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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #include <obd_support.h>
  39 #include <lustre_net.h>
  40 #include <lustre_lib.h>
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include "ptlrpc_internal.h"
  44
  45 /**
  46  * Helper function. Sends \a len bytes from \a base at offset \a offset
  47  * over \a conn connection to portal \a portal.
  48  * Returns 0 on success or error code.
  49  */
  50 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  51                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  52                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  53                          unsigned int offset)
  54 {
  55         int              rc;
  56         lnet_md_t         md;
  57         ENTRY;
  58
  59         LASSERT (portal != 0);
  60         LASSERT (conn != NULL);
  61         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  62         md.start     = base;
  63         md.length    = len;
  64         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  65         md.options   = PTLRPC_MD_OPTIONS;
  66         md.user_ptr  = cbid;
  67         md.eq_handle = ptlrpc_eq_h;
  68
  69         if (unlikely(ack == LNET_ACK_REQ &&
  70                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  71                 /* don't ask for the ack to simulate failing client */
  72                 ack = LNET_NOACK_REQ;
  73         }
  74
  75         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  76         if (unlikely(rc != 0)) {
  77                 CERROR ("LNetMDBind failed: %d\n", rc);
  78                 LASSERT (rc == -ENOMEM);
  79                 RETURN (-ENOMEM);
  80         }
  81
  82         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  83                len, portal, xid, offset);
  84
  85         rc = LNetPut (conn->c_self, *mdh, ack,
  86                       conn->c_peer, portal, xid, offset, 0);
  87         if (unlikely(rc != 0)) {
  88                 int rc2;
  89                 /* We're going to get an UNLINK event when I unlink below,
  90                  * which will complete just like any other failed send, so
  91                  * I fall through and return success here! */
  92                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  93                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  94                 rc2 = LNetMDUnlink(*mdh);
  95                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  96         }
  97
  98         RETURN (0);
  99 }
 100
 101 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
 102 {
 103         int i;
 104
 105         for (i = 0; i < count; i++)
 106                 LNetMDUnlink(bd_mds[i]);
 107 }
 108
 109 #ifdef HAVE_SERVER_SUPPORT
 110 /**
 111  * Prepare bulk descriptor for specified incoming request \a req that
 112  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
 113  * the bulk to be sent. Used on server-side after request was already
 114  * received.
 115  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 116  * error.
 117  */
 118 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 119                                               unsigned nfrags, unsigned max_brw,
 120                                               unsigned int type,
 121                                               unsigned portal,
 122                                               const struct ptlrpc_bulk_frag_ops
 123                                                 *ops)
 124 {
 125         struct obd_export *exp = req->rq_export;
 126         struct ptlrpc_bulk_desc *desc;
 127
 128         ENTRY;
 129         LASSERT(ptlrpc_is_bulk_op_active(type));
 130
 131         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
 132         if (desc == NULL)
 133                 RETURN(NULL);
 134
 135         desc->bd_export = class_export_get(exp);
 136         desc->bd_req = req;
 137
 138         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 139         desc->bd_cbid.cbid_arg = desc;
 140
 141         /* NB we don't assign rq_bulk here; server-side requests are
 142          * re-used, and the handler frees the bulk desc explicitly. */
 143
 144         return desc;
 145 }
 146 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 147
 148 /**
 149  * Starts bulk transfer for descriptor \a desc on the server.
 150  * Returns 0 on success or error code.
 151  */
 152 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 153 {
 154         struct obd_export        *exp = desc->bd_export;
 155         struct ptlrpc_connection *conn = exp->exp_connection;
 156         int                       rc = 0;
 157         __u64                     mbits;
 158         int                       posted_md;
 159         int                       total_md;
 160         lnet_md_t                 md;
 161         ENTRY;
 162
 163         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 164                 RETURN(0);
 165
 166         /* NB no locking required until desc is on the network */
 167         LASSERT(desc->bd_md_count == 0);
 168         LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
 169
 170         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 171         LASSERT(desc->bd_cbid.cbid_arg == desc);
 172
 173         /* NB total length may be 0 for a read past EOF, so we send 0
 174          * length bulks, since the client expects bulk events.
