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, 2014, 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 npages * 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 npages, unsigned max_brw,
 120                                               unsigned type, unsigned portal)
 121 {
 122         struct obd_export *exp = req->rq_export;
 123         struct ptlrpc_bulk_desc *desc;
 124
 125         ENTRY;
 126         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 127
 128         desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
 129         if (desc == NULL)
 130                 RETURN(NULL);
 131
 132         desc->bd_export = class_export_get(exp);
 133         desc->bd_req = req;
 134
 135         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 136         desc->bd_cbid.cbid_arg = desc;
 137
 138         /* NB we don't assign rq_bulk here; server-side requests are
 139          * re-used, and the handler frees the bulk desc explicitly. */
 140
 141         return desc;
 142 }
 143 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
 144
 145 /**
 146  * Starts bulk transfer for descriptor \a desc on the server.
 147  * Returns 0 on success or error code.
 148  */
 149 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 150 {
 151         struct obd_export        *exp = desc->bd_export;
 152         struct ptlrpc_connection *conn = exp->exp_connection;
 153         int                       rc = 0;
 154         __u64                     xid;
 155         int                       posted_md;
 156         int                       total_md;
 157         lnet_md_t                 md;
 158         ENTRY;
 159
 160         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 161                 RETURN(0);
 162
 163         /* NB no locking required until desc is on the network */
 164         LASSERT(desc->bd_md_count == 0);
 165         LASSERT(desc->bd_type == BULK_PUT_SOURCE ||
 166                 desc->bd_type == BULK_GET_SINK);
 167
 168         LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
 169         LASSERT(desc->bd_cbid.cbid_arg == desc);
 170
 171         /* NB total length may be 0 for a read past EOF, so we send 0
 172          * length bulks, since the client expects bulk events.
 173          *
 174          * The client may not need all of the bulk XIDs for the RPC.  The RPC
 175          * used the XID of the highest bulk XID needed, and the server masks
 176          * off high bits to get bulk count for this RPC. LU-1431 */
 177         xid = desc->bd_req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 178         total_md = desc->bd_req->rq_xid - xid + 1;
 179
 180         desc->bd_md_count = total_md;
 181         desc->bd_failure = 0;
 182
 183         md.user_ptr = &desc->bd_cbid;
 184         md.eq_handle = ptlrpc_eq_h;
 185         md.threshold = 2; /* SENT and ACK/REPLY */
 186
 187         for (posted_md = 0; posted_md < total_md; xid++) {
 188                 md.options = PTLRPC_MD_OPTIONS;
 189
 190                 /* NB it's assumed that source and sink buffer frags are
 191                  * page-aligned. Otherwise we'd have to send client bulk
 192                  * sizes over and split server buffer accordingly */
 193                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 194                 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
 195                 if (rc != 0) {
 196                         CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
 197                                exp->exp_obd->obd_name, posted_md, rc);
 198                         LASSERT(rc == -ENOMEM);
 199                         if (posted_md == 0) {
 200                                 desc->bd_md_count = 0;
 201                                 RETURN(-ENOMEM);
 202                         }
 203                         break;
 204                 }
 205                 /* Network is about to get at the memory */
 206                 if (desc->bd_type == BULK_PUT_SOURCE)
 207                         rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
 208                                      LNET_ACK_REQ, conn->c_peer,
 209                                      desc->bd_portal, xid, 0, 0);
 210                 else
 211                         rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
 212                                      conn->c_peer, desc->bd_portal, xid, 0);
 213
 214                 posted_md++;
 215                 if (rc != 0) {
 216                         CERROR("%s: failed bulk transfer with %s:%u x"LPU64": "
 217                                "rc = %d\n", exp->exp_obd->obd_name,
 218                                libcfs_id2str(conn->c_peer), desc->bd_portal,
 219                                xid, rc);
 220                         break;
 221                 }
 222         }
 223
 224         if (rc != 0) {
 225                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 226                  * event this creates will signal completion with failure,
 227                  * so we return SUCCESS here! */
 228                 spin_lock(&desc->bd_lock);
 229                 desc->bd_md_count -= total_md - posted_md;
 230                 spin_unlock(&desc->bd_lock);
