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 #ifdef HAVE_UIDGID_HEADER
  40 # include <linux/uidgid.h>
  41 #endif
  42
  43 #include <obd.h>
  44 #include <obd_class.h>
  45 #include <obd_cksum.h>
  46 #include <lustre_lfsck.h>
  47 #include <lustre_nodemap.h>
  48 #include <lustre_acl.h>
  49
  50 #include "tgt_internal.h"
  51
  52 char *tgt_name(struct lu_target *tgt)
  53 {
  54         LASSERT(tgt->lut_obd != NULL);
  55         return tgt->lut_obd->obd_name;
  56 }
  57 EXPORT_SYMBOL(tgt_name);
  58
  59 /*
  60  * Generic code handling requests that have struct mdt_body passed in:
  61  *
  62  *  - extract mdt_body from request and save it in @tsi, if present;
  63  *
  64  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  65  *  @tsi;
  66  *
  67  *  - if HAS_BODY flag is set for this request type check whether object
  68  *  actually exists on storage (lu_object_exists()).
  69  *
  70  */
  71 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  72 {
  73         const struct mdt_body   *body;
  74         struct lu_object        *obj;
  75         struct req_capsule      *pill = tsi->tsi_pill;
  76         int                      rc;
  77
  78         ENTRY;
  79
  80         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  81         if (body == NULL)
  82                 RETURN(-EFAULT);
  83
  84         tsi->tsi_mdt_body = body;
  85
  86         if (!(body->mbo_valid & OBD_MD_FLID))
  87                 RETURN(0);
  88
  89         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  90          * in any case in pre-2.5 clients. Fix that here if needed */
  91         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  92                 RETURN(0);
  93
  94         if (!fid_is_sane(&body->mbo_fid1)) {
  95                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  96                        PFID(&body->mbo_fid1));
  97                 RETURN(-EINVAL);
  98         }
  99
 100         obj = lu_object_find(tsi->tsi_env,
 101                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 102                              &body->mbo_fid1, NULL);
 103         if (!IS_ERR(obj)) {
 104                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 105                         lu_object_put(tsi->tsi_env, obj);
 106                         rc = -ENOENT;
 107                 } else {
 108                         tsi->tsi_corpus = obj;
 109                         rc = 0;
 110                 }
 111         } else {
 112                 rc = PTR_ERR(obj);
 113         }
 114
 115         tsi->tsi_fid = body->mbo_fid1;
 116
 117         RETURN(rc);
 118 }
 119
 120 /**
 121  * Validate oa from client.
 122  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 123  * req are valid.
 124  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 125  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 126  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 127  *       whether this is oi_fid or real ostid. So it will check
 128  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 129  *    c. Old FID-disable osc will send IDIF.
 130  *    d. new FID-enable osc/osp will send normal FID.
 131  *
 132  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 133  * be used for LAST_ID file, and only being accessed inside OST now.
 134  */
 135 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 136 {
 137         struct ost_id   *oi     = &oa->o_oi;
 138         u64              seq    = ostid_seq(oi);
 139         u64              id     = ostid_id(oi);
 140         int              rc;
 141         ENTRY;
 142
 143         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 144                      fid_seq_is_echo(seq))) {
 145                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 146                  * during create, and we will reset this to 1, since this
 147                  * oi_id is basically useless in the following create process,
 148                  * but oi_id == 0 will make it difficult to tell whether it is
 149                  * real FID or ost_id. */
 150                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 151                 oi->oi_fid.f_oid = id ?: 1;
 152                 oi->oi_fid.f_ver = 0;
 153         } else {
 154                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 155
 156                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 157                         GOTO(out, rc = -EPROTO);
 158
 159                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 160                  * all of the requests are required to setup FLGROUP */
 161                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 162                         ostid_set_seq_mdt0(oi);
 163                         oa->o_valid |= OBD_MD_FLGROUP;
 164                         seq = ostid_seq(oi);
 165                 }
 166
 167                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 168                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 169                         GOTO(out, rc = -EPROTO);
 170
 171                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 172                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 173                 if (unlikely(rc != 0))
 174                         GOTO(out, rc);
 175
 176                 oi->oi_fid = tti->tti_fid1;
 177         }
 178
 179         RETURN(0);
 180
 181 out:
 182         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 183                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 184                seq, id, rc);
 185         return rc;
 186 }
 187 EXPORT_SYMBOL(tgt_validate_obdo);
 188
 189 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 190 {
 191         unsigned                 max_brw;
 192         struct niobuf_remote    *rnb;
 193         struct obd_ioobj        *ioo;
 194         int                      obj_count;
 195
 196         ENTRY;
 197
 198         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 199         if (ioo == NULL)
 200                 RETURN(-EPROTO);
 201
 202         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 203         if (rnb == NULL)
 204                 RETURN(-EPROTO);
 205
 206         max_brw = ioobj_max_brw_get(ioo);
 207         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 208                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 209                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 210                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 211                        POSTID(oi), -EPROTO);
 212                 RETURN(-EPROTO);
 213         }
 214         ioo->ioo_oid = *oi;
 215
 216         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 217                                         RCL_CLIENT) / sizeof(*ioo);
 218         if (obj_count == 0) {
 219                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 220                 RETURN(-EPROTO);
 221         } else if (obj_count > 1) {
 222                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 223                        obj_count);
 224                 RETURN(-EPROTO);
 225         }
 226
 227         if (ioo->ioo_bufcnt == 0) {
 228                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 229                 RETURN(-EPROTO);
 230         }
 231
 232         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 233                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 234                           "bulk has too many pages (%d)",
 235                           ioo->ioo_bufcnt);
 236                 RETURN(-EPROTO);
 237         }
 238
 239         RETURN(0);
 240 }
 241
 242 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 243 {
 244         struct ost_body         *body;
 245         struct req_capsule      *pill = tsi->tsi_pill;
 246         struct lu_nodemap       *nodemap;
 247         int                      rc;
 248
 249         ENTRY;
 250
 251         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 252         if (body == NULL)
 253                 RETURN(-EFAULT);
 254
 255         rc = tgt_validate_obdo(tsi, &body->oa);
 256         if (rc)
 257                 RETURN(rc);
 258
 259         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 260         if (IS_ERR(nodemap))
 261                 RETURN(PTR_ERR(nodemap));
 262
 263         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 264                                         NODEMAP_CLIENT_TO_FS,
 265                                         body->oa.o_uid);
 266         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 267                                         NODEMAP_CLIENT_TO_FS,
 268                                         body->oa.o_gid);
 269         nodemap_putref(nodemap);
 270
 271         tsi->tsi_ost_body = body;
 272         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 273
 274         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 275                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 276                 if (rc < 0)
 277                         RETURN(rc);
 278         }
 279
 280         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 281                 if (flags & HAS_BODY) {
 282                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 283                                tgt_name(tsi->tsi_tgt));
 284                         RETURN(-EPROTO);
 285                 } else {
 286                         RETURN(0);
 287                 }
 288         }
 289
 290         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 291
 292         /*
 293          * OST doesn't get object in advance for further use to prevent
 294          * situations with nested object_find which is potential deadlock.
 295          */
 296         tsi->tsi_corpus = NULL;
 297         RETURN(rc);
 298 }
 299
 300 /*
 301  * Do necessary preprocessing according to handler ->th_flags.