 175          *
 176          * The client may not need all of the bulk mbits for the RPC. The RPC
 177          * used the mbits of the highest bulk mbits needed, and the server masks
 178          * off high bits to get bulk count for this RPC. LU-1431 */
 179         mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
 180         total_md = desc->bd_req->rq_mbits - mbits + 1;
 181
 182         desc->bd_md_count = total_md;
 183         desc->bd_failure = 0;
 184
 185         md.user_ptr = &desc->bd_cbid;
 186         md.eq_handle = ptlrpc_eq_h;
 187         md.threshold = 2; /* SENT and ACK/REPLY */
 188
 189         for (posted_md = 0; posted_md < total_md; mbits++) {
 190                 md.options = PTLRPC_MD_OPTIONS;
 191
 192                 /* NB it's assumed that source and sink buffer frags are
 193                  * page-aligned. Otherwise we'd have to send client bulk
 194                  * sizes over and split server buffer accordingly */
 195                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 196                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 197                 if (rc != 0) {
 198                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 199                                exp->exp_obd->obd_name, posted_md, rc);
 200                         LASSERT(rc == -ENOMEM);
 201                         if (posted_md == 0) {
 202                                 desc->bd_md_count = 0;
 203                                 RETURN(-ENOMEM);
 204                         }
 205                         break;
 206                 }
 207
 208                 /* LU-6441: last md is not sent and desc->bd_md_count == 1 */
 209                 if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
 210                                          CFS_FAIL_ONCE) &&
 211                     posted_md == desc->bd_md_max_brw - 1) {
 212                         posted_md++;
 213                         continue;
 214                 }
 215
 216                 /* Network is about to get at the memory */
 217                 if (ptlrpc_is_bulk_put_source(desc->bd_type))
 218                         rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
 219                                      LNET_ACK_REQ, conn->c_peer,
 220                                      desc->bd_portal, mbits, 0, 0);
 221                 else
 222                         rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
 223                                      conn->c_peer, desc->bd_portal, mbits, 0);
 224
 225                 posted_md++;
 226                 if (rc != 0) {
 227                         CERROR("%s: failed bulk transfer with %s:%u x"LPU64": "
 228                                "rc = %d\n", exp->exp_obd->obd_name,
 229                                libcfs_id2str(conn->c_peer), desc->bd_portal,
 230                                mbits, rc);
 231                         break;
 232                 }
 233         }
 234
 235         if (rc != 0) {
 236                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 237                  * event this creates will signal completion with failure,
 238                  * so we return SUCCESS here! */
 239                 spin_lock(&desc->bd_lock);
 240                 desc->bd_md_count -= total_md - posted_md;
 241                 spin_unlock(&desc->bd_lock);
 242                 LASSERT(desc->bd_md_count >= 0);
 243
 244                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 245                 RETURN(0);
 246         }
 247
 248         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 249                "id %s mbits "LPX64"-"LPX64"\n", desc->bd_iov_count,
 250                desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
 251                mbits - posted_md, mbits - 1);
 252
 253         RETURN(0);
 254 }
 255
 256 /**
 257  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 258  * serialises with completion callback)
 259  */
 260 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 261 {
 262         struct l_wait_info       lwi;
 263         int                      rc;
 264
 265         LASSERT(!in_interrupt());           /* might sleep */
 266
 267         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 268                 return;                         /* never started */
 269
 270         /* We used to poison the pages with 0xab here because we did not want to
 271          * send any meaningful data over the wire for evicted clients (bug 9297)
 272          * However, this is no longer safe now that we use the page cache on the
 273          * OSS (bug 20560) */
 274
 275         /* The unlink ensures the callback happens ASAP and is the last
 276          * one.  If it fails, it must be because completion just happened,
 277          * but we must still l_wait_event() in this case, to give liblustre
 278          * a chance to run server_bulk_callback()*/
 279         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 280
 281         for (;;) {
 282                 /* Network access will complete in finite time but the HUGE
 283                  * timeout lets us CWARN for visibility of sluggish NALs */
 284                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 285                                            cfs_time_seconds(1), NULL, NULL);
 286                 rc = l_wait_event(desc->bd_waitq,
 287                                   !ptlrpc_server_bulk_active(desc), &lwi);
 288                 if (rc == 0)
 289                         return;
 290
 291                 LASSERT(rc == -ETIMEDOUT);
 292                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 293         }
 294 }
 295 #endif /* HAVE_SERVER_SUPPORT */
 296
 297 /**
 298  * Register bulk at the sender for later transfer.