 231                 LASSERT(desc->bd_md_count >= 0);
 232
 233                 mdunlink_iterate_helper(desc->bd_mds, posted_md);
 234                 RETURN(0);
 235         }
 236
 237         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 238                "id %s xid "LPX64"-"LPX64"\n", desc->bd_iov_count,
 239                desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
 240                xid - posted_md, xid - 1);
 241
 242         RETURN(0);
 243 }
 244
 245 /**
 246  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 247  * serialises with completion callback)
 248  */
 249 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 250 {
 251         struct l_wait_info       lwi;
 252         int                      rc;
 253
 254         LASSERT(!in_interrupt());           /* might sleep */
 255
 256         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 257                 return;                         /* never started */
 258
 259         /* We used to poison the pages with 0xab here because we did not want to
 260          * send any meaningful data over the wire for evicted clients (bug 9297)
 261          * However, this is no longer safe now that we use the page cache on the
 262          * OSS (bug 20560) */
 263
 264         /* The unlink ensures the callback happens ASAP and is the last
 265          * one.  If it fails, it must be because completion just happened,
 266          * but we must still l_wait_event() in this case, to give liblustre
 267          * a chance to run server_bulk_callback()*/
 268         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_count);
 269
 270         for (;;) {
 271                 /* Network access will complete in finite time but the HUGE
 272                  * timeout lets us CWARN for visibility of sluggish NALs */
 273                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 274                                            cfs_time_seconds(1), NULL, NULL);
 275                 rc = l_wait_event(desc->bd_waitq,
 276                                   !ptlrpc_server_bulk_active(desc), &lwi);
 277                 if (rc == 0)
 278                         return;
 279
 280                 LASSERT(rc == -ETIMEDOUT);
 281                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 282         }
 283 }
 284 #endif /* HAVE_SERVER_SUPPORT */
 285
 286 /**
 287  * Register bulk at the sender for later transfer.
 288  * Returns 0 on success or error code.
 289  */
 290 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 291 {
 292         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 293         lnet_process_id_t peer;
 294         int rc = 0;
 295         int rc2;
 296         int posted_md;
 297         int total_md;
 298         __u64 xid;
 299         lnet_handle_me_t  me_h;
 300         lnet_md_t         md;
 301         ENTRY;
 302
 303         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 304                 RETURN(0);
 305
 306         /* NB no locking required until desc is on the network */
 307         LASSERT(desc->bd_nob > 0);
 308         LASSERT(desc->bd_md_count == 0);
 309         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 310         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 311         LASSERT(desc->bd_req != NULL);
 312         LASSERT(desc->bd_type == BULK_PUT_SINK ||
 313                 desc->bd_type == BULK_GET_SOURCE);
 314
 315         /* cleanup the state of the bulk for it will be reused */
 316         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 317                 desc->bd_nob_transferred = 0;
 318         else
 319                 LASSERT(desc->bd_nob_transferred == 0);
 320
 321         desc->bd_failure = 0;
 322
 323         peer = desc->bd_import->imp_connection->c_peer;
 324
 325         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 326         LASSERT(desc->bd_cbid.cbid_arg == desc);
 327
 328         /* An XID is only used for a single request from the client.
 329          * For retried bulk transfers, a new XID will be allocated in
 330          * in ptlrpc_check_set() if it needs to be resent, so it is not
 331          * using the same RDMA match bits after an error.
 332          *
 333          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 334          * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
 335         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 336         LASSERTF(!(desc->bd_registered &&
 337                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 338                  xid != desc->bd_last_xid,
 339                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 340                  desc->bd_registered, xid, desc->bd_last_xid);
 341
 342         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 343         desc->bd_registered = 1;
 344         desc->bd_last_xid = xid;
 345         desc->bd_md_count = total_md;
 346         md.user_ptr = &desc->bd_cbid;
 347         md.eq_handle = ptlrpc_eq_h;
 348         md.threshold = 1;                       /* PUT or GET */
 349
 350         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 351                 md.options = PTLRPC_MD_OPTIONS |
 352                              ((desc->bd_type == BULK_GET_SOURCE) ?