 302  */
 303 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 304                                   struct tgt_handler *h,
 305                                   struct ptlrpc_request *req)
 306 {
 307         struct req_capsule      *pill = tsi->tsi_pill;
 308         __u32                    flags = h->th_flags;
 309         int                      rc = 0;
 310
 311         ENTRY;
 312
 313         if (tsi->tsi_preprocessed)
 314                 RETURN(0);
 315
 316         LASSERT(h->th_act != NULL);
 317         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 318         LASSERT(current->journal_info == NULL);
 319
 320         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 321                      h->th_fmt != NULL));
 322         if (h->th_fmt != NULL) {
 323                 req_capsule_set(pill, h->th_fmt);
 324                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
 325                         rc = tgt_mdt_body_unpack(tsi, flags);
 326                         if (rc < 0)
 327                                 RETURN(rc);
 328                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 329                                                  RCL_CLIENT)) {
 330                         rc = tgt_ost_body_unpack(tsi, flags);
 331                         if (rc < 0)
 332                                 RETURN(rc);
 333                 }
 334         }
 335
 336         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 337                 RETURN(-EROFS);
 338
 339         if (flags & HAS_KEY) {
 340                 struct ldlm_request *dlm_req;
 341
 342                 LASSERT(h->th_fmt != NULL);
 343
 344                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 345                 if (dlm_req != NULL) {
 346                         union ldlm_wire_policy_data *policy =
 347                                         &dlm_req->lock_desc.l_policy_data;
 348
 349                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 350                                      LDLM_IBITS &&
 351                                      (policy->l_inodebits.bits |
 352                                       policy->l_inodebits.try_bits) == 0)) {
 353                                 /*
 354                                  * Lock without inodebits makes no sense and
 355                                  * will oops later in ldlm. If client miss to
 356                                  * set such bits, do not trigger ASSERTION.
 357                                  *
 358                                  * For liblustre flock case, it maybe zero.
 359                                  */
 360                                 rc = -EPROTO;
 361                         } else {
 362                                 tsi->tsi_dlm_req = dlm_req;
 363                         }
 364                 } else {
 365                         rc = -EFAULT;
 366                 }
 367         }
 368         tsi->tsi_preprocessed = 1;
 369         RETURN(rc);
 370 }
 371
 372 /*
 373  * Invoke handler for this request opc. Also do necessary preprocessing
 374  * (according to handler ->th_flags), and post-processing (setting of
 375  * ->last_{xid,committed}).
 376  */
 377 static int tgt_handle_request0(struct tgt_session_info *tsi,
 378                                struct tgt_handler *h,
 379                                struct ptlrpc_request *req)
 380 {
 381         int      serious = 0;
 382         int      rc;
 383         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 384
 385         ENTRY;
 386
 387
 388         /* When dealing with sec context requests, no export is associated yet,
 389          * because these requests are sent before *_CONNECT requests.
 390          * A NULL req->rq_export means the normal *_common_slice handlers will
 391          * not be called, because there is no reference to the target.
 392          * So deal with them by hand and jump directly to target_send_reply().
 393          */
 394         switch (opc) {
 395         case SEC_CTX_INIT:
 396         case SEC_CTX_INIT_CONT:
 397         case SEC_CTX_FINI:
 398                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 399                 GOTO(out, rc = 0);
 400         }
 401
 402         /*
 403          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 404          * to put same checks into handlers like mdt_close(), mdt_reint(),
 405          * etc., without talking to mdt authors first. Checking same thing
 406          * there again is useless and returning 0 error without packing reply
 407          * is buggy! Handlers either pack reply or return error.
 408          *
 409          * We return 0 here and do not send any reply in order to emulate
 410          * network failure. Do not send any reply in case any of NET related
 411          * fail_id has occured.
 412          */
 413         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 414                 RETURN(0);
 415         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 416                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 417                 RETURN(0);
 418
 419         rc = tgt_request_preprocess(tsi, h, req);
 420         /* pack reply if reply format is fixed */
 421         if (rc == 0 && h->th_flags & HAS_REPLY) {
 422                 /* Pack reply */
 423                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 424                                           RCL_SERVER))
 425                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 426                                              RCL_SERVER,
 427                                              tsi->tsi_mdt_body->mbo_eadatasize);
 428                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 429                                           RCL_SERVER))
 430                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 431                                              RCL_SERVER, 0);
 432                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 433                         req_capsule_set_size(tsi->tsi_pill,
 434                                              &RMF_ACL, RCL_SERVER,
 435                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 436
 437                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 438                                           RCL_SERVER)) {
 439                         struct niobuf_remote *remote_nb =
 440                                 req_capsule_client_get(tsi->tsi_pill,
 441                                                        &RMF_NIOBUF_REMOTE);
 442                         struct ost_body *body = tsi->tsi_ost_body;
 443
 444                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 445                                          RCL_SERVER,
 446                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 447                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 448                                          remote_nb[0].rnb_len : 0);
 449                 }
 450
 451                 rc = req_capsule_server_pack(tsi->tsi_pill);
 452         }
 453
 454         if (likely(rc == 0)) {
 455                 /*
 456                  * Process request, there can be two types of rc:
 457                  * 1) errors with msg unpack/pack, other failures outside the
 458                  * operation itself. This is counted as serious errors;
 459                  * 2) errors during fs operation, should be placed in rq_status
 460                  * only
 461                  */
 462                 rc = h->th_act(tsi);
 463                 if (!is_serious(rc) &&
 464                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 465                         DEBUG_REQ(D_ERROR, req,
 466                                   "%s: %s handler did not pack reply but returned no error",
 467                                   tgt_name(tsi->tsi_tgt), h->th_name);
 468                         LBUG();
 469                 }
 470                 serious = is_serious(rc);
 471                 rc = clear_serious(rc);
 472         } else {
 473                 serious = 1;
 474         }
 475
 476         req->rq_status = rc;
 477
 478         /*
 479          * ELDLM_* codes which > 0 should be in rq_status only as well as
 480          * all non-serious errors.