 299  * Returns 0 on success or error code.
 300  */
 301 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 302 {
 303         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 304         lnet_process_id_t peer;
 305         int rc = 0;
 306         int rc2;
 307         int posted_md;
 308         int total_md;
 309         __u64 mbits;
 310         lnet_handle_me_t  me_h;
 311         lnet_md_t         md;
 312         ENTRY;
 313
 314         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 315                 RETURN(0);
 316
 317         /* NB no locking required until desc is on the network */
 318         LASSERT(desc->bd_nob > 0);
 319         LASSERT(desc->bd_md_count == 0);
 320         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 321         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 322         LASSERT(desc->bd_req != NULL);
 323         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 324
 325         /* cleanup the state of the bulk for it will be reused */
 326         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 327                 desc->bd_nob_transferred = 0;
 328         else
 329                 LASSERT(desc->bd_nob_transferred == 0);
 330
 331         desc->bd_failure = 0;
 332
 333         peer = desc->bd_import->imp_connection->c_peer;
 334
 335         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 336         LASSERT(desc->bd_cbid.cbid_arg == desc);
 337
 338         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 339         /* rq_mbits is matchbits of the final bulk */
 340         mbits = req->rq_mbits - total_md + 1;
 341
 342         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 343                  "first mbits = x"LPU64", last mbits = x"LPU64"\n",
 344                  mbits, req->rq_mbits);
 345         LASSERTF(!(desc->bd_registered &&
 346                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 347                  mbits != desc->bd_last_mbits,
 348                  "registered: %d  rq_mbits: "LPU64" bd_last_mbits: "LPU64"\n",
 349                  desc->bd_registered, mbits, desc->bd_last_mbits);
 350
 351         desc->bd_registered = 1;
 352         desc->bd_last_mbits = mbits;
 353         desc->bd_md_count = total_md;
 354         md.user_ptr = &desc->bd_cbid;
 355         md.eq_handle = ptlrpc_eq_h;
 356         md.threshold = 1;                       /* PUT or GET */
 357
 358         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 359                 md.options = PTLRPC_MD_OPTIONS |
 360                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 361                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 362                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 363
 364                 rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 365                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 366                 if (rc != 0) {
 367                         CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
 368                                desc->bd_import->imp_obd->obd_name, mbits,
 369                                posted_md, rc);
 370                         break;
 371                 }
 372
 373                 /* About to let the network at it... */
 374                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 375                                   &desc->bd_mds[posted_md]);
 376                 if (rc != 0) {
 377                         CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
 378                                desc->bd_import->imp_obd->obd_name, mbits,
 379                                posted_md, rc);
 380                         rc2 = LNetMEUnlink(me_h);
 381                         LASSERT(rc2 == 0);
 382                         break;
 383                 }
 384         }
 385
 386         if (rc != 0) {
 387                 LASSERT(rc == -ENOMEM);
 388                 spin_lock(&desc->bd_lock);
 389                 desc->bd_md_count -= total_md - posted_md;
 390                 spin_unlock(&desc->bd_lock);
 391                 LASSERT(desc->bd_md_count >= 0);
 392                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 393                 req->rq_status = -ENOMEM;
 394                 RETURN(-ENOMEM);
 395         }
 396
 397         spin_lock(&desc->bd_lock);
 398         /* Holler if peer manages to touch buffers before he knows the mbits */
 399         if (desc->bd_md_count != total_md)
 400                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 401                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 402                       total_md - desc->bd_md_count);
 403         spin_unlock(&desc->bd_lock);
 404
 405         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 406                "mbits x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
 407                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 408                desc->bd_iov_count, desc->bd_nob,
 409                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 410
 411         RETURN(0);
 412 }
 413
 414 /**
 415  * Disconnect a bulk desc from the network. Idempotent. Not
 416  * thread-safe (i.e. only interlocks with completion callback).
