lustre/target/tgt_handler.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, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2013, 2017, Intel Corporation.
  25  */
  26 /*
  27  * lustre/target/tgt_handler.c
  28  *
  29  * Lustre Unified Target request handler code
  30  *
  31  * Author: Brian Behlendorf <behlendorf1@llnl.gov>
  32  * Author: Mikhail Pershin <mike.pershin@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_CLASS
  36
  37 #include <linux/user_namespace.h>
  38 #include <linux/delay.h>
  39 #include <linux/uidgid.h>
  40
  41 #include <libcfs/linux/linux-mem.h>
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47 #include <lustre_acl.h>
  48
  49 #include "tgt_internal.h"
  50
  51 char *tgt_name(struct lu_target *tgt)
  52 {
  53         LASSERT(tgt->lut_obd != NULL);
  54         return tgt->lut_obd->obd_name;
  55 }
  56 EXPORT_SYMBOL(tgt_name);
  57
  58 /*
  59  * Generic code handling requests that have struct mdt_body passed in:
  60  *
  61  *  - extract mdt_body from request and save it in @tsi, if present;
  62  *
  63  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  64  *  @tsi;
  65  *
  66  *  - if HAS_BODY flag is set for this request type check whether object
  67  *  actually exists on storage (lu_object_exists()).
  68  *
  69  */
  70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  71 {
  72         const struct mdt_body   *body;
  73         struct lu_object        *obj;
  74         struct req_capsule      *pill = tsi->tsi_pill;
  75         int                      rc;
  76
  77         ENTRY;
  78
  79         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EFAULT);
  82
  83         tsi->tsi_mdt_body = body;
  84
  85         if (!(body->mbo_valid & OBD_MD_FLID))
  86                 RETURN(0);
  87
  88         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  89          * in any case in pre-2.5 clients. Fix that here if needed */
  90         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  91                 RETURN(0);
  92
  93         if (!fid_is_sane(&body->mbo_fid1)) {
  94                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  95                        PFID(&body->mbo_fid1));
  96                 RETURN(-EINVAL);
  97         }
  98
  99         obj = lu_object_find(tsi->tsi_env,
 100                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 101                              &body->mbo_fid1, NULL);
 102         if (!IS_ERR(obj)) {
 103                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 104                         lu_object_put(tsi->tsi_env, obj);
 105                         rc = -ENOENT;
 106                 } else {
 107                         tsi->tsi_corpus = obj;
 108                         rc = 0;
 109                 }
 110         } else {
 111                 rc = PTR_ERR(obj);
 112         }
 113
 114         tsi->tsi_fid = body->mbo_fid1;
 115
 116         RETURN(rc);
 117 }
 118
 119 /**
 120  * Validate oa from client.
 121  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 122  * req are valid.
 123  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 124  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 125  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 126  *       whether this is oi_fid or real ostid. So it will check
 127  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 128  *    c. Old FID-disable osc will send IDIF.
 129  *    d. new FID-enable osc/osp will send normal FID.
 130  *
 131  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 132  * be used for LAST_ID file, and only being accessed inside OST now.
 133  */
 134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 135 {
 136         struct ost_id   *oi     = &oa->o_oi;
 137         u64              seq    = ostid_seq(oi);
 138         u64              id     = ostid_id(oi);
 139         int              rc;
 140         ENTRY;
 141
 142         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 143                      fid_seq_is_echo(seq))) {
 144                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 145                  * during create, and we will reset this to 1, since this
 146                  * oi_id is basically useless in the following create process,
 147                  * but oi_id == 0 will make it difficult to tell whether it is
 148                  * real FID or ost_id. */
 149                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 150                 oi->oi_fid.f_oid = id ?: 1;
 151                 oi->oi_fid.f_ver = 0;
 152         } else {
 153                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 154
 155                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 156                         GOTO(out, rc = -EPROTO);
 157
 158                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 159                  * all of the requests are required to setup FLGROUP */
 160                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 161                         ostid_set_seq_mdt0(oi);
 162                         oa->o_valid |= OBD_MD_FLGROUP;
 163                         seq = ostid_seq(oi);
 164                 }
 165
 166                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 167                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 168                         GOTO(out, rc = -EPROTO);
 169
 170                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 171                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 172                 if (unlikely(rc != 0))
 173                         GOTO(out, rc);
 174
 175                 oi->oi_fid = tti->tti_fid1;
 176         }
 177
 178         RETURN(0);
 179
 180 out:
 181         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 182                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 183                seq, id, rc);
 184         return rc;
 185 }
 186 EXPORT_SYMBOL(tgt_validate_obdo);
 187
 188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 189 {
 190         unsigned                 max_brw;
 191         struct niobuf_remote    *rnb;
 192         struct obd_ioobj        *ioo;
 193         int                      obj_count;
 194
 195         ENTRY;
 196
 197         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 198         if (ioo == NULL)
 199                 RETURN(-EPROTO);
 200
 201         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 202         if (rnb == NULL)
 203                 RETURN(-EPROTO);
 204
 205         max_brw = ioobj_max_brw_get(ioo);
 206         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 207                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 208                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 209                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 210                        POSTID(oi), -EPROTO);
 211                 RETURN(-EPROTO);
 212         }
 213         ioo->ioo_oid = *oi;
 214
 215         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 216                                         RCL_CLIENT) / sizeof(*ioo);
 217         if (obj_count == 0) {
 218                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 219                 RETURN(-EPROTO);
 220         } else if (obj_count > 1) {
 221                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 222                        obj_count);
 223                 RETURN(-EPROTO);
 224         }
 225
 226         if (ioo->ioo_bufcnt == 0) {
 227                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 228                 RETURN(-EPROTO);
 229         }
 230
 231         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 232                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 233                           "bulk has too many pages (%d)",
 234                           ioo->ioo_bufcnt);
 235                 RETURN(-EPROTO);
 236         }
 237
 238         RETURN(0);
 239 }
 240
 241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 242 {
 243         struct ost_body         *body;
 244         struct req_capsule      *pill = tsi->tsi_pill;
 245         struct lu_nodemap       *nodemap;
 246         int                      rc;
 247
 248         ENTRY;
 249
 250         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 251         if (body == NULL)
 252                 RETURN(-EFAULT);
 253
 254         rc = tgt_validate_obdo(tsi, &body->oa);
 255         if (rc)
 256                 RETURN(rc);
 257
 258         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 259         if (IS_ERR(nodemap))
 260                 RETURN(PTR_ERR(nodemap));
 261
 262         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_uid);
 265         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 266                                         NODEMAP_CLIENT_TO_FS,
 267                                         body->oa.o_gid);
 268         body->oa.o_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
 269                                            NODEMAP_CLIENT_TO_FS,
 270                                            body->oa.o_projid);
 271         nodemap_putref(nodemap);
 272
 273         tsi->tsi_ost_body = body;
 274         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 275
 276         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 277                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 278                 if (rc < 0)
 279                         RETURN(rc);
 280         }
 281
 282         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 283                 if (flags & HAS_BODY) {
 284                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 285                                tgt_name(tsi->tsi_tgt));
 286                         RETURN(-EPROTO);
 287                 } else {
 288                         RETURN(0);
 289                 }
 290         }
 291
 292         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 293
 294         /*
 295          * OST doesn't get object in advance for further use to prevent
 296          * situations with nested object_find which is potential deadlock.
 297          */
 298         tsi->tsi_corpus = NULL;
 299         RETURN(rc);
 300 }
 301
 302 /*
 303  * Do necessary preprocessing according to handler ->th_flags.
 304  */
 305 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 306                                   struct tgt_handler *h,
 307                                   struct ptlrpc_request *req)
 308 {
 309         struct req_capsule      *pill = tsi->tsi_pill;
 310         __u32                    flags = h->th_flags;
 311         int                      rc = 0;
 312
 313         ENTRY;
 314
 315         if (tsi->tsi_preprocessed)
 316                 RETURN(0);
 317
 318         LASSERT(h->th_act != NULL);
 319         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 320         LASSERT(current->journal_info == NULL);
 321
 322         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 323                      h->th_fmt != NULL));
 324         if (h->th_fmt != NULL) {
 325                 req_capsule_set(pill, h->th_fmt);
 326                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT) &&
 327                     req_capsule_field_present(pill, &RMF_MDT_BODY,
 328                                               RCL_CLIENT)) {
 329                         rc = tgt_mdt_body_unpack(tsi, flags);
 330                         if (rc < 0)
 331                                 RETURN(rc);
 332                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 333                                                  RCL_CLIENT) &&
 334                            req_capsule_field_present(pill, &RMF_OST_BODY,
 335                                                  RCL_CLIENT)) {
 336                         rc = tgt_ost_body_unpack(tsi, flags);
 337                         if (rc < 0)
 338                                 RETURN(rc);
 339                 }
 340         }
 341
 342         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 343                 RETURN(-EROFS);
 344
 345         if (flags & HAS_KEY) {
 346                 struct ldlm_request *dlm_req;
 347
 348                 LASSERT(h->th_fmt != NULL);
 349
 350                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 351                 if (dlm_req != NULL) {
 352                         union ldlm_wire_policy_data *policy =
 353                                         &dlm_req->lock_desc.l_policy_data;
 354
 355                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 356                                      LDLM_IBITS &&
 357                                      (policy->l_inodebits.bits |
 358                                       policy->l_inodebits.try_bits) == 0)) {
 359                                 /*
 360                                  * Lock without inodebits makes no sense and
 361                                  * will oops later in ldlm. If client miss to
 362                                  * set such bits, do not trigger ASSERTION.
 363                                  *
 364                                  * For liblustre flock case, it maybe zero.
 365                                  */
 366                                 rc = -EPROTO;
 367                         } else {
 368                                 tsi->tsi_dlm_req = dlm_req;
 369                         }
 370                 } else {
 371                         rc = -EFAULT;
 372                 }
 373         }
 374         tsi->tsi_preprocessed = 1;
 375         RETURN(rc);
 376 }
 377
 378 /*
 379  * Invoke handler for this request opc. Also do necessary preprocessing
 380  * (according to handler ->th_flags), and post-processing (setting of
 381  * ->last_{xid,committed}).
 382  */
 383 static int tgt_handle_request0(struct tgt_session_info *tsi,
 384                                struct tgt_handler *h,
 385                                struct ptlrpc_request *req)
 386 {
 387         int      serious = 0;
 388         int      rc;
 389         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 390
 391         ENTRY;
 392
 393
 394         /* When dealing with sec context requests, no export is associated yet,
 395          * because these requests are sent before *_CONNECT requests.
 396          * A NULL req->rq_export means the normal *_common_slice handlers will
 397          * not be called, because there is no reference to the target.
 398          * So deal with them by hand and jump directly to target_send_reply().
 399          */
 400         switch (opc) {
 401         case SEC_CTX_INIT:
 402         case SEC_CTX_INIT_CONT:
 403         case SEC_CTX_FINI:
 404                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 405                 GOTO(out, rc = 0);
 406         }
 407
 408         /*
 409          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 410          * to put same checks into handlers like mdt_close(), mdt_reint(),
 411          * etc., without talking to mdt authors first. Checking same thing
 412          * there again is useless and returning 0 error without packing reply
 413          * is buggy! Handlers either pack reply or return error.
 414          *
 415          * We return 0 here and do not send any reply in order to emulate
 416          * network failure. Do not send any reply in case any of NET related
 417          * fail_id has occured.