 481          */
 482         if (rc > 0 || !serious)
 483                 rc = 0;
 484
 485         LASSERT(current->journal_info == NULL);
 486
 487         if (likely(rc == 0 && req->rq_export))
 488                 target_committed_to_req(req);
 489
 490 out:
 491         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 492         RETURN(0);
 493 }
 494
 495 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 496                                        struct obd_device *obd, int *process)
 497 {
 498         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 499         case MDS_DISCONNECT:
 500         case OST_DISCONNECT:
 501         case OBD_IDX_READ:
 502                 *process = 1;
 503                 RETURN(0);
 504         case MDS_CLOSE:
 505         case MDS_SYNC: /* used in unmounting */
 506         case OBD_PING:
 507         case MDS_REINT:
 508         case OUT_UPDATE:
 509         case SEQ_QUERY:
 510         case FLD_QUERY:
 511         case FLD_READ:
 512         case LDLM_ENQUEUE:
 513         case OST_CREATE:
 514         case OST_DESTROY:
 515         case OST_PUNCH:
 516         case OST_SETATTR:
 517         case OST_SYNC:
 518         case OST_WRITE:
 519         case MDS_HSM_PROGRESS:
 520         case MDS_HSM_STATE_SET:
 521         case MDS_HSM_REQUEST:
 522                 *process = target_queue_recovery_request(req, obd);
 523                 RETURN(0);
 524
 525         default:
 526                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 527                 *process = -EAGAIN;
 528                 RETURN(0);
 529         }
 530 }
 531
 532 /*
 533  * Handle recovery. Return:
 534  *        +1: continue request processing;
 535  *       -ve: abort immediately with the given error code;
 536  *         0: send reply with error code in req->rq_status;
 537  */
 538 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 539 {
 540         ENTRY;
 541
 542         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 543         case MDS_CONNECT:
 544         case OST_CONNECT:
 545         case MGS_CONNECT:
 546         case SEC_CTX_INIT:
 547         case SEC_CTX_INIT_CONT:
 548         case SEC_CTX_FINI:
 549                 RETURN(+1);
 550         }
 551
 552         if (!req->rq_export->exp_obd->obd_replayable)
 553                 RETURN(+1);
 554
 555         /* sanity check: if the xid matches, the request must be marked as a
 556          * resent or replayed */
 557         if (req_can_reconstruct(req, NULL) == 1) {
 558                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 559                       (MSG_RESENT | MSG_REPLAY))) {
 560                         DEBUG_REQ(D_WARNING, req,
 561                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 562                                   req->rq_xid,
 563                                   lustre_msg_get_flags(req->rq_reqmsg));
 564                         req->rq_status = -ENOTCONN;
 565                         RETURN(-ENOTCONN);
 566                 }
 567         }
 568         /* else: note the opposite is not always true; a RESENT req after a
 569          * failover will usually not match the last_xid, since it was likely
 570          * never committed. A REPLAYed request will almost never match the
 571          * last xid, however it could for a committed, but still retained,
 572          * open. */
 573
 574         /* Check for aborted recovery... */
 575         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 576                 int rc;
 577                 int should_process;
 578
 579                 DEBUG_REQ(D_INFO, req, "Got new replay");
 580                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 581                                                  &should_process);
 582                 if (rc != 0 || !should_process)
 583                         RETURN(rc);
 584                 else if (should_process < 0) {
 585                         req->rq_status = should_process;
 586                         rc = ptlrpc_error(req);
 587                         RETURN(rc);
 588                 }
 589         }
 590         RETURN(+1);
 591 }
 592
 593 /* Initial check for request, it is validation mostly */
 594 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 595 {
 596         struct tgt_handler      *h;
 597         struct tgt_opc_slice    *s;
 598         struct lu_target        *tgt;
 599         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 600
 601         ENTRY;
 602
 603         tgt = class_exp2tgt(req->rq_export);
 604         if (unlikely(tgt == NULL)) {
 605                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 606                           class_exp2obd(req->rq_export)->obd_name);
 607                 RETURN(ERR_PTR(-EINVAL));
 608         }
 609
 610         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 611                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 612                         break;
 613
 614         /* opcode was not found in slice */
 615         if (unlikely(s->tos_hs == NULL)) {
 616                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 617                        opc);
 618                 RETURN(ERR_PTR(-ENOTSUPP));
 619         }
 620
 621         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 622         h = s->tos_hs + (opc - s->tos_opc_start);
 623         if (unlikely(h->th_opc == 0)) {
 624                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 625                 RETURN(ERR_PTR(-ENOTSUPP));
 626         }
 627
 628         RETURN(h);
 629 }
 630
 631 static int process_req_last_xid(struct ptlrpc_request *req)
 632 {
 633         __u64   last_xid;
 634         int rc = 0;
 635         struct obd_export *exp = req->rq_export;
 636         struct tg_export_data *ted = &exp->exp_target_data;
 637         bool need_lock = tgt_is_multimodrpcs_client(exp);
 638         ENTRY;
 639
 640         if (need_lock)
 641                 mutex_lock(&ted->ted_lcd_lock);
 642         /* check request's xid is consistent with export's last_xid */
 643         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 644         if (last_xid > exp->exp_last_xid)
 645                 exp->exp_last_xid = last_xid;
 646
 647         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 648                 /* Some request is allowed to be sent during replay,
 649                  * such as OUT update requests, FLD requests, so it
 650                  * is possible that replay requests has smaller XID
 651                  * than the exp_last_xid.
 652                  *
 653                  * Some non-replay requests may have smaller XID as
 654                  * well:
 655                  *
 656                  * - Client send a no_resend RPC, like statfs;
 657                  * - The RPC timedout (or some other error) on client,
 658                  *   then it's removed from the unreplied list;
 659                  * - Client send some other request to bump the
 660                  *   exp_last_xid on server;
 661                  * - The former RPC got chance to be processed;
 662                  */
 663                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 664                         rc = -EPROTO;
 665
 666                 DEBUG_REQ(D_WARNING, req,
 667                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 668                           req->rq_xid, exp->exp_last_xid, rc);
 669                 if (rc)
 670                         GOTO(out, rc);
 671         }
 672
 673         /* The "last_xid" is the minimum xid among unreplied requests,
 674          * if the request is from the previous connection, its xid can
 675          * still be larger than "exp_last_xid", then the above check of
 676          * xid is not enough to determine whether the request is delayed.
 677          *
 678          * For example, if some replay request was delayed and caused
 679          * timeout at client and the replay is restarted, the delayed
 680          * replay request will have the larger xid than "exp_last_xid"
 681          */
 682         if (req->rq_export->exp_conn_cnt >
 683             lustre_msg_get_conn_cnt(req->rq_reqmsg))
 684                 GOTO(out, rc = -ESTALE);
 685
 686         /* try to release in-memory reply data */
 687         if (tgt_is_multimodrpcs_client(exp)) {
 688                 tgt_handle_received_xid(exp, last_xid);
 689                 rc = tgt_handle_tag(req);
 690         }
 691
 692 out:
 693         if (need_lock)
 694                 mutex_unlock(&ted->ted_lcd_lock);
 695
 696         RETURN(rc);
 697 }
 698
 699 int tgt_request_handle(struct ptlrpc_request *req)
 700 {
 701         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 702
 703         struct lustre_msg       *msg = req->rq_reqmsg;
 704         struct tgt_handler      *h;
 705         struct lu_target        *tgt;
 706         int                      request_fail_id = 0;
 707         __u32                    opc = lustre_msg_get_opc(msg);
 708         struct obd_device       *obd;
 709         int                      rc;
 710         bool                     is_connect = false;
 711         ENTRY;
 712
 713         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 714                 if (cfs_fail_val == 0 &&
 715                     lustre_msg_get_opc(msg) != OBD_PING &&
 716                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 717                         cfs_fail_val = 1;
 718                         cfs_race_state = 0;
 719                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 720                 }
 721         }
 722
 723         /* Refill the context, to make sure all thread keys are allocated */
 724         lu_env_refill(req->rq_svc_thread->t_env);
 725
 726         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 727         tsi->tsi_pill = &req->rq_pill;
 728         tsi->tsi_env = req->rq_svc_thread->t_env;
 729
 730         /* if request has export then get handlers slice from corresponding
 731          * target, otherwise that should be connect operation */
 732         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 733             opc == MGS_CONNECT) {
 734                 is_connect = true;
 735                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 736                 rc = target_handle_connect(req);
 737                 if (rc != 0) {
 738                         rc = ptlrpc_error(req);
 739                         GOTO(out, rc);
 740                 }
 741                 /* recovery-small test 18c asks to drop connect reply */
 742                 if (unlikely(opc == OST_CONNECT &&
 743                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 744                         GOTO(out, rc = 0);
 745         }
 746
 747         if (unlikely(!class_connected_export(req->rq_export))) {
 748                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 749                     opc == SEC_CTX_FINI) {
 750                         /* sec context initialization has to be handled
 751                          * by hand in tgt_handle_request0() */
 752                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 753                         h = NULL;
 754                         GOTO(handle_recov, rc = 0);
 755                 }