 417  * Returns 1 on success or 0 if network unregistration failed for whatever
 418  * reason.
 419  */
 420 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 421 {
 422         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 423         struct l_wait_info       lwi;
 424         int                      rc;
 425         ENTRY;
 426
 427         LASSERT(!in_interrupt());     /* might sleep */
 428
 429         /* Let's setup deadline for reply unlink. */
 430         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 431             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 432                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 433
 434         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 435                 RETURN(1);                              /* never registered */
 436
 437         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 438
 439         /* the unlink ensures the callback happens ASAP and is the last
 440          * one.  If it fails, it must be because completion just happened,
 441          * but we must still l_wait_event() in this case to give liblustre
 442          * a chance to run client_bulk_callback() */
 443         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 444
 445         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 446                 RETURN(1);                              /* never registered */
 447
 448         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 449         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 450
 451         /* Do not wait for unlink to finish. */
 452         if (async)
 453                 RETURN(0);
 454
 455         for (;;) {
 456                 /* The wq argument is ignored by user-space wait_event macros */
 457                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 458                                         &req->rq_set->set_waitq :
 459                                         &req->rq_reply_waitq;
 460                 /* Network access will complete in finite time but the HUGE
 461                  * timeout lets us CWARN for visibility of sluggish NALs */
 462                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 463                                            cfs_time_seconds(1), NULL, NULL);
 464                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 465                 if (rc == 0) {
 466                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 467                         RETURN(1);
 468                 }
 469
 470                 LASSERT(rc == -ETIMEDOUT);
 471                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 472                           desc);
 473         }
 474         RETURN(0);
 475 }
 476
 477 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 478 {
 479         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 480         struct ptlrpc_service           *svc = svcpt->scp_service;
 481         int service_time = max_t(int, cfs_time_current_sec() -
 482                                  req->rq_arrival_time.tv_sec, 1);
 483
 484         if (!(flags & PTLRPC_REPLY_EARLY) &&
 485             (req->rq_type != PTL_RPC_MSG_ERR) &&
 486             (req->rq_reqmsg != NULL) &&
 487             !(lustre_msg_get_flags(req->rq_reqmsg) &
 488               (MSG_RESENT | MSG_REPLAY |
 489                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 490                 /* early replies, errors and recovery requests don't count
 491                  * toward our service time estimate */
 492                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 493
 494                 if (oldse != 0) {
 495                         DEBUG_REQ(D_ADAPTTO, req,
 496                                   "svc %s changed estimate from %d to %d",
 497                                   svc->srv_name, oldse,
 498                                   at_get(&svcpt->scp_at_estimate));
 499                 }
 500         }
 501         /* Report actual service time for client latency calc */
 502         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 503         /* Report service time estimate for future client reqs, but report 0
 504          * (to be ignored by client) if it's an error reply during recovery.
 505          * (bz15815) */
 506         if (req->rq_type == PTL_RPC_MSG_ERR &&
 507             (req->rq_export == NULL ||
 508              req->rq_export->exp_obd->obd_recovering)) {
 509                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 510         } else {
 511                 __u32 timeout;
 512
 513                 if (req->rq_export && req->rq_reqmsg != NULL &&
 514                     lustre_msg_get_flags(req->rq_reqmsg) &
 515                     (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))
 516                         timeout = cfs_time_current_sec() -
 517                                 req->rq_arrival_time.tv_sec +
 518                                 min(at_extra,
 519                                     req->rq_export->exp_obd->
 520                                     obd_recovery_timeout / 4);
 521                 else
 522                         timeout = at_get(&svcpt->scp_at_estimate);
 523                 lustre_msg_set_timeout(req->rq_repmsg, timeout);
 524         }
 525
 526         if (req->rq_reqmsg &&
 527             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 528                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 529                        "req_flags=%#x magic=%x/%x len=%d\n",
 530                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 531                        lustre_msg_get_magic(req->rq_reqmsg),
 532                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 533         }
 534 }
 535
 536 /**
 537  * Send request reply from request \a req reply buffer.
