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