 756                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 757                        opc, libcfs_id2str(req->rq_peer));
 758                 req->rq_status = -ENOTCONN;
 759                 rc = ptlrpc_error(req);
 760                 GOTO(out, rc);
 761         }
 762
 763         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 764         tsi->tsi_exp = req->rq_export;
 765         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 766                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 767         else
 768                 tsi->tsi_jobid = NULL;
 769
 770         if (tgt == NULL) {
 771                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 772                           class_exp2obd(req->rq_export)->obd_name);
 773                 req->rq_status = -EINVAL;
 774                 rc = ptlrpc_error(req);
 775                 GOTO(out, rc);
 776         }
 777
 778         /* Skip last_xid processing for the recovery thread, otherwise, the
 779          * last_xid on same request could be processed twice: first time when
 780          * processing the incoming request, second time when the request is
 781          * being processed by recovery thread. */
 782         obd = class_exp2obd(req->rq_export);
 783         if (is_connect) {
 784                 /* reset the exp_last_xid on each connection. */
 785                 req->rq_export->exp_last_xid = 0;
 786         } else if (obd->obd_recovery_data.trd_processing_task !=
 787                    current_pid()) {
 788                 rc = process_req_last_xid(req);
 789                 if (rc) {
 790                         req->rq_status = rc;
 791                         rc = ptlrpc_error(req);
 792                         GOTO(out, rc);
 793                 }
 794         }
 795
 796         request_fail_id = tgt->lut_request_fail_id;
 797         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 798
 799         h = tgt_handler_find_check(req);
 800         if (IS_ERR(h)) {
 801                 req->rq_status = PTR_ERR(h);
 802                 rc = ptlrpc_error(req);
 803                 GOTO(out, rc);
 804         }
 805
 806         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 807                  h->th_opc, opc);
 808
 809         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 810              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 811                 GOTO(out, rc = 0);
 812
 813         rc = lustre_msg_check_version(msg, h->th_version);
 814         if (unlikely(rc)) {
 815                 DEBUG_REQ(D_ERROR, req,
 816                           "%s: drop malformed request version=%08x expect=%08x",
 817                           tgt_name(tgt), lustre_msg_get_version(msg),
 818                           h->th_version);
 819                 req->rq_status = -EINVAL;
 820                 rc = ptlrpc_error(req);
 821                 GOTO(out, rc);
 822         }
 823
 824 handle_recov:
 825         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 826         if (likely(rc == 1)) {
 827                 rc = tgt_handle_request0(tsi, h, req);
 828                 if (rc)
 829                         GOTO(out, rc);
 830         }
 831         EXIT;
 832 out:
 833         req_capsule_fini(tsi->tsi_pill);
 834         if (tsi->tsi_corpus != NULL) {
 835                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 836                 tsi->tsi_corpus = NULL;
 837         }
 838         return rc;
 839 }
 840 EXPORT_SYMBOL(tgt_request_handle);
 841
 842 /** Assign high priority operations to the request if needed. */
 843 int tgt_hpreq_handler(struct ptlrpc_request *req)
 844 {
 845         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 846         struct tgt_handler      *h;
 847         int                      rc;
 848
 849         ENTRY;
 850
 851         if (req->rq_export == NULL)
 852                 RETURN(0);
 853
 854         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 855         tsi->tsi_pill = &req->rq_pill;
 856         tsi->tsi_env = req->rq_svc_thread->t_env;
 857         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 858         tsi->tsi_exp = req->rq_export;
 859
 860         h = tgt_handler_find_check(req);
 861         if (IS_ERR(h)) {
 862                 rc = PTR_ERR(h);
 863                 RETURN(rc);
 864         }
 865
 866         rc = tgt_request_preprocess(tsi, h, req);
 867         if (unlikely(rc != 0))
 868                 RETURN(rc);
 869
 870         if (h->th_hp != NULL)
 871                 h->th_hp(tsi);
 872         RETURN(0);
 873 }
 874 EXPORT_SYMBOL(tgt_hpreq_handler);
 875
 876 void tgt_counter_incr(struct obd_export *exp, int opcode)
 877 {
 878         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 879         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 880                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 881 }
 882 EXPORT_SYMBOL(tgt_counter_incr);
 883
 884 /*
 885  * Unified target generic handlers.
 886  */
 887
 888 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 889 {
 890         struct lu_target        *tgt = class_exp2tgt(exp);
 891         struct sptlrpc_flavor    flvr;
 892         int                      rc = 0;
 893
 894         LASSERT(tgt);
 895         LASSERT(tgt->lut_obd);
 896         LASSERT(tgt->lut_slice);
 897
 898         /* always allow ECHO client */
 899         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 900                             LUSTRE_ECHO_NAME) == 0)) {
 901                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 902                 return 0;
 903         }
 904
 905         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 906                 read_lock(&tgt->lut_sptlrpc_lock);
 907                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 908                                              req->rq_sp_from,
 909                                              req->rq_peer.nid,
 910                                              &flvr);
 911                 read_unlock(&tgt->lut_sptlrpc_lock);
 912
 913                 spin_lock(&exp->exp_lock);
 914                 exp->exp_sp_peer = req->rq_sp_from;
 915                 exp->exp_flvr = flvr;
 916
 917                 /* when on mgs, if no restriction is set, or if client
 918                  * is loopback, allow any flavor */
 919                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 920                            LUSTRE_MGS_NAME) == 0) &&
 921                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 922                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 923                       == LOLND))
 924                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 925
 926                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 927                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 928                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 929                                "expect %x\n", tgt_name(tgt),
 930                                req->rq_flvr.sf_rpc,
 931                                libcfs_nid2str(req->rq_peer.nid),
 932                                exp->exp_flvr.sf_rpc);
 933                         rc = -EACCES;
 934                 }
 935                 spin_unlock(&exp->exp_lock);
 936         } else {
 937                 if (exp->exp_sp_peer != req->rq_sp_from) {
 938                         CERROR("%s: RPC source %s doesn't match %s\n",
 939                                tgt_name(tgt),
 940                                sptlrpc_part2name(req->rq_sp_from),
 941                                sptlrpc_part2name(exp->exp_sp_peer));
 942                         rc = -EACCES;
 943                 } else {
 944                         rc = sptlrpc_target_export_check(exp, req);
 945                 }
 946         }
 947
 948         return rc;
 949 }
 950
 951 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 952 {
 953         struct sptlrpc_rule_set  tmp_rset;
 954         int                      rc;
 955
 956         if (unlikely(tgt == NULL)) {
 957                 CERROR("No target passed");
 958                 return -EINVAL;
 959         }
 960
 961         sptlrpc_rule_set_init(&tmp_rset);
 962         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 963         if (rc) {
 964                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 965                        tgt_name(tgt), rc);
 966                 return rc;
 967         }
 968
 969         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 970
 971         write_lock(&tgt->lut_sptlrpc_lock);
 972         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 973         tgt->lut_sptlrpc_rset = tmp_rset;
 974         write_unlock(&tgt->lut_sptlrpc_lock);
 975
 976         return 0;
 977 }
 978 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 979
 980 int tgt_connect(struct tgt_session_info *tsi)
 981 {
 982         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 983         struct obd_connect_data *reply;
 984         int                      rc;
 985
 986         ENTRY;
 987
 988         /* XXX: better to call this check right after getting new export but
 989          * before last_rcvd slot allocation to avoid server load upon insecure
 990          * connects. This is to be fixed after unifiyng all targets.
 991          */
 992         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 993         if (rc)
 994                 GOTO(out, rc);
 995
 996         /* To avoid exposing partially initialized connection flags, changes up
 997          * to this point have been staged in reply->ocd_connect_flags. Now that
 998          * connection handling has completed successfully, atomically update
 999          * the connect flags in the shared export data structure. LU-1623 */
1000         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1001         spin_lock(&tsi->tsi_exp->exp_lock);
1002         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1003         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1004                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1005                         reply->ocd_connect_flags2;
1006         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1007         spin_unlock(&tsi->tsi_exp->exp_lock);
1008
1009         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1010                    LUSTRE_MDT_NAME) == 0) {
1011                 rc = req_check_sepol(tsi->tsi_pill);
1012                 if (rc)
1013                         GOTO(out, rc);
1014         }
1015
1016         RETURN(0);
1017 out:
1018         obd_disconnect(class_export_get(tsi->tsi_exp));
1019         return rc;
1020 }
1021 EXPORT_SYMBOL(tgt_connect);
1022
1023 int tgt_disconnect(struct tgt_session_info *tsi)
1024 {
1025         int rc;
1026
1027         ENTRY;
1028
1029         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1030
1031         rc = target_handle_disconnect(tgt_ses_req(tsi));
1032         if (rc)
1033                 RETURN(err_serious(rc));
1034
1035         RETURN(rc);
1036 }
1037 EXPORT_SYMBOL(tgt_disconnect);
1038
1039 /*
1040  * Unified target OBD handlers
1041  */
1042 int tgt_obd_ping(struct tgt_session_info *tsi)
1043 {
1044         int rc;
1045
1046         ENTRY;
1047
1048         /* The target-specific part of OBD_PING request handling.