 538  * \a flags defines reply types
 539  * Returns 0 on success or error code
 540  */
 541 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 542 {
 543         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 544         struct ptlrpc_connection  *conn;
 545         int                        rc;
 546
 547         /* We must already have a reply buffer (only ptlrpc_error() may be
 548          * called without one). The reply generated by sptlrpc layer (e.g.
 549          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 550          * have a request buffer which is either the actual (swabbed) incoming
 551          * request, or a saved copy if this is a req saved in
 552          * target_queue_final_reply().
 553          */
 554         LASSERT (req->rq_no_reply == 0);
 555         LASSERT (req->rq_reqbuf != NULL);
 556         LASSERT (rs != NULL);
 557         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 558         LASSERT (req->rq_repmsg != NULL);
 559         LASSERT (req->rq_repmsg == rs->rs_msg);
 560         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 561         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 562
 563         /* There may be no rq_export during failover */
 564
 565         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 566                      req->rq_export->exp_obd->obd_fail)) {
 567                 /* Failed obd's only send ENODEV */
 568                 req->rq_type = PTL_RPC_MSG_ERR;
 569                 req->rq_status = -ENODEV;
 570                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 571                        req->rq_export->exp_obd->obd_minor);
 572         }
 573
 574         /* In order to keep interoprability with the client (< 2.3) which
 575          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 576          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 577          * reply buffer on client will be overflow.
 578          *
 579          * XXX Remove this whenver we drop the interoprability with such client.
 580          */
 581         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 582                                            sizeof(struct ptlrpc_body_v2), 1);
 583
 584         if (req->rq_type != PTL_RPC_MSG_ERR)
 585                 req->rq_type = PTL_RPC_MSG_REPLY;
 586
 587         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 588         lustre_msg_set_status(req->rq_repmsg,
 589                               ptlrpc_status_hton(req->rq_status));
 590         lustre_msg_set_opc(req->rq_repmsg,
 591                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 592
 593         target_pack_pool_reply(req);
 594
 595         ptlrpc_at_set_reply(req, flags);
 596
 597         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 598                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 599         else
 600                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 601
 602         if (unlikely(conn == NULL)) {
 603                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 604                 return -ENOTCONN;
 605         }
 606         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 607
 608         rc = sptlrpc_svc_wrap_reply(req);
 609         if (unlikely(rc))
 610                 goto out;
 611
 612         req->rq_sent = cfs_time_current_sec();
 613
 614         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 615                            (rs->rs_difficult && !rs->rs_no_ack) ?
 616                            LNET_ACK_REQ : LNET_NOACK_REQ,
 617                            &rs->rs_cb_id, conn,
 618                            ptlrpc_req2svc(req)->srv_rep_portal,
 619                            req->rq_xid, req->rq_reply_off);
 620 out:
 621         if (unlikely(rc != 0))
 622                 ptlrpc_req_drop_rs(req);
 623         ptlrpc_connection_put(conn);
 624         return rc;
 625 }
 626
 627 int ptlrpc_reply (struct ptlrpc_request *req)
 628 {
 629         if (req->rq_no_reply)
 630                 return 0;
 631         else
 632                 return (ptlrpc_send_reply(req, 0));
 633 }
 634
 635 /**
 636  * For request \a req send an error reply back. Create empty
 637  * reply buffers if necessary.