 418          */
 419         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 420                 RETURN(0);
 421         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 422                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 423                 RETURN(0);
 424
 425         /* drop OUT_UPDATE rpc */
 426         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == OUT_UPDATE &&
 427                      OBD_FAIL_CHECK(OBD_FAIL_OUT_UPDATE_DROP)))
 428                 RETURN(0);
 429
 430         rc = tgt_request_preprocess(tsi, h, req);
 431         /* pack reply if reply format is fixed */
 432         if (rc == 0 && h->th_flags & HAS_REPLY) {
 433                 /* Pack reply */
 434                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 435                                           RCL_SERVER))
 436                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 437                                              RCL_SERVER,
 438                                              tsi->tsi_mdt_body->mbo_eadatasize);
 439                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 440                                           RCL_SERVER))
 441                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 442                                              RCL_SERVER, 0);
 443                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 444                         req_capsule_set_size(tsi->tsi_pill,
 445                                              &RMF_ACL, RCL_SERVER,
 446                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 447
 448                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 449                                           RCL_SERVER)) {
 450                         struct niobuf_remote *remote_nb =
 451                                 req_capsule_client_get(tsi->tsi_pill,
 452                                                        &RMF_NIOBUF_REMOTE);
 453                         struct ost_body *body = tsi->tsi_ost_body;
 454
 455                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 456                                          RCL_SERVER,
 457                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 458                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 459                                          remote_nb[0].rnb_len : 0);
 460                 }
 461
 462                 rc = req_capsule_server_pack(tsi->tsi_pill);
 463         }
 464
 465         if (likely(rc == 0)) {
 466                 /*
 467                  * Process request, there can be two types of rc:
 468                  * 1) errors with msg unpack/pack, other failures outside the
 469                  * operation itself. This is counted as serious errors;
 470                  * 2) errors during fs operation, should be placed in rq_status
 471                  * only
 472                  */
 473                 rc = h->th_act(tsi);
 474                 if (!is_serious(rc) &&
 475                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 476                         DEBUG_REQ(D_ERROR, req,
 477                                   "%s: %s handler did not pack reply but returned no error",
 478                                   tgt_name(tsi->tsi_tgt), h->th_name);
 479                         LBUG();
 480                 }
 481                 serious = is_serious(rc);
 482                 rc = clear_serious(rc);
 483         } else {
 484                 serious = 1;
 485         }
 486
 487         req->rq_status = rc;
 488
 489         /*
 490          * ELDLM_* codes which > 0 should be in rq_status only as well as
 491          * all non-serious errors.
 492          */
 493         if (rc > 0 || !serious)
 494                 rc = 0;
 495
 496         LASSERT(current->journal_info == NULL);
 497
 498         if (likely(rc == 0 && req->rq_export))
 499                 target_committed_to_req(req);
 500
 501 out:
 502         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 503         RETURN(0);
 504 }
 505
 506 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 507                                        struct obd_device *obd, int *process)
 508 {
 509         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 510         case MDS_DISCONNECT:
 511         case OST_DISCONNECT:
 512         case OBD_IDX_READ:
 513                 *process = 1;
 514                 RETURN(0);
 515         case MDS_CLOSE:
 516         case MDS_SYNC: /* used in unmounting */
 517         case OBD_PING:
 518         case MDS_REINT:
 519         case OUT_UPDATE:
 520         case SEQ_QUERY:
 521         case FLD_QUERY:
 522         case FLD_READ:
 523         case LDLM_ENQUEUE:
 524         case OST_CREATE:
 525         case OST_DESTROY:
 526         case OST_PUNCH:
 527         case OST_SETATTR:
 528         case OST_SYNC:
 529         case OST_WRITE:
 530         case MDS_HSM_PROGRESS:
 531         case MDS_HSM_STATE_SET:
 532         case MDS_HSM_REQUEST:
 533         case OST_FALLOCATE:
 534                 *process = target_queue_recovery_request(req, obd);
 535                 RETURN(0);
 536
 537         default:
 538                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 539                 *process = -EAGAIN;
 540                 RETURN(0);
 541         }
 542 }
 543
 544 /*
 545  * Handle recovery. Return:
 546  *        +1: continue request processing;
 547  *       -ve: abort immediately with the given error code;
 548  *         0: send reply with error code in req->rq_status;
 549  */
 550 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 551 {
 552         ENTRY;
 553
 554         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 555         case MDS_CONNECT:
 556         case OST_CONNECT:
 557         case MGS_CONNECT:
 558         case SEC_CTX_INIT:
 559         case SEC_CTX_INIT_CONT:
 560         case SEC_CTX_FINI:
 561                 RETURN(+1);
 562         }
 563
 564         if (!req->rq_export->exp_obd->obd_replayable)
 565                 RETURN(+1);
 566
 567         /* sanity check: if the xid matches, the request must be marked as a
 568          * resent or replayed */
 569         if (req_can_reconstruct(req, NULL) == 1) {
 570                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 571                       (MSG_RESENT | MSG_REPLAY))) {
 572                         DEBUG_REQ(D_WARNING, req,
 573                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 574                                   req->rq_xid,
 575                                   lustre_msg_get_flags(req->rq_reqmsg));
 576                         req->rq_status = -ENOTCONN;
 577                         RETURN(-ENOTCONN);
 578                 }
 579         }
 580         /* else: note the opposite is not always true; a RESENT req after a
 581          * failover will usually not match the last_xid, since it was likely
 582          * never committed. A REPLAYed request will almost never match the
 583          * last xid, however it could for a committed, but still retained,
 584          * open. */
 585
 586         /* Check for aborted recovery... */
 587         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 588                 int rc;
 589                 int should_process;
 590
 591                 DEBUG_REQ(D_INFO, req, "Got new replay");
 592                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 593                                                  &should_process);
 594                 if (rc != 0 || !should_process)
 595                         RETURN(rc);
 596                 else if (should_process < 0) {
 597                         req->rq_status = should_process;
 598                         rc = ptlrpc_error(req);
 599                         RETURN(rc);
 600                 }
 601         }
 602         RETURN(+1);
 603 }
 604
 605 /* Initial check for request, it is validation mostly */
 606 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 607 {
 608         struct tgt_handler      *h;
 609         struct tgt_opc_slice    *s;
 610         struct lu_target        *tgt;
 611         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 612
 613         ENTRY;
 614
 615         tgt = class_exp2tgt(req->rq_export);
 616         if (unlikely(tgt == NULL)) {
 617                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 618                           class_exp2obd(req->rq_export)->obd_name);
 619                 RETURN(ERR_PTR(-EINVAL));
 620         }
 621
 622         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 623                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 624                         break;
 625
 626         /* opcode was not found in slice */
 627         if (unlikely(s->tos_hs == NULL)) {
 628                 static bool printed;
 629
 630                 /* don't spew error messages for unhandled RPCs */
 631                 if (!printed) {
 632                         CERROR("%s: no handler for opcode 0x%x from %s\n",
 633                                tgt_name(tgt), opc, libcfs_id2str(req->rq_peer));
 634                         printed = true;
 635                 }
 636                 RETURN(ERR_PTR(-ENOTSUPP));
 637         }
 638
 639         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 640         h = s->tos_hs + (opc - s->tos_opc_start);
 641         if (unlikely(h->th_opc == 0)) {
 642                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 643                 RETURN(ERR_PTR(-ENOTSUPP));
 644         }
 645
 646         RETURN(h);
 647 }
 648
 649 static int process_req_last_xid(struct ptlrpc_request *req)
 650 {
 651         __u64   last_xid;
 652         int rc = 0;
 653         struct obd_export *exp = req->rq_export;
 654         struct tg_export_data *ted = &exp->exp_target_data;
 655         bool need_lock = tgt_is_multimodrpcs_client(exp);
 656         ENTRY;
 657
 658         if (need_lock)
 659                 mutex_lock(&ted->ted_lcd_lock);
 660         /* check request's xid is consistent with export's last_xid */
 661         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 662         if (last_xid > exp->exp_last_xid)
 663                 exp->exp_last_xid = last_xid;
 664
 665         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 666                 /* Some request is allowed to be sent during replay,
 667                  * such as OUT update requests, FLD requests, so it
 668                  * is possible that replay requests has smaller XID
 669                  * than the exp_last_xid.
 670                  *
 671                  * Some non-replay requests may have smaller XID as
 672                  * well:
 673                  *
 674                  * - Client send a no_resend RPC, like statfs;
 675                  * - The RPC timedout (or some other error) on client,
 676                  *   then it's removed from the unreplied list;
 677                  * - Client send some other request to bump the
 678                  *   exp_last_xid on server;
 679                  * - The former RPC got chance to be processed;
 680                  */
 681                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 682                         rc = -EPROTO;
 683
 684                 DEBUG_REQ(D_WARNING, req,
 685                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 686                           req->rq_xid, exp->exp_last_xid, rc);
 687                 if (rc)
 688                         GOTO(out, rc);
 689         }
 690
 691         /* The "last_xid" is the minimum xid among unreplied requests,
 692          * if the request is from the previous connection, its xid can
 693          * still be larger than "exp_last_xid", then the above check of
 694          * xid is not enough to determine whether the request is delayed.