1049          * It controls Filter Modification Data (FMD) expiration each time
1050          * PING is received.
1051          *
1052          * Valid only for replayable targets, e.g. MDT and OFD
1053          */
1054         if (tsi->tsi_exp->exp_obd->obd_replayable)
1055                 tgt_fmd_expire(tsi->tsi_exp);
1056
1057         rc = req_capsule_server_pack(tsi->tsi_pill);
1058         if (rc)
1059                 RETURN(err_serious(rc));
1060
1061         RETURN(rc);
1062 }
1063 EXPORT_SYMBOL(tgt_obd_ping);
1064
1065 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1066 {
1067         return err_serious(-EOPNOTSUPP);
1068 }
1069
1070 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1071 {
1072         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1073         struct obd_export       *exp = req->rq_export;
1074         struct ptlrpc_bulk_desc *desc;
1075         int                      i;
1076         int                      rc;
1077
1078         ENTRY;
1079
1080         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1081                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1082                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1083         if (desc == NULL)
1084                 RETURN(-ENOMEM);
1085
1086         for (i = 0; i < rdbuf->rb_nbufs; i++)
1087                 desc->bd_frag_ops->add_iov_frag(desc,
1088                                         rdbuf->rb_bufs[i].lb_buf,
1089                                         rdbuf->rb_bufs[i].lb_len);
1090
1091         rc = target_bulk_io(exp, desc);
1092         ptlrpc_free_bulk(desc);
1093         RETURN(rc);
1094 }
1095 EXPORT_SYMBOL(tgt_send_buffer);
1096
1097 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1098 {
1099         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1100         struct obd_export       *exp = req->rq_export;
1101         struct ptlrpc_bulk_desc *desc;
1102         int                      tmpcount;
1103         int                      tmpsize;
1104         int                      i;
1105         int                      rc;
1106
1107         ENTRY;
1108
1109         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1110                                     PTLRPC_BULK_PUT_SOURCE |
1111                                         PTLRPC_BULK_BUF_KIOV,
1112                                     MDS_BULK_PORTAL,
1113                                     &ptlrpc_bulk_kiov_pin_ops);
1114         if (desc == NULL)
1115                 RETURN(-ENOMEM);
1116
1117         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1118                 /* old client requires reply size in it's PAGE_SIZE,
1119                  * which is rdpg->rp_count */
1120                 nob = rdpg->rp_count;
1121
1122         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1123              i++, tmpcount -= tmpsize) {
1124                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1125                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1126                                                  tmpsize);
1127         }
1128
1129         LASSERT(desc->bd_nob == nob);
1130         rc = target_bulk_io(exp, desc);
1131         ptlrpc_free_bulk(desc);
1132         RETURN(rc);
1133 }
1134 EXPORT_SYMBOL(tgt_sendpage);
1135
1136 /*
1137  * OBD_IDX_READ handler
1138  */
1139 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1140 {
1141         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1142         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1143         struct idx_info         *req_ii, *rep_ii;
1144         int                      rc, i;
1145
1146         ENTRY;
1147
1148         memset(rdpg, 0, sizeof(*rdpg));
1149         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1150
1151         /* extract idx_info buffer from request & reply */
1152         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1153         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1154                 RETURN(err_serious(-EPROTO));
1155
1156         rc = req_capsule_server_pack(tsi->tsi_pill);
1157         if (rc)
1158                 RETURN(err_serious(rc));
1159
1160         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1161         if (rep_ii == NULL)
1162                 RETURN(err_serious(-EFAULT));
1163         rep_ii->ii_magic = IDX_INFO_MAGIC;
1164
1165         /* extract hash to start with */
1166         rdpg->rp_hash = req_ii->ii_hash_start;
1167
1168         /* extract requested attributes */
1169         rdpg->rp_attrs = req_ii->ii_attrs;
1170
1171         /* check that fid packed in request is valid and supported */
1172         if (!fid_is_sane(&req_ii->ii_fid))
1173                 RETURN(-EINVAL);
1174         rep_ii->ii_fid = req_ii->ii_fid;
1175
1176         /* copy flags */
1177         rep_ii->ii_flags = req_ii->ii_flags;
1178
1179         /* compute number of pages to allocate, ii_count is the number of 4KB
1180          * containers */
1181         if (req_ii->ii_count <= 0)
1182                 GOTO(out, rc = -EFAULT);
1183         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1184                                exp_max_brw_size(tsi->tsi_exp));
1185         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1186
1187         /* allocate pages to store the containers */
1188         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1189         if (rdpg->rp_pages == NULL)
1190                 GOTO(out, rc = -ENOMEM);
1191         for (i = 0; i < rdpg->rp_npages; i++) {
1192                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1193                 if (rdpg->rp_pages[i] == NULL)
1194                         GOTO(out, rc = -ENOMEM);
1195         }
1196
1197         /* populate pages with key/record pairs */
1198         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1199         if (rc < 0)
1200                 GOTO(out, rc);
1201
1202         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1203                  "asked %d > %d\n", rc, rdpg->rp_count);
1204
1205         /* send pages to client */
1206         rc = tgt_sendpage(tsi, rdpg, rc);
1207         if (rc)
1208                 GOTO(out, rc);
1209         EXIT;
1210 out:
1211         if (rdpg->rp_pages) {
1212                 for (i = 0; i < rdpg->rp_npages; i++)
1213                         if (rdpg->rp_pages[i])
1214                                 __free_page(rdpg->rp_pages[i]);
1215                 OBD_FREE(rdpg->rp_pages,
1216                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1217         }
1218         return rc;
1219 }
1220
1221 struct tgt_handler tgt_obd_handlers[] = {
1222 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1223 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1224 };
1225 EXPORT_SYMBOL(tgt_obd_handlers);
1226
1227 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1228              struct dt_object *obj, __u64 start, __u64 end)
1229 {
1230         int rc = 0;
1231
1232         ENTRY;
1233
1234         /* if no objid is specified, it means "sync whole filesystem" */
1235         if (obj == NULL) {
1236                 rc = dt_sync(env, tgt->lut_bottom);
1237         } else if (dt_version_get(env, obj) >
1238                    tgt->lut_obd->obd_last_committed) {
1239                 rc = dt_object_sync(env, obj, start, end);
1240         }
1241         atomic_inc(&tgt->lut_sync_count);
1242
1243         RETURN(rc);
1244 }
1245 EXPORT_SYMBOL(tgt_sync);
1246 /*
1247  * Unified target DLM handlers.
1248  */
1249
1250 /**
1251  * Unified target BAST
1252  *
1253  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1254  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1255  * but if it's cross-MDT lock, because remote object version is not set, a
1256  * filesystem sync is needed.