 353                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 354                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 355
 356                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 357                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 358                 if (rc != 0) {
 359                         CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
 360                                desc->bd_import->imp_obd->obd_name, xid,
 361                                posted_md, rc);
 362                         break;
 363                 }
 364
 365                 /* About to let the network at it... */
 366                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 367                                   &desc->bd_mds[posted_md]);
 368                 if (rc != 0) {
 369                         CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
 370                                desc->bd_import->imp_obd->obd_name, xid,
 371                                posted_md, rc);
 372                         rc2 = LNetMEUnlink(me_h);
 373                         LASSERT(rc2 == 0);
 374                         break;
 375                 }
 376         }
 377
 378         if (rc != 0) {
 379                 LASSERT(rc == -ENOMEM);
 380                 spin_lock(&desc->bd_lock);
 381                 desc->bd_md_count -= total_md - posted_md;
 382                 spin_unlock(&desc->bd_lock);
 383                 LASSERT(desc->bd_md_count >= 0);
 384                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 385                 req->rq_status = -ENOMEM;
 386                 RETURN(-ENOMEM);
 387         }
 388
 389         /* Set rq_xid to matchbits of the final bulk so that server can
 390          * infer the number of bulks that were prepared */
 391         req->rq_xid = --xid;
 392         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 393                  "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n",
 394                  desc->bd_last_xid, req->rq_xid);
 395
 396         spin_lock(&desc->bd_lock);
 397         /* Holler if peer manages to touch buffers before he knows the xid */
 398         if (desc->bd_md_count != total_md)
 399                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 400                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 401                       total_md - desc->bd_md_count);
 402         spin_unlock(&desc->bd_lock);
 403
 404         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 405                "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
 406                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 407                desc->bd_iov_count, desc->bd_nob,
 408                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 409
 410         RETURN(0);
 411 }
 412
 413 /**
 414  * Disconnect a bulk desc from the network. Idempotent. Not
 415  * thread-safe (i.e. only interlocks with completion callback).
 416  * Returns 1 on success or 0 if network unregistration failed for whatever
 417  * reason.
 418  */
 419 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 420 {
 421         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 422         struct l_wait_info       lwi;
 423         int                      rc;
 424         ENTRY;
 425
 426         LASSERT(!in_interrupt());     /* might sleep */
 427
 428         /* Let's setup deadline for reply unlink. */
 429         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 430             async && req->rq_bulk_deadline == 0)
 431                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 432
 433         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 434                 RETURN(1);                              /* never registered */
 435
 436         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 437
 438         /* the unlink ensures the callback happens ASAP and is the last
 439          * one.  If it fails, it must be because completion just happened,
 440          * but we must still l_wait_event() in this case to give liblustre
 441          * a chance to run client_bulk_callback() */
 442         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 443
 444         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 445                 RETURN(1);                              /* never registered */
 446
 447         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 448         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 449
 450         /* Do not wait for unlink to finish. */
 451         if (async)
 452                 RETURN(0);
 453
 454         for (;;) {
 455                 /* The wq argument is ignored by user-space wait_event macros */
 456                 wait_queue_head_t *wq = (req->rq_set != NULL) ?