 638  */
 639 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 640 {
 641         int rc;
 642         ENTRY;
 643
 644         if (req->rq_no_reply)
 645                 RETURN(0);
 646
 647         if (!req->rq_repmsg) {
 648                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 649                 if (rc)
 650                         RETURN(rc);
 651         }
 652
 653         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 654             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 655             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 656                 req->rq_type = PTL_RPC_MSG_ERR;
 657
 658         rc = ptlrpc_send_reply(req, may_be_difficult);
 659         RETURN(rc);
 660 }
 661
 662 int ptlrpc_error(struct ptlrpc_request *req)
 663 {
 664         return ptlrpc_send_error(req, 0);
 665 }
 666
 667 /**
 668  * Send request \a request.
 669  * if \a noreply is set, don't expect any reply back and don't set up
 670  * reply buffers.
 671  * Returns 0 on success or error code.
 672  */
 673 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 674 {
 675         int rc;
 676         int rc2;
 677         int mpflag = 0;
 678         struct ptlrpc_connection *connection;
 679         lnet_handle_me_t  reply_me_h;
 680         lnet_md_t         reply_md;
 681         struct obd_import *imp = request->rq_import;
 682         struct obd_device *obd = imp->imp_obd;
 683         ENTRY;
 684
 685         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 686                 RETURN(0);
 687
 688         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 689         LASSERT(request->rq_wait_ctx == 0);
 690
 691         /* If this is a re-transmit, we're required to have disengaged
 692          * cleanly from the previous attempt */
 693         LASSERT(!request->rq_receiving_reply);
 694         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 695                 (imp->imp_state == LUSTRE_IMP_FULL)));
 696
 697         if (unlikely(obd != NULL && obd->obd_fail)) {
 698                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 699                         obd->obd_name);
 700                 /* this prevents us from waiting in ptlrpc_queue_wait */
 701                 spin_lock(&request->rq_lock);
 702                 request->rq_err = 1;
 703                 spin_unlock(&request->rq_lock);
 704                 request->rq_status = -ENODEV;
 705                 RETURN(-ENODEV);
 706         }
 707
 708         connection = imp->imp_connection;
 709
 710         lustre_msg_set_handle(request->rq_reqmsg,
 711                               &imp->imp_remote_handle);
 712         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 713         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 714                                 imp->imp_conn_cnt);
 715         lustre_msghdr_set_flags(request->rq_reqmsg,
 716                                 imp->imp_msghdr_flags);
 717
 718         /* If it's the first time to resend the request for EINPROGRESS,
 719          * we need to allocate a new XID (see after_reply()), it's different
 720          * from the resend for reply timeout. */
 721         if (request->rq_nr_resend != 0 &&
 722             list_empty(&request->rq_unreplied_list)) {
 723                 __u64 min_xid = 0;
 724                 /* resend for EINPROGRESS, allocate new xid to avoid reply
 725                  * reconstruction */
 726                 spin_lock(&imp->imp_lock);
 727                 ptlrpc_assign_next_xid_nolock(request);
 728                 min_xid = ptlrpc_known_replied_xid(imp);
 729                 spin_unlock(&imp->imp_lock);
 730
 731                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 732                 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
 733                           "resend on EINPROGRESS");
 734         }
 735
 736         if (request->rq_bulk != NULL) {
 737                 ptlrpc_set_bulk_mbits(request);
 738                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 739         }
 740
 741         if (list_empty(&request->rq_unreplied_list) ||
 742             request->rq_xid <= imp->imp_known_replied_xid) {
 743                 DEBUG_REQ(D_ERROR, request, "xid: "LPU64", replied: "LPU64", "
 744                           "list_empty:%d\n", request->rq_xid,
 745                           imp->imp_known_replied_xid,
 746                           list_empty(&request->rq_unreplied_list));
 747                 LBUG();
 748         }
 749
 750         /** For enabled AT all request should have AT_SUPPORT in the
 751          * FULL import state when OBD_CONNECT_AT is set */