1257  *
1258  * \param lock server side lock
1259  * \param desc lock desc
1260  * \param data ldlm_cb_set_arg
1261  * \param flag  indicates whether this cancelling or blocking callback
1262  * \retval      0 on success
1263  * \retval      negative number on error
1264  */
1265 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1266                             void *data, int flag)
1267 {
1268         struct lu_env            env;
1269         struct lu_target        *tgt;
1270         struct dt_object        *obj = NULL;
1271         struct lu_fid            fid;
1272         int                      rc = 0;
1273
1274         ENTRY;
1275
1276         tgt = class_exp2tgt(lock->l_export);
1277
1278         if (unlikely(tgt == NULL)) {
1279                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1280                        class_exp2obd(lock->l_export)->obd_name);
1281                 RETURN(-EINVAL);
1282         }
1283
1284         if (flag == LDLM_CB_CANCELING &&
1285             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1286             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1287              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1288               ldlm_is_cbpending(lock))) &&
1289             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1290              lock->l_resource->lr_type == LDLM_EXTENT)) {
1291                 __u64 start = 0;
1292                 __u64 end = OBD_OBJECT_EOF;
1293
1294                 rc = lu_env_init(&env, LCT_DT_THREAD);
1295                 if (unlikely(rc != 0))
1296                         GOTO(err, rc);
1297
1298                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1299                                    tgt->lut_lsd.lsd_osd_index);
1300
1301                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1302                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1303                         if (IS_ERR(obj))
1304                                 GOTO(err_env, rc = PTR_ERR(obj));
1305
1306                         if (!dt_object_exists(obj))
1307                                 GOTO(err_put, rc = -ENOENT);
1308
1309                         start = lock->l_policy_data.l_extent.start;
1310                         end = lock->l_policy_data.l_extent.end;
1311                 }
1312
1313                 rc = tgt_sync(&env, tgt, obj, start, end);
1314                 if (rc < 0) {
1315                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1316                                "cancel: rc = %d\n",
1317                                tgt_name(tgt), PFID(&fid),
1318                                lock->l_policy_data.l_extent.start,
1319                                lock->l_policy_data.l_extent.end, rc);
1320                 }
1321 err_put:
1322                 if (obj != NULL)
1323                         dt_object_put(&env, obj);
1324 err_env:
1325                 lu_env_fini(&env);
1326         }
1327 err:
1328         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1329         RETURN(rc);
1330 }
1331
1332 static struct ldlm_callback_suite tgt_dlm_cbs = {
1333         .lcs_completion = ldlm_server_completion_ast,
1334         .lcs_blocking   = tgt_blocking_ast,
1335         .lcs_glimpse    = ldlm_server_glimpse_ast
1336 };
1337
1338 int tgt_enqueue(struct tgt_session_info *tsi)
1339 {
1340         struct ptlrpc_request *req = tgt_ses_req(tsi);
1341         int rc;
1342
1343         ENTRY;
1344         /*
1345          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1346          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1347          */
1348         LASSERT(tsi->tsi_dlm_req != NULL);
1349         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1350                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1351         if (rc)
1352                 RETURN(err_serious(rc));
1353
1354         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1355         case LUT_FAIL_MDT:
1356                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1357                 break;
1358         case LUT_FAIL_OST:
1359                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1360                 break;
1361         case LUT_FAIL_MGT:
1362                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1363                 break;
1364         default:
1365                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1366                 break;
1367         }
1368         RETURN(req->rq_status);
1369 }
1370 EXPORT_SYMBOL(tgt_enqueue);
1371
1372 int tgt_convert(struct tgt_session_info *tsi)
1373 {
1374         struct ptlrpc_request *req = tgt_ses_req(tsi);
1375         int rc;
1376
1377         ENTRY;
1378         LASSERT(tsi->tsi_dlm_req);
1379         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1380         if (rc)
1381                 RETURN(err_serious(rc));
1382
1383         RETURN(req->rq_status);
1384 }
1385
1386 int tgt_bl_callback(struct tgt_session_info *tsi)
1387 {
1388         return err_serious(-EOPNOTSUPP);
1389 }
1390
1391 int tgt_cp_callback(struct tgt_session_info *tsi)
1392 {
1393         return err_serious(-EOPNOTSUPP);
1394 }
1395
1396 /* generic LDLM target handler */
1397 struct tgt_handler tgt_dlm_handlers[] = {
1398 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1399 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1400 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1401 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1402 };
1403 EXPORT_SYMBOL(tgt_dlm_handlers);
1404
1405 /*
1406  * Unified target LLOG handlers.
1407  */
1408 int tgt_llog_open(struct tgt_session_info *tsi)
1409 {
1410         int rc;
1411
1412         ENTRY;
1413
1414         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1415
1416         RETURN(rc);
1417 }
1418 EXPORT_SYMBOL(tgt_llog_open);
1419
1420 int tgt_llog_read_header(struct tgt_session_info *tsi)
1421 {
1422         int rc;
1423
1424         ENTRY;
1425
1426         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1427
1428         RETURN(rc);
1429 }
1430 EXPORT_SYMBOL(tgt_llog_read_header);
1431
1432 int tgt_llog_next_block(struct tgt_session_info *tsi)
1433 {
1434         int rc;
1435
1436         ENTRY;
1437
1438         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1439
1440         RETURN(rc);
1441 }
1442 EXPORT_SYMBOL(tgt_llog_next_block);
1443
1444 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1445 {
1446         int rc;
1447
1448         ENTRY;
1449
1450         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1451
1452         RETURN(rc);
1453 }
1454 EXPORT_SYMBOL(tgt_llog_prev_block);
1455
1456 /* generic llog target handler */
1457 struct tgt_handler tgt_llog_handlers[] = {
1458 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1459 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1460 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1461 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1462 };
1463 EXPORT_SYMBOL(tgt_llog_handlers);
1464
1465 /*
1466  * sec context handlers
1467  */
1468 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1469 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1470 {
1471         return 0;
1472 }
1473
1474 struct tgt_handler tgt_sec_ctx_handlers[] = {
1475 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1476 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1477 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1478 };
1479 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1480
1481 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1482                                  struct dt_device *key,
1483                                  struct lfsck_req_local *lrl,
1484                                  struct thandle *th) = NULL;
1485
1486 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1487                                                       struct dt_device *,
1488                                                       struct lfsck_req_local *,
1489                                                       struct thandle *))
1490 {
1491         tgt_lfsck_in_notify_local = notify;
1492 }
1493 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1494
1495 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1496                            struct dt_device *key,
1497                            struct lfsck_request *lr) = NULL;
1498
1499 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1500                                                 struct dt_device *,
1501                                                 struct lfsck_request *))
1502 {
1503         tgt_lfsck_in_notify = notify;
1504 }
1505 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1506
1507 static int (*tgt_lfsck_query)(const struct lu_env *env,
1508                               struct dt_device *key,
1509                               struct lfsck_request *req,
1510                               struct lfsck_reply *rep,
1511                               struct lfsck_query *que) = NULL;
1512
1513 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1514                                            struct dt_device *,
1515                                            struct lfsck_request *,
1516                                            struct lfsck_reply *,
1517                                            struct lfsck_query *))
1518 {
1519         tgt_lfsck_query = query;
1520 }
1521 EXPORT_SYMBOL(tgt_register_lfsck_query);
1522
1523 /* LFSCK request handlers */
1524 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1525 {
1526         const struct lu_env     *env = tsi->tsi_env;
1527         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1528         struct lfsck_request    *lr;
1529         int                      rc;
1530         ENTRY;
1531
1532         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1533         if (lr == NULL)
1534                 RETURN(-EPROTO);
1535
1536         rc = tgt_lfsck_in_notify(env, key, lr);
1537
1538         RETURN(rc);
1539 }
1540
1541 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1542 {
1543         struct lfsck_request    *request;
1544         struct lfsck_reply      *reply;
1545         int                      rc;
1546         ENTRY;
1547
1548         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1549         if (request == NULL)
1550                 RETURN(-EPROTO);
1551
1552         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1553         if (reply == NULL)
1554                 RETURN(-ENOMEM);
1555
1556         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1557                              request, reply, NULL);
1558
1559         RETURN(rc < 0 ? rc : 0);
1560 }
1561
1562 struct tgt_handler tgt_lfsck_handlers[] = {
1563 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1564 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1565 };
1566 EXPORT_SYMBOL(tgt_lfsck_handlers);
1567
1568 /*
1569  * initialize per-thread page pool (bug 5137).