 695          *
 696          * For example, if some replay request was delayed and caused
 697          * timeout at client and the replay is restarted, the delayed
 698          * replay request will have the larger xid than "exp_last_xid"
 699          */
 700         if (req->rq_export->exp_conn_cnt >
 701             lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 702                 CDEBUG(D_RPCTRACE,
 703                        "Dropping request %llu from an old epoch %u/%u\n",
 704                        req->rq_xid,
 705                        lustre_msg_get_conn_cnt(req->rq_reqmsg),
 706                        req->rq_export->exp_conn_cnt);
 707                 req->rq_no_reply = 1;
 708                 GOTO(out, rc = -ESTALE);
 709         }
 710
 711         /* try to release in-memory reply data */
 712         if (tgt_is_multimodrpcs_client(exp)) {
 713                 tgt_handle_received_xid(exp, last_xid);
 714                 rc = tgt_handle_tag(req);
 715         }
 716
 717 out:
 718         if (need_lock)
 719                 mutex_unlock(&ted->ted_lcd_lock);
 720
 721         RETURN(rc);
 722 }
 723
 724 int tgt_request_handle(struct ptlrpc_request *req)
 725 {
 726         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 727
 728         struct lustre_msg       *msg = req->rq_reqmsg;
 729         struct tgt_handler      *h;
 730         struct lu_target        *tgt;
 731         int                      request_fail_id = 0;
 732         __u32                    opc = lustre_msg_get_opc(msg);
 733         struct obd_device       *obd;
 734         int                      rc;
 735         bool                     is_connect = false;
 736         ENTRY;
 737
 738         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 739                 if (cfs_fail_val == 0 &&
 740                     lustre_msg_get_opc(msg) != OBD_PING &&
 741                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 742                         cfs_fail_val = 1;
 743                         cfs_race_state = 0;
 744                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 745                 }
 746         }
 747
 748         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 749         tsi->tsi_pill = &req->rq_pill;
 750         tsi->tsi_env = req->rq_svc_thread->t_env;
 751
 752         /* if request has export then get handlers slice from corresponding
 753          * target, otherwise that should be connect operation */
 754         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 755             opc == MGS_CONNECT) {
 756                 is_connect = true;
 757                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 758                 rc = target_handle_connect(req);
 759                 if (rc != 0) {
 760                         rc = ptlrpc_error(req);
 761                         GOTO(out, rc);
 762                 }
 763                 /* recovery-small test 18c asks to drop connect reply */
 764                 if (unlikely(opc == OST_CONNECT &&
 765                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 766                         GOTO(out, rc = 0);
 767         }
 768
 769         if (unlikely(!class_connected_export(req->rq_export))) {
 770                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 771                     opc == SEC_CTX_FINI) {
 772                         /* sec context initialization has to be handled
 773                          * by hand in tgt_handle_request0() */
 774                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 775                         h = NULL;
 776                         GOTO(handle_recov, rc = 0);
 777                 }
 778                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 779                        opc, libcfs_id2str(req->rq_peer));
 780                 req->rq_status = -ENOTCONN;
 781                 rc = ptlrpc_error(req);
 782                 GOTO(out, rc);
 783         }
 784
 785         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 786         tsi->tsi_exp = req->rq_export;
 787         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 788                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 789         else
 790                 tsi->tsi_jobid = NULL;
 791
 792         if (tgt == NULL) {
 793                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 794                           class_exp2obd(req->rq_export)->obd_name);
 795                 req->rq_status = -EINVAL;
 796                 rc = ptlrpc_error(req);
 797                 GOTO(out, rc);
 798         }
 799
 800         /* Skip last_xid processing for the recovery thread, otherwise, the
 801          * last_xid on same request could be processed twice: first time when
 802          * processing the incoming request, second time when the request is
 803          * being processed by recovery thread. */
 804         obd = class_exp2obd(req->rq_export);
 805         if (is_connect) {
 806                 /* reset the exp_last_xid on each connection. */
 807                 req->rq_export->exp_last_xid = 0;
 808         } else if (obd->obd_recovery_data.trd_processing_task !=
 809                    current->pid) {
 810                 rc = process_req_last_xid(req);
 811                 if (rc) {
 812                         req->rq_status = rc;
 813                         rc = ptlrpc_error(req);
 814                         GOTO(out, rc);
 815                 }
 816         }
 817
 818         request_fail_id = tgt->lut_request_fail_id;
 819         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 820
 821         h = tgt_handler_find_check(req);
 822         if (IS_ERR(h)) {
 823                 req->rq_status = PTR_ERR(h);
 824                 rc = ptlrpc_error(req);
 825                 GOTO(out, rc);
 826         }
 827
 828         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 829                  h->th_opc, opc);
 830
 831         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 832              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 833                 GOTO(out, rc = 0);
 834
 835         rc = lustre_msg_check_version(msg, h->th_version);
 836         if (unlikely(rc)) {
 837                 DEBUG_REQ(D_ERROR, req,
 838                           "%s: drop malformed request version=%08x expect=%08x",
 839                           tgt_name(tgt), lustre_msg_get_version(msg),
 840                           h->th_version);
 841                 req->rq_status = -EINVAL;
 842                 rc = ptlrpc_error(req);
 843                 GOTO(out, rc);
 844         }
 845
 846 handle_recov:
 847         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 848         if (likely(rc == 1)) {
 849                 rc = tgt_handle_request0(tsi, h, req);
 850                 if (rc)
 851                         GOTO(out, rc);
 852         }
 853         EXIT;
 854 out:
 855         req_capsule_fini(tsi->tsi_pill);
 856         if (tsi->tsi_corpus != NULL) {
 857                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 858                 tsi->tsi_corpus = NULL;
 859         }
 860         return rc;
 861 }
 862 EXPORT_SYMBOL(tgt_request_handle);
 863
 864 /** Assign high priority operations to the request if needed. */
 865 int tgt_hpreq_handler(struct ptlrpc_request *req)
 866 {
 867         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 868         struct tgt_handler      *h;
 869         int                      rc;
 870
 871         ENTRY;
 872
 873         if (req->rq_export == NULL)
 874                 RETURN(0);
 875
 876         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 877         tsi->tsi_pill = &req->rq_pill;
 878         tsi->tsi_env = req->rq_svc_thread->t_env;
 879         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 880         tsi->tsi_exp = req->rq_export;
 881
 882         h = tgt_handler_find_check(req);
 883         if (IS_ERR(h)) {
 884                 rc = PTR_ERR(h);
 885                 RETURN(rc);
 886         }
 887
 888         rc = tgt_request_preprocess(tsi, h, req);
 889         if (unlikely(rc != 0))
 890                 RETURN(rc);
 891
 892         if (h->th_hp != NULL)
 893                 h->th_hp(tsi);
 894         RETURN(0);
 895 }
 896 EXPORT_SYMBOL(tgt_hpreq_handler);
 897
 898 void tgt_counter_incr(struct obd_export *exp, int opcode)
 899 {
 900         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 901         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 902                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 903 }
 904 EXPORT_SYMBOL(tgt_counter_incr);
 905
 906 /*
 907  * Unified target generic handlers.
 908  */
 909
 910 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 911 {
 912         struct lu_target *tgt = class_exp2tgt(exp);
 913         struct sptlrpc_flavor flvr;
 914         int rc = 0;
 915
 916         LASSERT(tgt);
 917         LASSERT(tgt->lut_obd);
 918         LASSERT(tgt->lut_slice);
 919
 920         /* always allow ECHO client */
 921         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 922                             LUSTRE_ECHO_NAME) == 0)) {
 923                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 924                 return 0;
 925         }
 926
 927         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 928                 read_lock(&tgt->lut_sptlrpc_lock);
 929                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 930                                              req->rq_sp_from,
 931                                              req->rq_peer.nid,
 932                                              &flvr);
 933                 read_unlock(&tgt->lut_sptlrpc_lock);
 934
 935                 spin_lock(&exp->exp_lock);
 936                 exp->exp_sp_peer = req->rq_sp_from;
 937                 exp->exp_flvr = flvr;
 938
 939                 /* when on mgs, if no restriction is set, or if the client
 940                  * NID is on the local node, allow any flavor
 941                  */
 942                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 943                            LUSTRE_MGS_NAME) == 0) &&
 944                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 945                       LNetIsPeerLocal(exp->exp_connection->c_peer.nid)))
 946                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 947
 948                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 949                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 950                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 951                                "expect %x\n", tgt_name(tgt),
 952                                req->rq_flvr.sf_rpc,
 953                                libcfs_nid2str(req->rq_peer.nid),
 954                                exp->exp_flvr.sf_rpc);
 955                         rc = -EACCES;
 956                 }
 957                 spin_unlock(&exp->exp_lock);
 958         } else {
 959                 if (exp->exp_sp_peer != req->rq_sp_from) {
 960                         CERROR("%s: RPC source %s doesn't match %s\n",
 961                                tgt_name(tgt),
 962                                sptlrpc_part2name(req->rq_sp_from),
 963                                sptlrpc_part2name(exp->exp_sp_peer));
 964                         rc = -EACCES;
 965                 } else {
 966                         rc = sptlrpc_target_export_check(exp, req);
 967                 }
 968         }
 969
 970         return rc;
 971 }
 972
 973 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 974 {
 975         struct sptlrpc_rule_set  tmp_rset;
 976         int                      rc;
 977
 978         if (unlikely(tgt == NULL)) {
 979                 CERROR("No target passed\n");
 980                 return -EINVAL;
 981         }
 982
 983         sptlrpc_rule_set_init(&tmp_rset);
 984         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 985         if (rc) {
 986                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 987                        tgt_name(tgt), rc);
 988                 return rc;
 989         }
 990
 991         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 992
 993         write_lock(&tgt->lut_sptlrpc_lock);
 994         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 995         tgt->lut_sptlrpc_rset = tmp_rset;
 996         write_unlock(&tgt->lut_sptlrpc_lock);
 997
 998         return 0;
 999 }
1000 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
1001
1002 int tgt_connect(struct tgt_session_info *tsi)
1003 {
1004         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1005         struct obd_connect_data *reply;
1006         int                      rc;
1007
1008         ENTRY;
1009
1010         /* XXX: better to call this check right after getting new export but
1011          * before last_rcvd slot allocation to avoid server load upon insecure
1012          * connects. This is to be fixed after unifiyng all targets.
1013          */
1014         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1015         if (rc)
1016                 GOTO(out, rc);
1017
1018         /* To avoid exposing partially initialized connection flags, changes up
1019          * to this point have been staged in reply->ocd_connect_flags. Now that
1020          * connection handling has completed successfully, atomically update
1021          * the connect flags in the shared export data structure. LU-1623 */
1022         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1023         spin_lock(&tsi->tsi_exp->exp_lock);
1024         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1025         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1026                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1027                         reply->ocd_connect_flags2;
1028         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1029         spin_unlock(&tsi->tsi_exp->exp_lock);
1030
1031         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1032                    LUSTRE_MDT_NAME) == 0) {
1033                 rc = req_check_sepol(tsi->tsi_pill);
1034                 if (rc)
1035                         GOTO(out, rc);
1036
1037                 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1038                     reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT &&
1039                     tsi->tsi_pill->rc_req->rq_export) {
1040                         bool forbid_encrypt = true;
1041                         struct lu_nodemap *nm =
1042                          nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1043
1044                         if (!nm) {
1045                                 /* nodemap_get_from_exp returns NULL in case
1046                                  * nodemap is not active, so we do not forbid
1047                                  */
1048                                 forbid_encrypt = false;
1049                         } else if (!IS_ERR(nm)) {
1050                                 forbid_encrypt = nm->nmf_forbid_encryption;
1051                                 nodemap_putref(nm);
1052                         }
1053
1054                         if (forbid_encrypt)
1055                                 GOTO(out, rc = -EACCES);
1056                 }
1057         }
1058
1059         RETURN(0);
1060 out:
1061         obd_disconnect(class_export_get(tsi->tsi_exp));
1062         return rc;
1063 }
1064 EXPORT_SYMBOL(tgt_connect);
1065
1066 int tgt_disconnect(struct tgt_session_info *tsi)
1067 {
1068         int rc;
1069
1070         ENTRY;
1071
1072         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1073
1074         rc = target_handle_disconnect(tgt_ses_req(tsi));
1075         if (rc)
1076                 RETURN(err_serious(rc));
1077
1078         RETURN(rc);
1079 }
1080 EXPORT_SYMBOL(tgt_disconnect);
1081
1082 /*
1083  * Unified target OBD handlers
1084  */
1085 int tgt_obd_ping(struct tgt_session_info *tsi)
1086 {
1087         int rc;
1088
1089         ENTRY;
1090
1091         /* The target-specific part of OBD_PING request handling.
1092          * It controls Filter Modification Data (FMD) expiration each time
1093          * PING is received.