 457                                         &req->rq_set->set_waitq :
 458                                         &req->rq_reply_waitq;
 459                 /* Network access will complete in finite time but the HUGE
 460                  * timeout lets us CWARN for visibility of sluggish NALs */
 461                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 462                                            cfs_time_seconds(1), NULL, NULL);
 463                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 464                 if (rc == 0) {
 465                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 466                         RETURN(1);
 467                 }
 468
 469                 LASSERT(rc == -ETIMEDOUT);
 470                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 471                           desc);
 472         }
 473         RETURN(0);
 474 }
 475
 476 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 477 {
 478         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 479         struct ptlrpc_service           *svc = svcpt->scp_service;
 480         int service_time = max_t(int, cfs_time_current_sec() -
 481                                  req->rq_arrival_time.tv_sec, 1);
 482
 483         if (!(flags & PTLRPC_REPLY_EARLY) &&
 484             (req->rq_type != PTL_RPC_MSG_ERR) &&
 485             (req->rq_reqmsg != NULL) &&
 486             !(lustre_msg_get_flags(req->rq_reqmsg) &
 487               (MSG_RESENT | MSG_REPLAY |
 488                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 489                 /* early replies, errors and recovery requests don't count
 490                  * toward our service time estimate */
 491                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 492
 493                 if (oldse != 0) {
 494                         DEBUG_REQ(D_ADAPTTO, req,
 495                                   "svc %s changed estimate from %d to %d",
 496                                   svc->srv_name, oldse,
 497                                   at_get(&svcpt->scp_at_estimate));
 498                 }
 499         }
 500         /* Report actual service time for client latency calc */
 501         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 502         /* Report service time estimate for future client reqs, but report 0
 503          * (to be ignored by client) if it's an error reply during recovery.
 504          * (bz15815) */
 505         if (req->rq_type == PTL_RPC_MSG_ERR &&
 506             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 507                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 508         else
 509                 lustre_msg_set_timeout(req->rq_repmsg,
 510                                        at_get(&svcpt->scp_at_estimate));
 511
 512         if (req->rq_reqmsg &&
 513             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 514                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 515                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 516                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 517                        lustre_msg_is_v1(req->rq_reqmsg),
 518                        lustre_msg_get_magic(req->rq_reqmsg),
 519                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 520         }
 521 }
 522
 523 /**
 524  * Send request reply from request \a req reply buffer.
 525  * \a flags defines reply types
 526  * Returns 0 on success or error code
 527  */
 528 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 529 {
 530         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 531         struct ptlrpc_connection  *conn;
 532         int                        rc;
 533
 534         /* We must already have a reply buffer (only ptlrpc_error() may be
 535          * called without one). The reply generated by sptlrpc layer (e.g.
 536          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 537          * have a request buffer which is either the actual (swabbed) incoming
 538          * request, or a saved copy if this is a req saved in
 539          * target_queue_final_reply().
 540          */
 541         LASSERT (req->rq_no_reply == 0);
 542         LASSERT (req->rq_reqbuf != NULL);
 543         LASSERT (rs != NULL);
 544         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 545         LASSERT (req->rq_repmsg != NULL);
 546         LASSERT (req->rq_repmsg == rs->rs_msg);
 547         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 548         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 549
 550         /* There may be no rq_export during failover */
 551
 552         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 553                      req->rq_export->exp_obd->obd_fail)) {
 554                 /* Failed obd's only send ENODEV */
 555                 req->rq_type = PTL_RPC_MSG_ERR;
 556                 req->rq_status = -ENODEV;
 557                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 558                        req->rq_export->exp_obd->obd_minor);
 559         }
 560
 561         /* In order to keep interoprability with the client (< 2.3) which
 562          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 563          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 564          * reply buffer on client will be overflow.
 565          *
 566          * XXX Remove this whenver we drop the interoprability with such client.
 567          */
 568         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 569                                            sizeof(struct ptlrpc_body_v2), 1);
 570
 571         if (req->rq_type != PTL_RPC_MSG_ERR)
 572                 req->rq_type = PTL_RPC_MSG_REPLY;
 573
 574         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 575         lustre_msg_set_status(req->rq_repmsg,
 576                               ptlrpc_status_hton(req->rq_status));
 577         lustre_msg_set_opc(req->rq_repmsg,
 578                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 579
 580         target_pack_pool_reply(req);
 581
 582         ptlrpc_at_set_reply(req, flags);
 583
 584         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 585                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 586         else
 587                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 588
 589         if (unlikely(conn == NULL)) {
 590                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 591                 return -ENOTCONN;
 592         }
 593         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 594
 595         rc = sptlrpc_svc_wrap_reply(req);
 596         if (unlikely(rc))
 597                 goto out;
 598
 599         req->rq_sent = cfs_time_current_sec();
 600
 601         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 602                            (rs->rs_difficult && !rs->rs_no_ack) ?