1570  */
1571 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1572 {
1573         struct tgt_thread_big_cache *tbc;
1574
1575         ENTRY;
1576
1577         LASSERT(thread != NULL);
1578         LASSERT(thread->t_data == NULL);
1579
1580         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1581         if (tbc == NULL)
1582                 RETURN(-ENOMEM);
1583         thread->t_data = tbc;
1584         RETURN(0);
1585 }
1586 EXPORT_SYMBOL(tgt_io_thread_init);
1587
1588 /*
1589  * free per-thread pool created by tgt_thread_init().
1590  */
1591 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1592 {
1593         struct tgt_thread_big_cache *tbc;
1594
1595         ENTRY;
1596
1597         LASSERT(thread != NULL);
1598
1599         /*
1600          * be prepared to handle partially-initialized pools (because this is
1601          * called from ost_io_thread_init() for cleanup.
1602          */
1603         tbc = thread->t_data;
1604         if (tbc != NULL) {
1605                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1606                 thread->t_data = NULL;
1607         }
1608         EXIT;
1609 }
1610 EXPORT_SYMBOL(tgt_io_thread_done);
1611
1612 /**
1613  * Helper function for getting Data-on-MDT file server DLM lock
1614  * if asked by client.
1615  */
1616 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1617                       struct lustre_handle *lh, int mode, __u64 *flags)
1618 {
1619         union ldlm_policy_data policy = {
1620                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1621         };
1622         int rc;
1623
1624         ENTRY;
1625
1626         LASSERT(lh != NULL);
1627         LASSERT(ns != NULL);
1628         LASSERT(!lustre_handle_is_used(lh));
1629
1630         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1631                                     flags, ldlm_blocking_ast,
1632                                     ldlm_completion_ast, ldlm_glimpse_ast,
1633                                     NULL, 0, LVB_T_NONE, NULL, lh);
1634
1635         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1636 }
1637 EXPORT_SYMBOL(tgt_mdt_data_lock);
1638
1639 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1640 {
1641         LASSERT(lustre_handle_is_used(lh));
1642         ldlm_lock_decref(lh, mode);
1643 }
1644 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1645
1646 /**
1647  * Helper function for getting server side [start, start+count] DLM lock
1648  * if asked by client.
1649  */
1650 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1651                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1652                     struct lustre_handle *lh, int mode, __u64 *flags)
1653 {
1654         union ldlm_policy_data policy;
1655         int rc;
1656
1657         ENTRY;
1658
1659         LASSERT(lh != NULL);
1660         LASSERT(ns != NULL);
1661         LASSERT(!lustre_handle_is_used(lh));
1662
1663         policy.l_extent.gid = 0;
1664         policy.l_extent.start = start & PAGE_MASK;
1665
1666         /*
1667          * If ->o_blocks is EOF it means "lock till the end of the file".
1668          * Otherwise, it's size of an extent or hole being punched (in bytes).
1669          */
1670         if (end == OBD_OBJECT_EOF || end < start)
1671                 policy.l_extent.end = OBD_OBJECT_EOF;
1672         else
1673                 policy.l_extent.end = end | ~PAGE_MASK;
1674
1675         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1676                                     mode, flags, ldlm_blocking_ast,
1677                                     ldlm_completion_ast, ldlm_glimpse_ast,
1678                                     NULL, 0, LVB_T_NONE, NULL, lh);
1679         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1680 }
1681 EXPORT_SYMBOL(tgt_extent_lock);
1682
1683 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1684 {
1685         LASSERT(lustre_handle_is_used(lh));
1686         ldlm_lock_decref(lh, mode);
1687 }
1688 EXPORT_SYMBOL(tgt_extent_unlock);
1689
1690 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1691                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1692                         struct niobuf_remote *nb, struct lustre_handle *lh,
1693                         enum ldlm_mode mode)
1694 {
1695         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1696         __u64                    flags = 0;
1697         int                      nrbufs = obj->ioo_bufcnt;
1698         int                      i;
1699         int                      rc;
1700
1701         ENTRY;
1702
1703         LASSERT(mode == LCK_PR || mode == LCK_PW);
1704         LASSERT(!lustre_handle_is_used(lh));
1705
1706         if (ns->ns_obd->obd_recovering)
1707                 RETURN(0);
1708
1709         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1710                 RETURN(0);
1711
1712         for (i = 1; i < nrbufs; i++)
1713                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1714                         RETURN(-EFAULT);
1715
1716         /* MDT IO for data-on-mdt */
1717         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1718                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1719         else
1720                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1721                                      nb[nrbufs - 1].rnb_offset +
1722                                      nb[nrbufs - 1].rnb_len - 1,
1723                                      lh, mode, &flags);
1724         RETURN(rc);
1725 }
1726
1727 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1728                            struct lustre_handle *lh, enum ldlm_mode mode)
1729 {
1730         ENTRY;
1731
1732         LASSERT(mode == LCK_PR || mode == LCK_PW);
1733         LASSERT((obj->ioo_bufcnt > 0 &&
1734                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1735                 lustre_handle_is_used(lh));
1736
1737         if (lustre_handle_is_used(lh))
1738                 tgt_extent_unlock(lh, mode);
1739         EXIT;
1740 }
1741 static int tgt_checksum_niobuf(struct lu_target *tgt,
1742                                  struct niobuf_local *local_nb, int npages,
1743                                  int opc, enum cksum_types cksum_type,
1744                                  __u32 *cksum)
1745 {
1746         struct ahash_request           *req;
1747         unsigned int                    bufsize;
1748         int                             i, err;
1749         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1750
1751         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1752         if (IS_ERR(req)) {
1753                 CERROR("%s: unable to initialize checksum hash %s\n",
1754                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1755                 return PTR_ERR(req);
1756         }
1757
1758         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1759         for (i = 0; i < npages; i++) {
1760                 /* corrupt the data before we compute the checksum, to
1761                  * simulate a client->OST data error */
1762                 if (i == 0 && opc == OST_WRITE &&
1763                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1764                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1765                         int len = local_nb[i].lnb_len;
1766                         struct page *np = tgt_page_to_corrupt;
1767
1768                         if (np) {
1769                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1770                                                         KM_USER0);
1771                                 char *ptr2 = page_address(np);
1772
1773                                 memcpy(ptr2 + off, ptr + off, len);
1774                                 memcpy(ptr2 + off, "bad3", min(4, len));
1775                                 ll_kunmap_atomic(ptr, KM_USER0);
1776
1777                                 /* LU-8376 to preserve original index for
1778                                  * display in dump_all_bulk_pages() */
1779                                 np->index = i;
1780
1781                                 cfs_crypto_hash_update_page(req, np, off,
1782                                                             len);
1783                                 continue;
1784                         } else {
1785                                 CERROR("%s: can't alloc page for corruption\n",
1786                                        tgt_name(tgt));
1787                         }
1788                 }
1789                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1790                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1791                                   local_nb[i].lnb_len);
1792
1793                  /* corrupt the data after we compute the checksum, to
1794                  * simulate an OST->client data error */
1795                 if (i == 0 && opc == OST_READ &&
1796                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1797                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1798                         int len = local_nb[i].lnb_len;
1799                         struct page *np = tgt_page_to_corrupt;
1800
1801                         if (np) {
1802                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1803                                                         KM_USER0);
1804                                 char *ptr2 = page_address(np);
1805
1806                                 memcpy(ptr2 + off, ptr + off, len);
1807                                 memcpy(ptr2 + off, "bad4", min(4, len));
1808                                 ll_kunmap_atomic(ptr, KM_USER0);
1809
1810                                 /* LU-8376 to preserve original index for
1811                                  * display in dump_all_bulk_pages() */
1812                                 np->index = i;
1813
1814                                 cfs_crypto_hash_update_page(req, np, off,
1815                                                             len);
1816                                 continue;
1817                         } else {
1818                                 CERROR("%s: can't alloc page for corruption\n",
1819                                        tgt_name(tgt));
1820                         }
1821                 }
1822         }
1823
1824         bufsize = sizeof(*cksum);
1825         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1826
1827         return 0;
1828 }
1829
1830 char dbgcksum_file_name[PATH_MAX];
1831
1832 static void dump_all_bulk_pages(struct obdo *oa, int count,
1833                                 struct niobuf_local *local_nb,
1834                                 __u32 server_cksum, __u32 client_cksum)
1835 {