1094          *
1095          * Valid only for replayable targets, e.g. MDT and OFD
1096          */
1097         if (tsi->tsi_exp->exp_obd->obd_replayable)
1098                 tgt_fmd_expire(tsi->tsi_exp);
1099
1100         rc = req_capsule_server_pack(tsi->tsi_pill);
1101         if (rc)
1102                 RETURN(err_serious(rc));
1103
1104         RETURN(rc);
1105 }
1106 EXPORT_SYMBOL(tgt_obd_ping);
1107
1108 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1109 {
1110         return err_serious(-EOPNOTSUPP);
1111 }
1112
1113 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1114 {
1115         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1116         struct obd_export       *exp = req->rq_export;
1117         struct ptlrpc_bulk_desc *desc;
1118         int                      i;
1119         int                      rc;
1120         int                      pages = 0;
1121
1122         ENTRY;
1123
1124         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1125                 unsigned int offset;
1126
1127                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1128                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1129                                       PAGE_SIZE);
1130         }
1131
1132         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1133                                   PTLRPC_BULK_PUT_SOURCE,
1134                                     MDS_BULK_PORTAL,
1135                                     &ptlrpc_bulk_kiov_nopin_ops);
1136         if (desc == NULL)
1137                 RETURN(-ENOMEM);
1138
1139         for (i = 0; i < rdbuf->rb_nbufs; i++)
1140                 desc->bd_frag_ops->add_iov_frag(desc,
1141                                         rdbuf->rb_bufs[i].lb_buf,
1142                                         rdbuf->rb_bufs[i].lb_len);
1143
1144         rc = target_bulk_io(exp, desc);
1145         ptlrpc_free_bulk(desc);
1146         RETURN(rc);
1147 }
1148 EXPORT_SYMBOL(tgt_send_buffer);
1149
1150 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1151 {
1152         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1153         struct obd_export       *exp = req->rq_export;
1154         struct ptlrpc_bulk_desc *desc;
1155         int                      tmpcount;
1156         int                      tmpsize;
1157         int                      i;
1158         int                      rc;
1159
1160         ENTRY;
1161
1162         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1163                                     PTLRPC_BULK_PUT_SOURCE,
1164                                     MDS_BULK_PORTAL,
1165                                     &ptlrpc_bulk_kiov_pin_ops);
1166         if (desc == NULL)
1167                 RETURN(-ENOMEM);
1168
1169         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1170                 /* old client requires reply size in it's PAGE_SIZE,
1171                  * which is rdpg->rp_count */
1172                 nob = rdpg->rp_count;
1173
1174         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1175              i++, tmpcount -= tmpsize) {
1176                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1177                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1178                                                  tmpsize);
1179         }
1180
1181         LASSERT(desc->bd_nob == nob);
1182         rc = target_bulk_io(exp, desc);
1183         ptlrpc_free_bulk(desc);
1184         RETURN(rc);
1185 }
1186 EXPORT_SYMBOL(tgt_sendpage);
1187
1188 /*
1189  * OBD_IDX_READ handler
1190  */
1191 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1192 {
1193         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1194         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1195         struct idx_info         *req_ii, *rep_ii;
1196         int                      rc, i;
1197
1198         ENTRY;
1199
1200         memset(rdpg, 0, sizeof(*rdpg));
1201         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1202
1203         /* extract idx_info buffer from request & reply */
1204         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1205         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1206                 RETURN(err_serious(-EPROTO));
1207
1208         rc = req_capsule_server_pack(tsi->tsi_pill);
1209         if (rc)
1210                 RETURN(err_serious(rc));
1211
1212         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1213         if (rep_ii == NULL)
1214                 RETURN(err_serious(-EFAULT));
1215         rep_ii->ii_magic = IDX_INFO_MAGIC;
1216
1217         /* extract hash to start with */
1218         rdpg->rp_hash = req_ii->ii_hash_start;
1219
1220         /* extract requested attributes */
1221         rdpg->rp_attrs = req_ii->ii_attrs;
1222
1223         /* check that fid packed in request is valid and supported */
1224         if (!fid_is_sane(&req_ii->ii_fid))
1225                 RETURN(-EINVAL);
1226         rep_ii->ii_fid = req_ii->ii_fid;
1227
1228         /* copy flags */
1229         rep_ii->ii_flags = req_ii->ii_flags;
1230
1231         /* compute number of pages to allocate, ii_count is the number of 4KB
1232          * containers */
1233         if (req_ii->ii_count <= 0)
1234                 GOTO(out, rc = -EFAULT);
1235         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1236                                exp_max_brw_size(tsi->tsi_exp));
1237         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1238
1239         /* allocate pages to store the containers */
1240         OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1241         if (rdpg->rp_pages == NULL)
1242                 GOTO(out, rc = -ENOMEM);
1243         for (i = 0; i < rdpg->rp_npages; i++) {
1244                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1245                 if (rdpg->rp_pages[i] == NULL)
1246                         GOTO(out, rc = -ENOMEM);
1247         }
1248
1249         /* populate pages with key/record pairs */
1250         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1251         if (rc < 0)
1252                 GOTO(out, rc);
1253
1254         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1255                  "asked %d > %d\n", rc, rdpg->rp_count);
1256
1257         /* send pages to client */
1258         rc = tgt_sendpage(tsi, rdpg, rc);
1259         if (rc)
1260                 GOTO(out, rc);
1261         EXIT;
1262 out:
1263         if (rdpg->rp_pages) {
1264                 for (i = 0; i < rdpg->rp_npages; i++)
1265                         if (rdpg->rp_pages[i])
1266                                 __free_page(rdpg->rp_pages[i]);
1267                 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1268         }
1269         return rc;
1270 }
1271
1272 struct tgt_handler tgt_obd_handlers[] = {
1273 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1274 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1275 };
1276 EXPORT_SYMBOL(tgt_obd_handlers);
1277
1278 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1279              struct dt_object *obj, __u64 start, __u64 end)
1280 {
1281         int rc = 0;
1282
1283         ENTRY;
1284
1285         /* if no objid is specified, it means "sync whole filesystem" */
1286         if (obj == NULL) {
1287                 rc = dt_sync(env, tgt->lut_bottom);
1288         } else if (dt_version_get(env, obj) >
1289                    tgt->lut_obd->obd_last_committed) {
1290                 rc = dt_object_sync(env, obj, start, end);
1291         }
1292         atomic_inc(&tgt->lut_sync_count);
1293
1294         RETURN(rc);
1295 }
1296 EXPORT_SYMBOL(tgt_sync);
1297 /*
1298  * Unified target DLM handlers.
1299  */
1300
1301 /**
1302  * Unified target BAST
1303  *
1304  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1305  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1306  * but if it's cross-MDT lock, because remote object version is not set, a
1307  * filesystem sync is needed.
1308  *
1309  * \param lock server side lock
1310  * \param desc lock desc
1311  * \param data ldlm_cb_set_arg
1312  * \param flag  indicates whether this cancelling or blocking callback
1313  * \retval      0 on success
1314  * \retval      negative number on error
1315  */
1316 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1317                             void *data, int flag)
1318 {
1319         struct lu_env            env;
1320         struct lu_target        *tgt;
1321         struct dt_object        *obj = NULL;
1322         struct lu_fid            fid;
1323         int                      rc = 0;
1324
1325         ENTRY;
1326
1327         tgt = class_exp2tgt(lock->l_export);
1328
1329         if (unlikely(tgt == NULL)) {
1330                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1331                        class_exp2obd(lock->l_export)->obd_name);
1332                 RETURN(-EINVAL);
1333         }
1334
1335         if (flag == LDLM_CB_CANCELING &&
1336             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1337             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1338              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1339               ldlm_is_cbpending(lock))) &&
1340             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1341              lock->l_resource->lr_type == LDLM_EXTENT)) {
1342                 __u64 start = 0;
1343                 __u64 end = OBD_OBJECT_EOF;
1344
1345                 rc = lu_env_init(&env, LCT_DT_THREAD);
1346                 if (unlikely(rc != 0))
1347                         GOTO(err, rc);
1348
1349                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1350                                    tgt->lut_lsd.lsd_osd_index);
1351
1352                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1353                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1354                         if (IS_ERR(obj))
1355                                 GOTO(err_env, rc = PTR_ERR(obj));
1356
1357                         if (!dt_object_exists(obj))
1358                                 GOTO(err_put, rc = -ENOENT);
1359
1360                         start = lock->l_policy_data.l_extent.start;
1361                         end = lock->l_policy_data.l_extent.end;
1362                 }
1363
1364                 rc = tgt_sync(&env, tgt, obj, start, end);
1365                 if (rc < 0) {
1366                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1367                                "cancel: rc = %d\n",
1368                                tgt_name(tgt), PFID(&fid),
1369                                lock->l_policy_data.l_extent.start,
1370                                lock->l_policy_data.l_extent.end, rc);
1371                 }
1372 err_put:
1373                 if (obj != NULL)
1374                         dt_object_put(&env, obj);
1375 err_env:
1376                 lu_env_fini(&env);
1377         }
1378 err:
1379         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1380         RETURN(rc);
1381 }
1382
1383 static struct ldlm_callback_suite tgt_dlm_cbs = {
1384         .lcs_completion = ldlm_server_completion_ast,
1385         .lcs_blocking   = tgt_blocking_ast,
1386         .lcs_glimpse    = ldlm_server_glimpse_ast
1387 };
1388
1389 int tgt_enqueue(struct tgt_session_info *tsi)
1390 {
1391         struct ptlrpc_request *req = tgt_ses_req(tsi);
1392         int rc;
1393
1394         ENTRY;
1395         /*
1396          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1397          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1398          */
1399         LASSERT(tsi->tsi_dlm_req != NULL);
1400         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1401                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1402         if (rc)
1403                 RETURN(err_serious(rc));
1404
1405         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1406         case LUT_FAIL_MDT:
1407                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1408                 break;
1409         case LUT_FAIL_OST:
1410                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1411                 break;
1412         case LUT_FAIL_MGT:
1413                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1414                 break;
1415         default:
1416                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1417                 break;
1418         }
1419         RETURN(req->rq_status);
1420 }
1421 EXPORT_SYMBOL(tgt_enqueue);
1422
1423 int tgt_convert(struct tgt_session_info *tsi)
1424 {
1425         struct ptlrpc_request *req = tgt_ses_req(tsi);
1426         int rc;
1427
1428         ENTRY;
1429         LASSERT(tsi->tsi_dlm_req);
1430         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1431         if (rc)
1432                 RETURN(err_serious(rc));
1433
1434         RETURN(req->rq_status);
1435 }
1436
1437 int tgt_bl_callback(struct tgt_session_info *tsi)
1438 {
1439         return err_serious(-EOPNOTSUPP);
1440 }
1441
1442 int tgt_cp_callback(struct tgt_session_info *tsi)
1443 {
1444         return err_serious(-EOPNOTSUPP);
1445 }
1446
1447 /* generic LDLM target handler */
1448 struct tgt_handler tgt_dlm_handlers[] = {
1449 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1450 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1451 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1452 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1453 };
1454 EXPORT_SYMBOL(tgt_dlm_handlers);
1455
1456 /*
1457  * Unified target LLOG handlers.
