h;
 954
 955         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 956                service->srv_req_portal);
 957
 958         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 959                 return (-ENOMEM);
 960
 961         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 962          * which means buffer can only be attached on local CPT, and LND
 963          * threads can find it by grabbing a local lock */
 964         rc = LNetMEAttach(service->srv_req_portal,
 965                           match_id, 0, ~0, LNET_UNLINK,
 966                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 967                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 968         if (rc != 0) {
 969                 CERROR("LNetMEAttach failed: %d\n", rc);
 970                 return (-ENOMEM);
 971         }
 972
 973         LASSERT(rqbd->rqbd_refcount == 0);
 974         rqbd->rqbd_refcount = 1;
 975
 976         md.start     = rqbd->rqbd_buffer;
 977         md.length    = service->srv_buf_size;
 978         md.max_size  = service->srv_max_req_size;
 979         md.threshold = LNET_MD_THRESH_INF;
 980         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 981         md.user_ptr  = &rqbd->rqbd_cbid;
 982         md.eq_handle = ptlrpc_eq_h;
 983
 984         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 985         if (rc == 0)
 986                 return (0);
 987
 988         CERROR("LNetMDAttach failed: %d; \n", rc);
 989         LASSERT (rc == -ENOMEM);
 990         rc = LNetMEUnlink (me_h);
 991         LASSERT (rc == 0);
 992         rqbd->rqbd_refcount = 0;
 993
 994         return (-ENOMEM);
 995 }