 603                            LNET_ACK_REQ : LNET_NOACK_REQ,
 604                            &rs->rs_cb_id, conn,
 605                            ptlrpc_req2svc(req)->srv_rep_portal,
 606                            req->rq_xid, req->rq_reply_off);
 607 out:
 608         if (unlikely(rc != 0))
 609                 ptlrpc_req_drop_rs(req);
 610         ptlrpc_connection_put(conn);
 611         return rc;
 612 }
 613
 614 int ptlrpc_reply (struct ptlrpc_request *req)
 615 {
 616         if (req->rq_no_reply)
 617                 return 0;
 618         else
 619                 return (ptlrpc_send_reply(req, 0));
 620 }
 621
 622 /**
 623  * For request \a req send an error reply back. Create empty
 624  * reply buffers if necessary.
 625  */
 626 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 627 {
 628         int rc;
 629         ENTRY;
 630
 631         if (req->rq_no_reply)
 632                 RETURN(0);
 633
 634         if (!req->rq_repmsg) {
 635                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 636                 if (rc)
 637                         RETURN(rc);
 638         }
 639
 640         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 641             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 642             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 643                 req->rq_type = PTL_RPC_MSG_ERR;
 644
 645         rc = ptlrpc_send_reply(req, may_be_difficult);
 646         RETURN(rc);
 647 }
 648
 649 int ptlrpc_error(struct ptlrpc_request *req)
 650 {
 651         return ptlrpc_send_error(req, 0);
 652 }
 653
 654 /**
 655  * Send request \a request.
 656  * if \a noreply is set, don't expect any reply back and don't set up
 657  * reply buffers.
 658  * Returns 0 on success or error code.
 659  */
 660 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 661 {
 662         int rc;
 663         int rc2;
 664         int mpflag = 0;
 665         struct ptlrpc_connection *connection;
 666         lnet_handle_me_t  reply_me_h;
 667         lnet_md_t         reply_md;
 668         struct obd_import *imp = request->rq_import;
 669         struct obd_device *obd = imp->imp_obd;
 670         ENTRY;
 671
 672         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 673                 RETURN(0);
 674
 675         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 676         LASSERT(request->rq_wait_ctx == 0);
 677
 678         /* If this is a re-transmit, we're required to have disengaged
 679          * cleanly from the previous attempt */
 680         LASSERT(!request->rq_receiving_reply);
 681         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 682                 (imp->imp_state == LUSTRE_IMP_FULL)));
 683
 684         if (unlikely(obd != NULL && obd->obd_fail)) {
 685                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 686                         obd->obd_name);
 687                 /* this prevents us from waiting in ptlrpc_queue_wait */
 688                 spin_lock(&request->rq_lock);
 689                 request->rq_err = 1;
 690                 spin_unlock(&request->rq_lock);
 691                 request->rq_status = -ENODEV;
 692                 RETURN(-ENODEV);
 693         }
 694
 695         connection = imp->imp_connection;
 696
 697         lustre_msg_set_handle(request->rq_reqmsg,
 698                               &imp->imp_remote_handle);
 699         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 700         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 701                                 imp->imp_conn_cnt);
 702         lustre_msghdr_set_flags(request->rq_reqmsg,
 703                                 imp->imp_msghdr_flags);
 704
 705         /** For enabled AT all request should have AT_SUPPORT in the
 706          * FULL import state when OBD_CONNECT_AT is set */
 707         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 708                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 709                 !(imp->imp_connect_data.ocd_connect_flags &
 710                 OBD_CONNECT_AT));
 711
 712         if (request->rq_resend) {
 713                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 714                 if (request->rq_resend_cb != NULL)
 715                         request->rq_resend_cb(request, &request->rq_async_args);
 716         }
 717         if (request->rq_memalloc)