 752         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 753                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 754                 !(imp->imp_connect_data.ocd_connect_flags &
 755                 OBD_CONNECT_AT));
 756
 757         if (request->rq_resend) {
 758                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 759                 if (request->rq_resend_cb != NULL)
 760                         request->rq_resend_cb(request, &request->rq_async_args);
 761         }
 762         if (request->rq_memalloc)
 763                 mpflag = cfs_memory_pressure_get_and_set();
 764
 765         rc = sptlrpc_cli_wrap_request(request);
 766         if (rc == -ENOMEM)
 767                 /* set rq_sent so that this request is treated
 768                  * as a delayed send in the upper layers */
 769                 request->rq_sent = cfs_time_current_sec();
 770         if (rc)
 771                 GOTO(out, rc);
 772
 773         /* bulk register should be done after wrap_request() */
 774         if (request->rq_bulk != NULL) {
 775                 rc = ptlrpc_register_bulk (request);
 776                 if (rc != 0)
 777                         GOTO(out, rc);
 778         }
 779
 780         if (!noreply) {
 781                 LASSERT (request->rq_replen != 0);
 782                 if (request->rq_repbuf == NULL) {
 783                         LASSERT(request->rq_repdata == NULL);
 784                         LASSERT(request->rq_repmsg == NULL);
 785                         rc = sptlrpc_cli_alloc_repbuf(request,
 786                                                       request->rq_replen);
 787                         if (rc) {
 788                                 /* this prevents us from looping in
 789                                  * ptlrpc_queue_wait */
 790                                 spin_lock(&request->rq_lock);
 791                                 request->rq_err = 1;
 792                                 spin_unlock(&request->rq_lock);
 793                                 request->rq_status = rc;
 794                                 GOTO(cleanup_bulk, rc);
 795                         }
 796                 } else {
 797                         request->rq_repdata = NULL;
 798                         request->rq_repmsg = NULL;
 799                 }
 800
 801                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 802                                   connection->c_peer, request->rq_xid, 0,
 803                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 804                 if (rc != 0) {
 805                         CERROR("LNetMEAttach failed: %d\n", rc);
 806                         LASSERT (rc == -ENOMEM);
 807                         GOTO(cleanup_bulk, rc = -ENOMEM);
 808                 }
 809         }
 810
 811         spin_lock(&request->rq_lock);
 812         /* We are responsible for unlinking the reply buffer */
 813         request->rq_reply_unlinked = noreply;
 814         request->rq_receiving_reply = !noreply;
 815         /* Clear any flags that may be present from previous sends. */
 816         request->rq_req_unlinked = 0;
 817         request->rq_replied = 0;
 818         request->rq_err = 0;
 819         request->rq_timedout = 0;
 820         request->rq_net_err = 0;
 821         request->rq_resend = 0;
 822         request->rq_restart = 0;
 823         request->rq_reply_truncated = 0;
 824         spin_unlock(&request->rq_lock);
 825
 826         if (!noreply) {
 827                 reply_md.start     = request->rq_repbuf;
 828                 reply_md.length    = request->rq_repbuf_len;
 829                 /* Allow multiple early replies */
 830                 reply_md.threshold = LNET_MD_THRESH_INF;
 831                 /* Manage remote for early replies */
 832                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 833                         LNET_MD_MANAGE_REMOTE |
 834                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 835                 reply_md.user_ptr  = &request->rq_reply_cbid;
 836                 reply_md.eq_handle = ptlrpc_eq_h;
 837
 838                 /* We must see the unlink callback to set rq_reply_unlinked,
 839                  * so we can't auto-unlink */
 840                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 841                                   &request->rq_reply_md_h);
 842                 if (rc != 0) {
 843                         CERROR("LNetMDAttach failed: %d\n", rc);
 844                         LASSERT (rc == -ENOMEM);
 845                         spin_lock(&request->rq_lock);
 846                         /* ...but the MD attach didn't succeed... */
 847                         request->rq_receiving_reply = 0;
 848                         spin_unlock(&request->rq_lock);