1458  */
1459 int tgt_llog_open(struct tgt_session_info *tsi)
1460 {
1461         int rc;
1462
1463         ENTRY;
1464
1465         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1466
1467         RETURN(rc);
1468 }
1469 EXPORT_SYMBOL(tgt_llog_open);
1470
1471 int tgt_llog_read_header(struct tgt_session_info *tsi)
1472 {
1473         int rc;
1474
1475         ENTRY;
1476
1477         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1478
1479         RETURN(rc);
1480 }
1481 EXPORT_SYMBOL(tgt_llog_read_header);
1482
1483 int tgt_llog_next_block(struct tgt_session_info *tsi)
1484 {
1485         int rc;
1486
1487         ENTRY;
1488
1489         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1490
1491         RETURN(rc);
1492 }
1493 EXPORT_SYMBOL(tgt_llog_next_block);
1494
1495 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1496 {
1497         int rc;
1498
1499         ENTRY;
1500
1501         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1502
1503         RETURN(rc);
1504 }
1505 EXPORT_SYMBOL(tgt_llog_prev_block);
1506
1507 /* generic llog target handler */
1508 struct tgt_handler tgt_llog_handlers[] = {
1509 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1510 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1511 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1512 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1513 };
1514 EXPORT_SYMBOL(tgt_llog_handlers);
1515
1516 /*
1517  * sec context handlers
1518  */
1519 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1520 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1521 {
1522         return 0;
1523 }
1524
1525 struct tgt_handler tgt_sec_ctx_handlers[] = {
1526 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1527 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1528 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1529 };
1530 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1531
1532 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1533                                  struct dt_device *key,
1534                                  struct lfsck_req_local *lrl,
1535                                  struct thandle *th) = NULL;
1536
1537 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1538                                                       struct dt_device *,
1539                                                       struct lfsck_req_local *,
1540                                                       struct thandle *))
1541 {
1542         tgt_lfsck_in_notify_local = notify;
1543 }
1544 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1545
1546 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1547                            struct dt_device *key,
1548                            struct lfsck_request *lr) = NULL;
1549
1550 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1551                                                 struct dt_device *,
1552                                                 struct lfsck_request *))
1553 {
1554         tgt_lfsck_in_notify = notify;
1555 }
1556 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1557
1558 static int (*tgt_lfsck_query)(const struct lu_env *env,
1559                               struct dt_device *key,
1560                               struct lfsck_request *req,
1561                               struct lfsck_reply *rep,
1562                               struct lfsck_query *que) = NULL;
1563
1564 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1565                                            struct dt_device *,
1566                                            struct lfsck_request *,
1567                                            struct lfsck_reply *,
1568                                            struct lfsck_query *))
1569 {
1570         tgt_lfsck_query = query;
1571 }
1572 EXPORT_SYMBOL(tgt_register_lfsck_query);
1573
1574 /* LFSCK request handlers */
1575 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1576 {
1577         const struct lu_env     *env = tsi->tsi_env;
1578         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1579         struct lfsck_request    *lr;
1580         int                      rc;
1581         ENTRY;
1582
1583         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1584         if (lr == NULL)
1585                 RETURN(-EPROTO);
1586
1587         rc = tgt_lfsck_in_notify(env, key, lr);
1588
1589         RETURN(rc);
1590 }
1591
1592 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1593 {
1594         struct lfsck_request    *request;
1595         struct lfsck_reply      *reply;
1596         int                      rc;
1597         ENTRY;
1598
1599         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1600         if (request == NULL)
1601                 RETURN(-EPROTO);
1602
1603         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1604         if (reply == NULL)
1605                 RETURN(-ENOMEM);
1606
1607         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1608                              request, reply, NULL);
1609
1610         RETURN(rc < 0 ? rc : 0);
1611 }
1612
1613 struct tgt_handler tgt_lfsck_handlers[] = {
1614 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1615 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1616 };
1617 EXPORT_SYMBOL(tgt_lfsck_handlers);
1618
1619 /*
1620  * initialize per-thread page pool (bug 5137).
1621  */
1622 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1623 {
1624         struct tgt_thread_big_cache *tbc;
1625
1626         ENTRY;
1627
1628         LASSERT(thread != NULL);
1629         LASSERT(thread->t_data == NULL);
1630
1631         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1632         if (tbc == NULL)
1633                 RETURN(-ENOMEM);
1634         thread->t_data = tbc;
1635         RETURN(0);
1636 }
1637 EXPORT_SYMBOL(tgt_io_thread_init);
1638
1639 /*
1640  * free per-thread pool created by tgt_thread_init().
1641  */
1642 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1643 {
1644         struct tgt_thread_big_cache *tbc;
1645
1646         ENTRY;
1647
1648         LASSERT(thread != NULL);
1649
1650         /*
1651          * be prepared to handle partially-initialized pools (because this is
1652          * called from ost_io_thread_init() for cleanup.
1653          */
1654         tbc = thread->t_data;
1655         if (tbc != NULL) {
1656                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1657                 thread->t_data = NULL;
1658         }
1659         EXIT;
1660 }
1661 EXPORT_SYMBOL(tgt_io_thread_done);
1662
1663 /**
1664  * Helper function for getting Data-on-MDT file server DLM lock
1665  * if asked by client.
1666  */
1667 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1668                       struct lustre_handle *lh, int mode, __u64 *flags)
1669 {
1670         union ldlm_policy_data policy = {
1671                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1672         };
1673         int rc;
1674
1675         ENTRY;
1676
1677         LASSERT(lh != NULL);
1678         LASSERT(ns != NULL);
1679         LASSERT(!lustre_handle_is_used(lh));
1680
1681         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1682                                     flags, ldlm_blocking_ast,
1683                                     ldlm_completion_ast, ldlm_glimpse_ast,
1684                                     NULL, 0, LVB_T_NONE, NULL, lh);
1685
1686         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1687 }
1688 EXPORT_SYMBOL(tgt_mdt_data_lock);
1689
1690 /**
1691  * Helper function for getting server side [start, start+count] DLM lock
1692  * if asked by client.
1693  */
1694 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1695                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1696                     struct lustre_handle *lh, int mode, __u64 *flags)
1697 {
1698         union ldlm_policy_data policy;
1699         int rc;
1700
1701         ENTRY;
1702
1703         LASSERT(lh != NULL);
1704         LASSERT(ns != NULL);
1705         LASSERT(!lustre_handle_is_used(lh));
1706
1707         policy.l_extent.gid = 0;
1708         policy.l_extent.start = start & PAGE_MASK;
1709
1710         /*
1711          * If ->o_blocks is EOF it means "lock till the end of the file".
1712          * Otherwise, it's size of an extent or hole being punched (in bytes).
1713          */
1714         if (end == OBD_OBJECT_EOF || end < start)
1715                 policy.l_extent.end = OBD_OBJECT_EOF;
1716         else
1717                 policy.l_extent.end = end | ~PAGE_MASK;
1718
1719         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1720                                     mode, flags, ldlm_blocking_ast,
1721                                     ldlm_completion_ast, ldlm_glimpse_ast,
1722                                     NULL, 0, LVB_T_NONE, NULL, lh);
1723         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1724 }
1725 EXPORT_SYMBOL(tgt_extent_lock);
1726
1727 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1728                          struct ldlm_res_id *res_id, __u64 start, __u64 end,
1729                          struct lustre_handle *lh, enum ldlm_mode mode)
1730 {
1731         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1732         __u64 flags = 0;
1733
1734         /* MDT IO for data-on-mdt */
1735         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1736                 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1737
1738         return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1739 }
1740
1741 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1742 {
1743         LASSERT(lustre_handle_is_used(lh));
1744         ldlm_lock_decref(lh, mode);
1745 }
1746 EXPORT_SYMBOL(tgt_data_unlock);
1747
1748 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1749                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1750                         struct niobuf_remote *nb, struct lustre_handle *lh,
1751                         enum ldlm_mode mode)
1752 {
1753         int nrbufs = obj->ioo_bufcnt;
1754         int i;
1755
1756         ENTRY;
1757
1758         LASSERT(mode == LCK_PR || mode == LCK_PW);
1759         LASSERT(!lustre_handle_is_used(lh));
1760
1761         if (exp->exp_obd->obd_recovering)
1762                 RETURN(0);
1763
1764         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1765                 RETURN(0);
1766
1767         for (i = 1; i < nrbufs; i++)
1768                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1769                         RETURN(-EFAULT);
1770
1771         return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1772                              nb[nrbufs - 1].rnb_offset +
1773                              nb[nrbufs - 1].rnb_len - 1, lh, mode);
1774 }
1775
1776 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1777                            struct niobuf_remote *niob,
1778                            struct lustre_handle *lh, enum ldlm_mode mode)
1779 {
1780         ENTRY;
1781
1782         LASSERT(mode == LCK_PR || mode == LCK_PW);
1783         LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1784                  niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1785                 lustre_handle_is_used(lh));
1786
1787         if (lustre_handle_is_used(lh))
1788                 tgt_data_unlock(lh, mode);
1789         EXIT;
1790 }
1791
1792 static int tgt_checksum_niobuf(struct lu_target *tgt,
1793                                  struct niobuf_local *local_nb, int npages,
1794                                  int opc, enum cksum_types cksum_type,
1795                                  __u32 *cksum)
1796 {
1797         struct ahash_request           *req;
1798         unsigned int                    bufsize;
1799         int                             i, err;
1800         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1801
1802         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1803         if (IS_ERR(req)) {
1804                 CERROR("%s: unable to initialize checksum hash %s\n",
1805                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1806                 return PTR_ERR(req);
1807         }
1808
1809         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1810         for (i = 0; i < npages; i++) {
1811                 /* corrupt the data before we compute the checksum, to
1812                  * simulate a client->OST data error */
1813                 if (i == 0 && opc == OST_WRITE &&
1814                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1815                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1816                         int len = local_nb[i].lnb_len;
1817                         struct page *np = tgt_page_to_corrupt;
1818
1819                         if (np) {
1820                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1821                                 char *ptr2 = page_address(np);
1822
1823                                 memcpy(ptr2 + off, ptr + off, len);
1824                                 memcpy(ptr2 + off, "bad3", min(4, len));
1825                                 kunmap_atomic(ptr);
1826
1827                                 /* LU-8376 to preserve original index for
1828                                  * display in dump_all_bulk_pages() */
1829                                 np->index = i;
1830
1831                                 cfs_crypto_hash_update_page(req, np, off,
1832                                                             len);
1833                                 continue;
1834                         } else {
1835                                 CERROR("%s: can't alloc page for corruption\n",
1836                                        tgt_name(tgt));
1837                         }
1838                 }
1839                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1840                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1841                                   local_nb[i].lnb_len);
1842
1843                  /* corrupt the data after we compute the checksum, to
1844                  * simulate an OST->client data error */
1845                 if (i == 0 && opc == OST_READ &&
1846                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1847                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1848                         int len = local_nb[i].lnb_len;
1849                         struct page *np = tgt_page_to_corrupt;
1850
1851                         if (np) {
1852                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1853                                 char *ptr2 = page_address(np);
1854
1855                                 memcpy(ptr2 + off, ptr + off, len);
1856                                 memcpy(ptr2 + off, "bad4", min(4, len));
1857                                 kunmap_atomic(ptr);
1858
1859                                 /* LU-8376 to preserve original index for
1860                                  * display in dump_all_bulk_pages() */
1861                                 np->index = i;
1862
1863                                 cfs_crypto_hash_update_page(req, np, off,
1864                                                             len);
1865                                 continue;
1866                         } else {
1867                                 CERROR("%s: can't alloc page for corruption\n",
1868                                        tgt_name(tgt));
1869                         }
1870                 }
1871         }
1872
1873         bufsize = sizeof(*cksum);
1874         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1875
1876         return 0;
1877 }
1878
1879 char dbgcksum_file_name[PATH_MAX];
1880
1881 static void dump_all_bulk_pages(struct obdo *oa, int count,
1882                                 struct niobuf_local *local_nb,
1883                                 __u32 server_cksum, __u32 client_cksum)
1884 {
1885         struct file *filp;
1886         int rc, i;
1887         unsigned int len;
1888         char *buf;
1889
1890         /* will only keep dump of pages on first error for the same range in
1891          * file/fid, not during the resends/retries. */
1892         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1893                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1894                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1895                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1896                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1897                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1898                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1899                  local_nb[0].lnb_file_offset,
1900                  local_nb[count-1].lnb_file_offset +
1901                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1902         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1903         filp = filp_open(dbgcksum_file_name,
1904                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1905         if (IS_ERR(filp)) {
1906                 rc = PTR_ERR(filp);
1907                 if (rc == -EEXIST)
1908                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1909                                "checksum error: rc = %d\n", dbgcksum_file_name,