 718                 mpflag = cfs_memory_pressure_get_and_set();
 719
 720         rc = sptlrpc_cli_wrap_request(request);
 721         if (rc)
 722                 GOTO(out, rc);
 723
 724         /* bulk register should be done after wrap_request() */
 725         if (request->rq_bulk != NULL) {
 726                 rc = ptlrpc_register_bulk (request);
 727                 if (rc != 0)
 728                         GOTO(out, rc);
 729         }
 730
 731         if (!noreply) {
 732                 LASSERT (request->rq_replen != 0);
 733                 if (request->rq_repbuf == NULL) {
 734                         LASSERT(request->rq_repdata == NULL);
 735                         LASSERT(request->rq_repmsg == NULL);
 736                         rc = sptlrpc_cli_alloc_repbuf(request,
 737                                                       request->rq_replen);
 738                         if (rc) {
 739                                 /* this prevents us from looping in
 740                                  * ptlrpc_queue_wait */
 741                                 spin_lock(&request->rq_lock);
 742                                 request->rq_err = 1;
 743                                 spin_unlock(&request->rq_lock);
 744                                 request->rq_status = rc;
 745                                 GOTO(cleanup_bulk, rc);
 746                         }
 747                 } else {
 748                         request->rq_repdata = NULL;
 749                         request->rq_repmsg = NULL;
 750                 }
 751
 752                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 753                                   connection->c_peer, request->rq_xid, 0,
 754                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 755                 if (rc != 0) {
 756                         CERROR("LNetMEAttach failed: %d\n", rc);
 757                         LASSERT (rc == -ENOMEM);
 758                         GOTO(cleanup_bulk, rc = -ENOMEM);
 759                 }
 760         }
 761
 762         spin_lock(&request->rq_lock);
 763         /* We are responsible for unlinking the reply buffer */
 764         request->rq_reply_unlinked = noreply;
 765         request->rq_receiving_reply = !noreply;
 766         /* Clear any flags that may be present from previous sends. */
 767         request->rq_req_unlinked = 0;
 768         request->rq_replied = 0;
 769         request->rq_err = 0;
 770         request->rq_timedout = 0;
 771         request->rq_net_err = 0;
 772         request->rq_resend = 0;
 773         request->rq_restart = 0;
 774         request->rq_reply_truncated = 0;
 775         spin_unlock(&request->rq_lock);
 776
 777         if (!noreply) {
 778                 reply_md.start     = request->rq_repbuf;
 779                 reply_md.length    = request->rq_repbuf_len;
 780                 /* Allow multiple early replies */
 781                 reply_md.threshold = LNET_MD_THRESH_INF;
 782                 /* Manage remote for early replies */
 783                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 784                         LNET_MD_MANAGE_REMOTE |
 785                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 786                 reply_md.user_ptr  = &request->rq_reply_cbid;
 787                 reply_md.eq_handle = ptlrpc_eq_h;
 788
 789                 /* We must see the unlink callback to set rq_reply_unlinked,
 790                  * so we can't auto-unlink */
 791                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 792                                   &request->rq_reply_md_h);
 793                 if (rc != 0) {
 794                         CERROR("LNetMDAttach failed: %d\n", rc);
 795                         LASSERT (rc == -ENOMEM);
 796                         spin_lock(&request->rq_lock);
 797                         /* ...but the MD attach didn't succeed... */
 798                         request->rq_receiving_reply = 0;
 799                         spin_unlock(&request->rq_lock);
 800                         GOTO(cleanup_me, rc = -ENOMEM);
 801                 }
 802
 803                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 804                        ", portal %u\n",
 805                        request->rq_repbuf_len, request->rq_xid,
 806                        request->rq_reply_portal);
 807         }
 808