 849                         GOTO(cleanup_me, rc = -ENOMEM);
 850                 }
 851
 852                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 853                        ", portal %u\n",
 854                        request->rq_repbuf_len, request->rq_xid,
 855                        request->rq_reply_portal);
 856         }
 857
 858         /* add references on request for request_out_callback */
 859         ptlrpc_request_addref(request);
 860         if (obd != NULL && obd->obd_svc_stats != NULL)
 861                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 862                         atomic_read(&imp->imp_inflight));
 863
 864         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 865
 866         do_gettimeofday(&request->rq_sent_tv);
 867         request->rq_sent = cfs_time_current_sec();
 868         /* We give the server rq_timeout secs to process the req, and
 869            add the network latency for our local timeout. */
 870         request->rq_deadline = request->rq_sent + request->rq_timeout +
 871                 ptlrpc_at_get_net_latency(request);
 872
 873         ptlrpc_pinger_sending_on_import(imp);
 874
 875         DEBUG_REQ(D_INFO, request, "send flg=%x",
 876                   lustre_msg_get_flags(request->rq_reqmsg));
 877         rc = ptl_send_buf(&request->rq_req_md_h,
 878                           request->rq_reqbuf, request->rq_reqdata_len,
 879                           LNET_NOACK_REQ, &request->rq_req_cbid,
 880                           connection,
 881                           request->rq_request_portal,
 882                           request->rq_xid, 0);
 883         if (likely(rc == 0))
 884                 GOTO(out, rc);
 885
 886         request->rq_req_unlinked = 1;
 887         ptlrpc_req_finished(request);
 888         if (noreply)
 889                 GOTO(out, rc);
 890
 891  cleanup_me:
 892         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 893          * nobody apart from the PUT's target has the right nid+XID to
 894          * access the reply buffer. */
 895         rc2 = LNetMEUnlink(reply_me_h);
 896         LASSERT (rc2 == 0);
 897         /* UNLINKED callback called synchronously */
 898         LASSERT(!request->rq_receiving_reply);
 899
 900  cleanup_bulk:
 901         /* We do sync unlink here as there was no real transfer here so
 902          * the chance to have long unlink to sluggish net is smaller here. */
 903         ptlrpc_unregister_bulk(request, 0);
 904  out:
 905         if (request->rq_memalloc)
 906                 cfs_memory_pressure_restore(mpflag);
 907         return rc;
 908 }
 909 EXPORT_SYMBOL(ptl_send_rpc);
 910
 911 /**
 912  * Register request buffer descriptor for request receiving.
 913  */
 914 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 915 {
 916         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 917         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 918         int                       rc;
 919         lnet_md_t                 md;
 920         lnet_handle_me_t          me_h;
 921
 922         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 923                service->srv_req_portal);
 924
 925         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 926                 return (-ENOMEM);
 927
 928         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 929          * which means buffer can only be attached on local CPT, and LND
 930          * threads can find it by grabbing a local lock */
 931         rc = LNetMEAttach(service->srv_req_portal,
 932                           match_id, 0, ~0, LNET_UNLINK,
 933                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 934                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 935         if (rc != 0) {
 936                 CERROR("LNetMEAttach failed: %d\n", rc);
 937                 return (-ENOMEM);
 938         }
 939
 940         LASSERT(rqbd->rqbd_refcount == 0);
 941         rqbd->rqbd_refcount = 1;
 942
 943         md.start     = rqbd->rqbd_buffer;
 944         md.length    = service->srv_buf_size;
 945         md.max_size  = service->srv_max_req_size;
 946         md.threshold = LNET_MD_THRESH_INF;
 947         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 948         md.user_ptr  = &rqbd->rqbd_cbid;
 949         md.eq_handle = ptlrpc_eq_h;
 950
 951         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 952         if (rc == 0)
 953                 return (0);
 954
 955         CERROR("LNetMDAttach failed: %d; \n", rc);
 956         LASSERT (rc == -ENOMEM);
 957         rc = LNetMEUnlink (me_h);
 958         LASSERT (rc == 0);
 959         rqbd->rqbd_refcount = 0;
 960
 961         return (-ENOMEM);
 962 }