1910                                rc);
1911                 else
1912                         CERROR("%s: can't open to dump pages with checksum "
1913                                "error: rc = %d\n", dbgcksum_file_name, rc);
1914                 return;
1915         }
1916
1917         for (i = 0; i < count; i++) {
1918                 len = local_nb[i].lnb_len;
1919                 buf = kmap(local_nb[i].lnb_page);
1920                 while (len != 0) {
1921                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1922                         if (rc < 0) {
1923                                 CERROR("%s: wanted to write %u but got %d "
1924                                        "error\n", dbgcksum_file_name, len, rc);
1925                                 break;
1926                         }
1927                         len -= rc;
1928                         buf += rc;
1929                 }
1930                 kunmap(local_nb[i].lnb_page);
1931         }
1932
1933         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1934         if (rc)
1935                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1936         filp_close(filp, NULL);
1937
1938         libcfs_debug_dumplog();
1939 }
1940
1941 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1942                                struct obd_export *exp, struct obdo *oa,
1943                                const struct lnet_process_id *peer,
1944                                __u32 client_cksum, __u32 server_cksum,
1945                                enum cksum_types server_cksum_type)
1946 {
1947         char *msg;
1948         enum cksum_types cksum_type;
1949         loff_t start, end;
1950
1951         /* unlikely to happen and only if resend does not occur due to cksum
1952          * control failure on Client */
1953         if (unlikely(server_cksum == client_cksum)) {
1954                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1955                        client_cksum);
1956                 return 0;
1957         }
1958
1959         if (exp->exp_obd->obd_checksum_dump)
1960                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1961                                     client_cksum);
1962
1963         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1964                                            oa->o_flags : 0);
1965
1966         if (cksum_type != server_cksum_type)
1967                 msg = "the server may have not used the checksum type specified"
1968                       " in the original request - likely a protocol problem";
1969         else
1970                 msg = "should have changed on the client or in transit";
1971
1972         start = local_nb[0].lnb_file_offset;
1973         end = local_nb[npages-1].lnb_file_offset +
1974                                         local_nb[npages-1].lnb_len - 1;
1975
1976         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1977                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1978                 " %x (type %x), server csum %x (type %x)\n",
1979                 exp->exp_obd->obd_name,
1980                 msg, libcfs_nid2str(peer->nid),
1981                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1982                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1983                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1984                 POSTID(&oa->o_oi),
1985                 start, end, client_cksum, cksum_type, server_cksum,
1986                 server_cksum_type);
1987
1988         return 1;
1989 }
1990
1991 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1992                              unsigned char *buf, int size)
1993 {
1994         int     i, off, len, copied = size;
1995         char    *ptr;
1996
1997         for (i = 0; i < npages; i++) {
1998                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1999                 len = local[i].lnb_len;
2000
2001                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2002                        i, off, len, size);
2003                 if (len > size)
2004                         return -EINVAL;
2005
2006                 ptr = kmap_atomic(local[i].lnb_page);
2007                 memcpy(buf, ptr + off, len);
2008                 kunmap_atomic(ptr);
2009                 buf += len;
2010                 size -= len;
2011         }
2012         return copied - size;
2013 }
2014
2015 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2016                                      struct niobuf_local *local_nb, int npages,
2017                                      int opc, obd_dif_csum_fn *fn,
2018                                      int sector_size, u32 *check_sum,
2019                                      bool resend)
2020 {
2021         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2022         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2023         const char *obd_name = tgt->lut_obd->obd_name;
2024         struct ahash_request *req;
2025         unsigned int bufsize;
2026         unsigned char *buffer;
2027         struct page *__page;
2028         __u16 *guard_start;
2029         int guard_number;
2030         int used_number = 0;
2031         __u32 cksum;
2032         int rc = 0;
2033         int used;
2034         int i;
2035
2036         __page = alloc_page(GFP_KERNEL);
2037         if (__page == NULL)
2038                 return -ENOMEM;
2039
2040         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2041         if (IS_ERR(req)) {
2042                 CERROR("%s: unable to initialize checksum hash %s\n",
2043                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2044                 return PTR_ERR(req);
2045         }
2046
2047         buffer = kmap(__page);
2048         guard_start = (__u16 *)buffer;
2049         guard_number = PAGE_SIZE / sizeof(*guard_start);
2050         if (unlikely(resend))
2051                 CDEBUG(D_PAGE | D_HA, "GRD tags per page = %u\n", guard_number);
2052         for (i = 0; i < npages; i++) {
2053                 bool use_t10_grd;
2054
2055                 /* corrupt the data before we compute the checksum, to
2056                  * simulate a client->OST data error */
2057                 if (i == 0 && opc == OST_WRITE &&
2058                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2059                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2060                         int len = local_nb[i].lnb_len;
2061                         struct page *np = tgt_page_to_corrupt;
2062
2063                         if (np) {
2064                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2065                                 char *ptr2 = page_address(np);
2066
2067                                 memcpy(ptr2 + off, ptr + off, len);
2068                                 memcpy(ptr2 + off, "bad3", min(4, len));
2069                                 kunmap_atomic(ptr);
2070
2071                                 /* LU-8376 to preserve original index for
2072                                  * display in dump_all_bulk_pages() */
2073                                 np->index = i;
2074
2075                                 cfs_crypto_hash_update_page(req, np, off,
2076                                                             len);
2077                                 continue;
2078                         } else {
2079                                 CERROR("%s: can't alloc page for corruption\n",
2080                                        tgt_name(tgt));
2081                         }
2082                 }
2083
2084                 /*
2085                  * The left guard number should be able to hold checksums of a
2086                  * whole page
2087                  */
2088                 use_t10_grd = t10_cksum_type && opc == OST_READ &&
2089                               local_nb[i].lnb_len == PAGE_SIZE &&
2090                               local_nb[i].lnb_guard_disk;
2091                 if (use_t10_grd) {
2092                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2093                         if (used > (guard_number - used_number)) {
2094                                 rc = -E2BIG;
2095                                 break;
2096                         }
2097                         memcpy(guard_start + used_number,
2098                                local_nb[i].lnb_guards,
2099                                used * sizeof(*guard_start));
2100                         if (unlikely(resend))
2101                                 CDEBUG(D_PAGE | D_HA,
2102                                        "lnb[%u]: used %u off %u+%u lnb checksum: %*phN\n",
2103                                        i, used,
2104                                        local_nb[i].lnb_page_offset,
2105                                        local_nb[i].lnb_len,
2106                                        (int)(used * sizeof(*guard_start)),
2107                                        guard_start + used_number);
2108                 }
2109                 if (!use_t10_grd || unlikely(resend)) {
2110                         __u16 guard_tmp[MAX_GUARD_NUMBER];
2111                         __u16 *guards = guard_start + used_number;
2112                         int used_tmp = -1, *usedp = &used;
2113
2114                         if (unlikely(use_t10_grd)) {
2115                                 guards = guard_tmp;
2116                                 usedp = &used_tmp;
2117                         }
2118                         rc = obd_page_dif_generate_buffer(obd_name,
2119                                 local_nb[i].lnb_page,
2120                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2121                                 local_nb[i].lnb_len, guards,
2122                                 guard_number - used_number, usedp, sector_size,
2123                                 fn);
2124                         if (unlikely(resend)) {
2125                                 bool bad = use_t10_grd &&
2126                                         memcmp(guard_tmp,
2127                                                local_nb[i].lnb_guards,
2128                                                used_tmp * sizeof(*guard_tmp));
2129
2130                                 if (bad)
2131                                         CERROR("lnb[%u]: used %u/%u off %u+%u tmp checksum: %*phN\n",
2132                                                i, used, used_tmp,
2133                                                local_nb[i].lnb_page_offset,
2134                                                local_nb[i].lnb_len,
2135                                                (int)(used_tmp * sizeof(*guard_start)),
2136                                                guard_tmp);
2137                                 CDEBUG_LIMIT(D_PAGE | D_HA | (bad ? D_ERROR : 0),
2138                                        "lnb[%u]: used %u/%u off %u+%u gen checksum: %*phN\n",
2139                                        i, used, used_tmp,
2140                                        local_nb[i].lnb_page_offset,
2141                                        local_nb[i].lnb_len,
2142                                        (int)(used * sizeof(*guard_start)),
2143                                        guard_start + used_number);
2144                         }
2145                         if (rc)
2146                                 break;
2147                 }
2148
2149                 LASSERT(used <= MAX_GUARD_NUMBER);
2150                 /*
2151                  * If disk support T10PI checksum, copy guards to local_nb.
2152                  * If the write is partial page, do not use the guards for bio
2153                  * submission since the data might not be full-sector. The bio
2154                  * guards will be generated later based on the full sectors. If
2155                  * the sector size is 512B rather than 4 KB, or the page size
2156                  * is larger than 4KB, this might drop some useful guards for
2157                  * partial page write, but it will only add minimal extra time
2158                  * of checksum calculation.
2159                  */
2160                 if (t10_cksum_type && opc == OST_WRITE &&
2161                     local_nb[i].lnb_len == PAGE_SIZE) {
2162                         local_nb[i].lnb_guard_rpc = 1;
2163                         memcpy(local_nb[i].lnb_guards,
2164                                guard_start + used_number,
2165                                used * sizeof(*local_nb[i].lnb_guards));
2166                 }
2167
2168                 used_number += used;
2169                 if (used_number == guard_number) {
2170                         cfs_crypto_hash_update_page(req, __page, 0,
2171                                 used_number * sizeof(*guard_start));
2172                         used_number = 0;
2173                 }
2174
2175                  /* corrupt the data after we compute the checksum, to
2176                  * simulate an OST->client data error */
2177                 if (unlikely(i == 0 && opc == OST_READ &&
2178                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2179                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2180                         int len = local_nb[i].lnb_len;
2181                         struct page *np = tgt_page_to_corrupt;
2182
2183                         if (np) {
2184                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2185                                 char *ptr2 = page_address(np);
2186
2187                                 memcpy(ptr2 + off, ptr + off, len);
2188                                 memcpy(ptr2 + off, "bad4", min(4, len));
2189                                 kunmap_atomic(ptr);
2190
2191                                 /* LU-8376 to preserve original index for
2192                                  * display in dump_all_bulk_pages() */
2193                                 np->index = i;
2194
2195                                 cfs_crypto_hash_update_page(req, np, off,
2196                                                             len);
2197                                 continue;
2198                         } else {
2199                                 CERROR("%s: can't alloc page for corruption\n",
2200                                        tgt_name(tgt));
2201                         }
2202                 }
2203         }
2204         kunmap(__page);
2205         if (rc)
2206                 GOTO(out, rc);
2207
2208         if (used_number != 0)
2209                 cfs_crypto_hash_update_page(req, __page, 0,
2210                         used_number * sizeof(*guard_start));
2211
2212         bufsize = sizeof(cksum);
2213         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2214
2215         if (rc == 0)
2216                 *check_sum = cksum;
2217 out:
2218         __free_page(__page);
2219         return rc;
2220 }
2221
2222 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2223                                   enum cksum_types cksum_type,
2224                                   struct niobuf_local *local_nb,
2225                                   int npages, int opc, u32 *check_sum,
2226                                   bool resend)
2227 {
2228         obd_dif_csum_fn *fn = NULL;
2229         int sector_size = 0;
2230         int rc;
2231
2232         ENTRY;
2233         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2234
2235         if (fn)
2236                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2237                                                opc, fn, sector_size,
2238                                                check_sum, resend);
2239         else
2240                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2241                                          cksum_type, check_sum);
2242
2243         RETURN(rc);
2244 }
2245
2246 int tgt_brw_read(struct tgt_session_info *tsi)
2247 {
2248         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2249         struct ptlrpc_bulk_desc *desc = NULL;
2250         struct obd_export       *exp = tsi->tsi_exp;
2251         struct niobuf_remote    *remote_nb;
2252         struct niobuf_local     *local_nb;
2253         struct obd_ioobj        *ioo;
2254         struct ost_body         *body, *repbody;
2255         struct lustre_handle     lockh = { 0 };
2256         int                      npages, nob = 0, rc, i, no_reply = 0,
2257                                  npages_read;
2258         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2259         const char *obd_name = exp->exp_obd->obd_name;
2260
2261         ENTRY;
2262
2263         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2264             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2265                 CERROR("%s: deny read request from %s to portal %u\n",
2266                        tgt_name(tsi->tsi_tgt),
2267                        obd_export_nid2str(req->rq_export),
2268                        ptlrpc_req2svc(req)->srv_req_portal);
2269                 RETURN(-EPROTO);
2270         }
2271
2272         req->rq_bulk_read = 1;
2273
2274         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) {
2275                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2276                  * this server to fail requests.