 809         /* add references on request for request_out_callback */
 810         ptlrpc_request_addref(request);
 811         if (obd != NULL && obd->obd_svc_stats != NULL)
 812                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 813                         atomic_read(&imp->imp_inflight));
 814
 815         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 816
 817         do_gettimeofday(&request->rq_sent_tv);
 818         request->rq_sent = cfs_time_current_sec();
 819         /* We give the server rq_timeout secs to process the req, and
 820            add the network latency for our local timeout. */
 821         request->rq_deadline = request->rq_sent + request->rq_timeout +
 822                 ptlrpc_at_get_net_latency(request);
 823
 824         ptlrpc_pinger_sending_on_import(imp);
 825
 826         DEBUG_REQ(D_INFO, request, "send flg=%x",
 827                   lustre_msg_get_flags(request->rq_reqmsg));
 828         rc = ptl_send_buf(&request->rq_req_md_h,
 829                           request->rq_reqbuf, request->rq_reqdata_len,
 830                           LNET_NOACK_REQ, &request->rq_req_cbid,
 831                           connection,
 832                           request->rq_request_portal,
 833                           request->rq_xid, 0);
 834         if (likely(rc == 0))
 835                 GOTO(out, rc);
 836
 837         request->rq_req_unlinked = 1;
 838         ptlrpc_req_finished(request);
 839         if (noreply)
 840                 GOTO(out, rc);
 841
 842  cleanup_me:
 843         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 844          * nobody apart from the PUT's target has the right nid+XID to
 845          * access the reply buffer. */
 846         rc2 = LNetMEUnlink(reply_me_h);
 847         LASSERT (rc2 == 0);
 848         /* UNLINKED callback called synchronously */
 849         LASSERT(!request->rq_receiving_reply);
 850
 851  cleanup_bulk:
 852         /* We do sync unlink here as there was no real transfer here so
 853          * the chance to have long unlink to sluggish net is smaller here. */
 854         ptlrpc_unregister_bulk(request, 0);
 855  out:
 856         if (request->rq_memalloc)
 857                 cfs_memory_pressure_restore(mpflag);
 858         return rc;
 859 }
 860 EXPORT_SYMBOL(ptl_send_rpc);
 861
 862 /**
 863  * Register request buffer descriptor for request receiving.
 864  */
 865 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 866 {
 867         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 868         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 869         int                       rc;
 870         lnet_md_t                 md;
 871         lnet_handle_me_t          me_h;
 872
 873         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 874                service->srv_req_portal);
 875
 876         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 877                 return (-ENOMEM);
 878
 879         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 880          * which means buffer can only be attached on local CPT, and LND
 881          * threads can find it by grabbing a local lock */
 882         rc = LNetMEAttach(service->srv_req_portal,
 883                           match_id, 0, ~0, LNET_UNLINK,
 884                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 885                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 886         if (rc != 0) {
 887                 CERROR("LNetMEAttach failed: %d\n", rc);
 888                 return (-ENOMEM);
 889         }
 890
 891         LASSERT(rqbd->rqbd_refcount == 0);
 892         rqbd->rqbd_refcount = 1;
 893
 894         md.start     = rqbd->rqbd_buffer;
 895         md.length    = service->srv_buf_size;
 896         md.max_size  = service->srv_max_req_size;
 897         md.threshold = LNET_MD_THRESH_INF;
 898         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 899         md.user_ptr  = &rqbd->rqbd_cbid;
 900         md.eq_handle = ptlrpc_eq_h;
 901
 902         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 903         if (rc == 0)
 904                 return (0);
 905
 906         CERROR("LNetMDAttach failed: %d; \n", rc);
 907         LASSERT (rc == -ENOMEM);
 908         rc = LNetMEUnlink (me_h);
 909         LASSERT (rc == 0);
 910         rqbd->rqbd_refcount = 0;
 911
 912         return (-ENOMEM);
 913 }