2277                  */
2278                 char fail_ost_name[MAX_OBD_NAME];
2279
2280                 if (cfs_fail_val > 0) {
2281                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2282                                  cfs_fail_val - 1);
2283
2284                         if (strstr(obd_name, fail_ost_name))
2285                                 RETURN(err_serious(-EIO));
2286                 } else {
2287                         RETURN(err_serious(-EIO));
2288                 }
2289         }
2290
2291         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2292                          cfs_fail_val : (obd_timeout + 1) / 4);
2293
2294         /* Check if there is eviction in progress, and if so, wait for it to
2295          * finish */
2296         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2297                 /* We do not care how long it takes */
2298                 wait_event_idle(
2299                         exp->exp_obd->obd_evict_inprogress_waitq,
2300                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2301         }
2302
2303         /* There must be big cache in current thread to process this request
2304          * if it is NULL then something went wrong and it wasn't allocated,
2305          * report -ENOMEM in that case */
2306         if (tbc == NULL)
2307                 RETURN(-ENOMEM);
2308
2309         body = tsi->tsi_ost_body;
2310         LASSERT(body != NULL);
2311
2312         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2313             body->oa.o_flags & OBD_FL_NORPC)
2314                 RETURN(0);
2315
2316         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2317         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2318
2319         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2320         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2321
2322         local_nb = tbc->local;
2323
2324         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2325                           &lockh, LCK_PR);
2326         if (rc != 0)
2327                 RETURN(rc);
2328
2329         /*
2330          * If getting the lock took more time than
2331          * client was willing to wait, drop it. b=11330
2332          */
2333         if (ktime_get_real_seconds() > req->rq_deadline ||
2334             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2335                 no_reply = 1;
2336                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2337                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2338                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2339                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2340                 GOTO(out_lock, rc = -ETIMEDOUT);
2341         }
2342
2343         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2344         repbody->oa = body->oa;
2345
2346         npages = PTLRPC_MAX_BRW_PAGES;
2347         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2348                         ioo, remote_nb, &npages, local_nb);
2349         if (rc != 0)
2350                 GOTO(out_lock, rc);
2351
2352         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2353             body->oa.o_flags & OBD_FL_SHORT_IO) {
2354                 desc = NULL;
2355         } else {
2356                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2357                                             PTLRPC_BULK_PUT_SOURCE,
2358                                             OST_BULK_PORTAL,
2359                                             &ptlrpc_bulk_kiov_nopin_ops);
2360                 if (desc == NULL)
2361                         GOTO(out_commitrw, rc = -ENOMEM);
2362         }
2363
2364         nob = 0;
2365         npages_read = npages;
2366         for (i = 0; i < npages; i++) {
2367                 int page_rc = local_nb[i].lnb_rc;
2368
2369                 if (page_rc < 0) {
2370                         rc = page_rc;
2371                         npages_read = i;
2372                         break;
2373                 }
2374
2375                 nob += page_rc;
2376                 if (page_rc != 0 && desc != NULL) { /* some data! */
2377                         LASSERT(local_nb[i].lnb_page != NULL);
2378                         desc->bd_frag_ops->add_kiov_frag
2379                           (desc, local_nb[i].lnb_page,
2380                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2381                            page_rc);
2382                 }
2383
2384                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2385                         local_nb[i].lnb_len = page_rc;
2386                         npages_read = i + (page_rc != 0 ? 1 : 0);
2387                         /* All subsequent pages should be 0 */
2388                         while (++i < npages)
2389                                 LASSERT(local_nb[i].lnb_rc == 0);
2390                         break;
2391                 }
2392         }
2393
2394         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2395                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2396                            body->oa.o_flags : 0;
2397                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2398                 bool resend = (body->oa.o_valid & OBD_MD_FLFLAGS) &&
2399                         (body->oa.o_flags & OBD_FL_RECOV_RESEND);
2400
2401                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2402                                                           cksum_type);
2403                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2404
2405                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2406                                             local_nb, npages_read, OST_READ,
2407                                             &repbody->oa.o_cksum, resend);
2408                 if (rc < 0)
2409                         GOTO(out_commitrw, rc);
2410                 CDEBUG(D_PAGE | (resend ? D_HA : 0),
2411                        "checksum at read origin: %x (%x)\n",
2412                        repbody->oa.o_cksum, cksum_type);
2413
2414                 /* if a resend it could be for a cksum error, so check Server
2415                  * cksum with returned Client cksum (this should even cover
2416                  * zero-cksum case) */
2417                 if (resend)
2418                         check_read_checksum(local_nb, npages_read, exp,
2419                                             &body->oa, &req->rq_peer,
2420                                             body->oa.o_cksum,
2421                                             repbody->oa.o_cksum, cksum_type);
2422         } else {
2423                 repbody->oa.o_valid = 0;
2424         }
2425         /* We're finishing using body->oa as an input variable */
2426
2427         /* Check if client was evicted while we were doing i/o before touching
2428          * network */
2429         if (rc == 0) {
2430                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2431                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2432                         unsigned char *short_io_buf;
2433                         int short_io_size;
2434
2435                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2436                                                               &RMF_SHORT_IO);
2437                         short_io_size = req_capsule_get_size(&req->rq_pill,
2438                                                              &RMF_SHORT_IO,
2439                                                              RCL_SERVER);
2440                         rc = tgt_pages2shortio(local_nb, npages_read,
2441                                                short_io_buf, short_io_size);
2442                         if (rc >= 0)
2443                                 req_capsule_shrink(&req->rq_pill,
2444                                                    &RMF_SHORT_IO, rc,
2445                                                    RCL_SERVER);
2446                         rc = rc > 0 ? 0 : rc;
2447                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2448                         rc = target_bulk_io(exp, desc);
2449                 }
2450                 no_reply = rc != 0;
2451         } else {
2452                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2453                     body->oa.o_flags & OBD_FL_SHORT_IO)
2454                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2455                                            RCL_SERVER);
2456         }
2457
2458 out_commitrw:
2459         /* Must commit after prep above in all cases */
2460         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2461                           remote_nb, npages, local_nb, rc);
2462 out_lock:
2463         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2464
2465         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2466                 ptlrpc_free_bulk(desc);
2467
2468         LASSERT(rc <= 0);
2469         if (rc == 0) {
2470                 rc = nob;
2471                 ptlrpc_lprocfs_brw(req, nob);
2472         } else if (no_reply) {
2473                 req->rq_no_reply = 1;
2474                 /* reply out callback would free */
2475                 ptlrpc_req_drop_rs(req);
2476                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2477                               "client will retry: rc %d\n",
2478                               obd_name,
2479                               obd_uuid2str(&exp->exp_client_uuid),
2480                               obd_export_nid2str(exp), rc);
2481         }
2482         /* send a bulk after reply to simulate a network delay or reordering
2483          * by a router - Note that !desc implies short io, so there is no bulk
2484          * to reorder. */
2485         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2486             desc) {
2487                 /* Calculate checksum before request transfer, original
2488                  * it is done by target_bulk_io() */
2489                 rc = sptlrpc_svc_wrap_bulk(req, desc);
2490                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
2491                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2492                 else /* old version, bulk matchbits is rq_xid */
2493                         req->rq_mbits = req->rq_xid;
2494
2495                 req->rq_status = rc;
2496                 target_committed_to_req(req);
2497                 target_send_reply(req, 0, 0);
2498
2499                 CDEBUG(D_INFO, "reorder BULK\n");
2500                 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
2501                                  cfs_fail_val ? : 3);
2502
2503                 target_bulk_io(exp, desc);
2504                 ptlrpc_free_bulk(desc);
2505         }
2506
2507         RETURN(rc);
2508 }
2509 EXPORT_SYMBOL(tgt_brw_read);
2510
2511 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2512                              unsigned char *buf, unsigned int size)
2513 {
2514         int     i, off, len;
2515         char    *ptr;
2516
2517         for (i = 0; i < npages; i++) {
2518                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2519                 len = local[i].lnb_len;
2520
2521                 if (len == 0)
2522                         continue;
2523
2524                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2525                        i, off, len, size);
2526                 ptr = kmap_atomic(local[i].lnb_page);
2527                 if (ptr == NULL)
2528                         return -EINVAL;
2529                 memcpy(ptr + off, buf, len < size ? len : size);
2530                 kunmap_atomic(ptr);
2531                 buf += len;
2532                 size -= len;
2533         }
2534         return 0;
2535 }
2536
2537 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2538                               struct ptlrpc_bulk_desc *desc,
2539                               struct niobuf_local *local_nb, int npages,
2540                               u32 client_cksum, u32 server_cksum,
2541                               bool mmap)
2542 {
2543         struct obd_export *exp = req->rq_export;
2544         struct ost_body *body;
2545         char *router = "";
2546         char *via = "";
2547
2548         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2549         LASSERT(body != NULL);
2550
2551         if (desc && req->rq_peer.nid != desc->bd_sender) {
2552                 via = " via ";
2553                 router = libcfs_nid2str(desc->bd_sender);
2554         }
2555
2556         if (exp->exp_obd->obd_checksum_dump)
2557                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2558                                     client_cksum);
2559
2560         if (mmap) {
2561                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2562                              client_cksum, server_cksum);
2563                 return;
2564         }
2565
2566         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2567                            DFID" object "DOSTID" extent [%llu-%llu"
2568                            "]: client csum %x, server csum %x\n",
2569                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2570                            via, router,
2571                            body->oa.o_valid & OBD_MD_FLFID ?
2572                            body->oa.o_parent_seq : (__u64)0,
2573                            body->oa.o_valid & OBD_MD_FLFID ?
2574                            body->oa.o_parent_oid : 0,
2575                            body->oa.o_valid & OBD_MD_FLFID ?
2576                            body->oa.o_parent_ver : 0,
2577                            POSTID(&body->oa.o_oi),
2578                            local_nb[0].lnb_file_offset,
2579                            local_nb[npages-1].lnb_file_offset +
2580                            local_nb[npages - 1].lnb_len - 1,
2581                            client_cksum, server_cksum);
2582 }
2583
2584 int tgt_brw_write(struct tgt_session_info *tsi)
2585 {
2586         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2587         struct ptlrpc_bulk_desc *desc = NULL;
2588         struct obd_export       *exp = req->rq_export;
2589         struct niobuf_remote    *remote_nb;
2590         struct niobuf_local     *local_nb;
2591         struct obd_ioobj        *ioo;
2592         struct ost_body         *body, *repbody;
2593         struct lustre_handle     lockh = {0};
2594         __u32                   *rcs;
2595         int                      objcount, niocount, npages;
2596         int                      rc = 0;
2597         int                      i, j;
2598         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2599         bool                     no_reply = false, mmap;
2600         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2601         bool wait_sync = false;
2602         const char *obd_name = exp->exp_obd->obd_name;
2603         /* '1' for consistency with code that checks !mpflag to restore */
2604         unsigned int mpflags = 1;
2605
2606         ENTRY;
2607
2608         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2609             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2610                 CERROR("%s: deny write request from %s to portal %u\n",
2611                        tgt_name(tsi->tsi_tgt),
2612                        obd_export_nid2str(req->rq_export),
2613                        ptlrpc_req2svc(req)->srv_req_portal);
2614                 RETURN(err_serious(-EPROTO));
2615         }
2616
2617         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2618                 RETURN(err_serious(-ENOSPC));
2619         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2620                 RETURN(err_serious(-EROFS));
2621
2622         req->rq_bulk_write = 1;
2623
2624         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2625                 rc = -EIO;
2626         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2627                 rc = -EFAULT;
2628         if (rc < 0) {
2629                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2630                  * this server to fail requests.
2631                  */
2632                 char fail_ost_name[MAX_OBD_NAME];
2633
2634                 if (cfs_fail_val > 0) {
2635                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2636                                  cfs_fail_val - 1);
2637
2638                         if (strstr(obd_name, fail_ost_name))
2639                                 RETURN(err_serious(rc));
2640                 } else {
2641                         RETURN(err_serious(rc));
2642                 }
2643         }
2644
2645         /* pause before transaction has been started */
2646         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2647                          cfs_fail_val : (obd_timeout + 1) / 4);
2648
2649         /* Delay write commit to show stale size information */
2650         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2651
2652         /* There must be big cache in current thread to process this request
2653          * if it is NULL then something went wrong and it wasn't allocated,
2654          * report -ENOMEM in that case */
2655         if (tbc == NULL)
2656                 RETURN(-ENOMEM);
2657
2658         body = tsi->tsi_ost_body;
2659         LASSERT(body != NULL);
2660
2661         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2662             body->oa.o_flags & OBD_FL_NORPC)
2663                 RETURN(0);
2664
2665
2666         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2667         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2668
2669         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2670                                         RCL_CLIENT) / sizeof(*ioo);
2671
2672         for (niocount = i = 0; i < objcount; i++)
2673                 niocount += ioo[i].ioo_bufcnt;
2674
2675         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2676         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2677         if (niocount != req_capsule_get_size(&req->rq_pill,
2678                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2679                         sizeof(*remote_nb))
2680                 RETURN(err_serious(-EPROTO));
2681
2682         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2683             ptlrpc_connection_is_local(exp->exp_connection))
2684                 mpflags = memalloc_noreclaim_save();
2685
2686         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2687                              niocount * sizeof(*rcs));
2688         rc = req_capsule_server_pack(&req->rq_pill);
2689         if (rc != 0)
2690                 GOTO(out, rc = err_serious(rc));
2691
2692         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2693         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2694
2695         local_nb = tbc->local;
2696
2697         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2698                           &lockh, LCK_PW);
2699         if (rc != 0)
2700                 GOTO(out, rc);
2701
2702         /*
2703          * If getting the lock took more time than
2704          * client was willing to wait, drop it. b=11330
2705          */
2706         if (ktime_get_real_seconds() > req->rq_deadline ||
2707             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2708                 no_reply = true;
2709                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2710                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2711                        POSTID(&ioo->ioo_oid),
2712                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2713                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2714                 GOTO(out_lock, rc = -ETIMEDOUT);
2715         }
2716
2717         /* Because we already sync grant info with client when reconnect,
2718          * grant info will be cleared for resent req, then fed_grant and
2719          * total_grant will not be modified in following preprw_write */
2720         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2721                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2722                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2723         }
2724
2725         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2726         if (repbody == NULL)
2727                 GOTO(out_lock, rc = -ENOMEM);
2728         repbody->oa = body->oa;
2729
2730         npages = PTLRPC_MAX_BRW_PAGES;
2731         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2732                         objcount, ioo, remote_nb, &npages, local_nb);
2733         if (rc < 0)
2734                 GOTO(out_lock, rc);
2735         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2736             body->oa.o_flags & OBD_FL_SHORT_IO) {
2737                 unsigned int short_io_size;
2738                 unsigned char *short_io_buf;
2739
2740                 short_io_size = req_capsule_get_size(&req->rq_pill,
2741                                                      &RMF_SHORT_IO,
2742                                                      RCL_CLIENT);
2743                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2744                                                       &RMF_SHORT_IO);
2745                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2746                                " size = %d\n", short_io_size);
2747
2748                 /* Copy short io buf to pages */
2749                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2750                                        short_io_size);
2751                 desc = NULL;
2752         } else {
2753                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2754                                             PTLRPC_BULK_GET_SINK,
2755                                             OST_BULK_PORTAL,
2756                                             &ptlrpc_bulk_kiov_nopin_ops);
2757                 if (desc == NULL)
2758                         GOTO(skip_transfer, rc = -ENOMEM);
2759
2760                 /* NB Having prepped, we must commit... */
2761                 for (i = 0; i < npages; i++)
2762                         desc->bd_frag_ops->add_kiov_frag(desc,
2763                                         local_nb[i].lnb_page,
2764                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2765                                         local_nb[i].lnb_len);
2766
2767                 rc = sptlrpc_svc_prep_bulk(req, desc);
2768                 if (rc != 0)
2769                         GOTO(skip_transfer, rc);
2770
2771                 rc = target_bulk_io(exp, desc);
2772         }
2773
2774         no_reply = rc != 0;
2775
2776 skip_transfer:
2777         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2778                 static int cksum_counter;
2779
2780                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2781                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2782
2783                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2784                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2785                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2786                                                            cksum_type);
2787
2788                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2789                                             local_nb, npages, OST_WRITE,
2790                                             &repbody->oa.o_cksum, false);
2791                 if (rc < 0)
2792                         GOTO(out_commitrw, rc);
2793
2794                 cksum_counter++;
2795
2796                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2797                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2798                                 body->oa.o_flags & OBD_FL_MMAP);
2799
2800                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2801                                           body->oa.o_cksum,
2802                                           repbody->oa.o_cksum, mmap);
2803                         cksum_counter = 0;
2804                 } else if ((cksum_counter & (-cksum_counter)) ==
2805                            cksum_counter) {
2806                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2807                                cksum_counter, libcfs_id2str(req->rq_peer),
2808                                repbody->oa.o_cksum);
2809                 }
2810         }
2811
2812         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2813
2814 out_commitrw:
2815         /* Must commit after prep above in all cases */
2816         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2817                           objcount, ioo, remote_nb, npages, local_nb, rc);
2818         if (rc == -ENOTCONN)
2819                 /* quota acquire process has been given up because
2820                  * either the client has been evicted or the client
2821                  * has timed out the request already */
2822                 no_reply = true;
2823
2824         for (i = 0; i < niocount; i++) {
2825                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2826                         wait_sync = true;
2827                         break;
2828                 }
2829         }
2830         /*
2831          * Disable sending mtime back to the client. If the client locked the
2832          * whole object, then it has already updated the mtime on its side,
2833          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2834          */
2835         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2836
2837         if (rc == 0) {
2838                 int nob = 0;
2839
2840                 /* set per-requested niobuf return codes */
2841                 for (i = j = 0; i < niocount; i++) {
2842                         int len = remote_nb[i].rnb_len;
2843
2844                         nob += len;
2845                         rcs[i] = 0;
2846                         do {
2847                                 LASSERT(j < npages);
2848                                 if (local_nb[j].lnb_rc < 0)
2849                                         rcs[i] = local_nb[j].lnb_rc;
2850                                 len -= local_nb[j].lnb_len;
2851                                 j++;
2852                         } while (len > 0);
2853                         LASSERT(len == 0);
2854                 }
2855                 LASSERT(j == npages);
2856                 ptlrpc_lprocfs_brw(req, nob);
2857         }
2858 out_lock:
2859         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2860         if (desc)
2861                 ptlrpc_free_bulk(desc);
2862 out:
2863         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2864                 req->rq_no_reply = 1;
2865                 /* reply out callback would free */
2866                 ptlrpc_req_drop_rs(req);
2867                 if (!exp->exp_obd->obd_no_transno)
2868                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2869                                       " client will retry: rc = %d\n",
2870                                       obd_name,
2871                                       obd_uuid2str(&exp->exp_client_uuid),
2872                                       obd_export_nid2str(exp), rc);
2873         }
2874
2875         if (mpflags)
2876                 memalloc_noreclaim_restore(mpflags);
2877
2878         RETURN(rc);
2879 }
2880 EXPORT_SYMBOL(tgt_brw_write);
2881
2882 /**
2883  * Common request handler for OST_SEEK RPC.
2884  *
2885  * Unified request handling for OST_SEEK RPC.
2886  * It takes object by its FID, does needed lseek and packs result
2887  * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2888  *
2889  * \param[in] tsi       target session environment for this request
2890  *
2891  * \retval              0 if successful
2892  * \retval              negative value on error
2893  */
2894 int tgt_lseek(struct tgt_session_info *tsi)
2895 {
2896         struct lustre_handle lh = { 0 };
2897         struct dt_object *dob;
2898         struct ost_body *repbody;
2899         loff_t offset = tsi->tsi_ost_body->oa.o_size;
2900         int whence = tsi->tsi_ost_body->oa.o_mode;
2901         bool srvlock;
2902         int rc = 0;
2903
2904         ENTRY;
2905
2906         if (whence != SEEK_HOLE && whence != SEEK_DATA)
2907                 RETURN(-EPROTO);
2908
2909         /* Negative offset is prohibited on wire and must be handled on client
2910          * prior sending RPC.
2911          */
2912         if (offset < 0)
2913                 RETURN(-EPROTO);
2914
2915         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2916         if (repbody == NULL)
2917                 RETURN(-ENOMEM);
2918         repbody->oa = tsi->tsi_ost_body->oa;
2919
2920         srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
2921                   tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
2922         if (srvlock) {
2923                 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
2924                                    offset, OBD_OBJECT_EOF, &lh, LCK_PR);
2925                 if (rc)
2926                         RETURN(rc);
2927         }
2928
2929         dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
2930         if (IS_ERR(dob))
2931                 GOTO(out, rc = PTR_ERR(dob));
2932
2933         if (!dt_object_exists(dob))
2934                 GOTO(obj_put, rc = -ENOENT);
2935
2936         repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
2937         rc = 0;
2938 obj_put:
2939         dt_object_put(tsi->tsi_env, dob);
2940 out:
2941         if (srvlock)
2942                 tgt_data_unlock(&lh, LCK_PR);
2943
2944         RETURN(rc);
2945 }
2946 EXPORT_SYMBOL(tgt_lseek);
2947
2948 /* Check if request can be reconstructed from saved reply data
2949  * A copy of the reply data is returned in @trd if the pointer is not NULL
2950  */
2951 int req_can_reconstruct(struct ptlrpc_request *req,
2952                          struct tg_reply_data *trd)
2953 {
2954         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2955         struct lsd_client_data *lcd = ted->ted_lcd;
2956         int found;
2957
2958         if (tgt_is_multimodrpcs_client(req->rq_export))
2959                 return tgt_lookup_reply(req, trd);
2960
2961         mutex_lock(&ted->ted_lcd_lock);
2962         found = req->rq_xid == lcd->lcd_last_xid ||
2963                 req->rq_xid == lcd->lcd_last_close_xid;
2964
2965         if (found && trd != NULL) {
2966                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2967                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2968                         trd->trd_reply.lrd_transno =
2969                                                 lcd->lcd_last_close_transno;
2970                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2971                 } else {
2972                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2973                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2974                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2975                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2976                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2977                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2978                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2979                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2980                 }
2981         }
2982         mutex_unlock(&ted->ted_lcd_lock);
2983
2984         return found;
2985 }
2986 EXPORT_SYMBOL(req_can_reconstruct);
2987