1836         struct file *filp;
1837         int rc, i;
1838         unsigned int len;
1839         char *buf;
1840
1841         /* will only keep dump of pages on first error for the same range in
1842          * file/fid, not during the resends/retries. */
1843         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1844                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1845                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1846                   libcfs_debug_file_path_arr :
1847                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1848                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1849                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1850                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1851                  local_nb[0].lnb_file_offset,
1852                  local_nb[count-1].lnb_file_offset +
1853                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1854         filp = filp_open(dbgcksum_file_name,
1855                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1856         if (IS_ERR(filp)) {
1857                 rc = PTR_ERR(filp);
1858                 if (rc == -EEXIST)
1859                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1860                                "checksum error: rc = %d\n", dbgcksum_file_name,
1861                                rc);
1862                 else
1863                         CERROR("%s: can't open to dump pages with checksum "
1864                                "error: rc = %d\n", dbgcksum_file_name, rc);
1865                 return;
1866         }
1867
1868         for (i = 0; i < count; i++) {
1869                 len = local_nb[i].lnb_len;
1870                 buf = kmap(local_nb[i].lnb_page);
1871                 while (len != 0) {
1872                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1873                         if (rc < 0) {
1874                                 CERROR("%s: wanted to write %u but got %d "
1875                                        "error\n", dbgcksum_file_name, len, rc);
1876                                 break;
1877                         }
1878                         len -= rc;
1879                         buf += rc;
1880                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1881                                dbgcksum_file_name, rc);
1882                 }
1883                 kunmap(local_nb[i].lnb_page);
1884         }
1885
1886         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1887         if (rc)
1888                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1889         filp_close(filp, NULL);
1890         return;
1891 }
1892
1893 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1894                                struct obd_export *exp, struct obdo *oa,
1895                                const struct lnet_process_id *peer,
1896                                __u32 client_cksum, __u32 server_cksum,
1897                                enum cksum_types server_cksum_type)
1898 {
1899         char *msg;
1900         enum cksum_types cksum_type;
1901         loff_t start, end;
1902
1903         /* unlikely to happen and only if resend does not occur due to cksum
1904          * control failure on Client */
1905         if (unlikely(server_cksum == client_cksum)) {
1906                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1907                        client_cksum);
1908                 return 0;
1909         }
1910
1911         if (exp->exp_obd->obd_checksum_dump)
1912                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1913                                     client_cksum);
1914
1915         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1916                                            oa->o_flags : 0);
1917
1918         if (cksum_type != server_cksum_type)
1919                 msg = "the server may have not used the checksum type specified"
1920                       " in the original request - likely a protocol problem";
1921         else
1922                 msg = "should have changed on the client or in transit";
1923
1924         start = local_nb[0].lnb_file_offset;
1925         end = local_nb[npages-1].lnb_file_offset +
1926                                         local_nb[npages-1].lnb_len - 1;
1927
1928         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1929                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1930                 " %x (type %x), server csum %x (type %x)\n",
1931                 exp->exp_obd->obd_name,
1932                 msg, libcfs_nid2str(peer->nid),
1933                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1934                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1935                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1936                 POSTID(&oa->o_oi),
1937                 start, end, client_cksum, cksum_type, server_cksum,
1938                 server_cksum_type);
1939
1940         return 1;
1941 }
1942
1943 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1944                              unsigned char *buf, int size)
1945 {
1946         int     i, off, len, copied = size;
1947         char    *ptr;
1948
1949         for (i = 0; i < npages; i++) {
1950                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1951                 len = local[i].lnb_len;
1952
1953                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1954                        i, off, len, size);
1955                 if (len > size)
1956                         return -EINVAL;
1957
1958                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1959                 memcpy(buf + off, ptr, len);
1960                 ll_kunmap_atomic(ptr, KM_USER0);
1961                 buf += len;
1962                 size -= len;
1963         }
1964         return copied - size;
1965 }
1966
1967 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1968                                      struct niobuf_local *local_nb,
1969                                      int npages, int opc,
1970                                      obd_dif_csum_fn *fn,
1971                                      int sector_size,
1972                                      u32 *check_sum)
1973 {
1974         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1975         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1976         const char *obd_name = tgt->lut_obd->obd_name;
1977         struct ahash_request *req;
1978         unsigned int bufsize;
1979         unsigned char *buffer;
1980         struct page *__page;
1981         __u16 *guard_start;
1982         int guard_number;
1983         int used_number = 0;
1984         __u32 cksum;
1985         int rc = 0;
1986         int used;
1987         int i;
1988
1989         __page = alloc_page(GFP_KERNEL);
1990         if (__page == NULL)
1991                 return -ENOMEM;
1992
1993         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1994         if (IS_ERR(req)) {
1995                 CERROR("%s: unable to initialize checksum hash %s\n",
1996                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1997                 return PTR_ERR(req);
1998         }
1999
2000         buffer = kmap(__page);
2001         guard_start = (__u16 *)buffer;
2002         guard_number = PAGE_SIZE / sizeof(*guard_start);
2003         for (i = 0; i < npages; i++) {
2004                 /* corrupt the data before we compute the checksum, to
2005                  * simulate a client->OST data error */
2006                 if (i == 0 && opc == OST_WRITE &&
2007                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2008                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2009                         int len = local_nb[i].lnb_len;
2010                         struct page *np = tgt_page_to_corrupt;
2011
2012                         if (np) {
2013                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2014                                                         KM_USER0);
2015                                 char *ptr2 = page_address(np);
2016
2017                                 memcpy(ptr2 + off, ptr + off, len);
2018                                 memcpy(ptr2 + off, "bad3", min(4, len));
2019                                 ll_kunmap_atomic(ptr, KM_USER0);
2020
2021                                 /* LU-8376 to preserve original index for
2022                                  * display in dump_all_bulk_pages() */
2023                                 np->index = i;
2024
2025                                 cfs_crypto_hash_update_page(req, np, off,
2026                                                             len);
2027                                 continue;
2028                         } else {
2029                                 CERROR("%s: can't alloc page for corruption\n",
2030                                        tgt_name(tgt));
2031                         }
2032                 }
2033
2034                 /*
2035                  * The left guard number should be able to hold checksums of a
2036                  * whole page
2037                  */
2038                 if (t10_cksum_type && opc == OST_READ &&
2039                     local_nb[i].lnb_guard_disk) {
2040                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2041                         if (used > (guard_number - used_number)) {
2042                                 rc = -E2BIG;
2043                                 break;
2044                         }
2045                         memcpy(guard_start + used_number,
2046                                local_nb[i].lnb_guards,
2047                                used * sizeof(*local_nb[i].lnb_guards));
2048                 } else {
2049                         rc = obd_page_dif_generate_buffer(obd_name,
2050                                 local_nb[i].lnb_page,
2051                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2052                                 local_nb[i].lnb_len, guard_start + used_number,
2053                                 guard_number - used_number, &used, sector_size,
2054                                 fn);
2055                         if (rc)
2056                                 break;
2057                 }
2058
2059                 LASSERT(used <= MAX_GUARD_NUMBER);
2060                 /*
2061                  * If disk support T10PI checksum, copy guards to local_nb.
2062                  * If the write is partial page, do not use the guards for bio
2063                  * submission since the data might not be full-sector. The bio
2064                  * guards will be generated later based on the full sectors. If
2065                  * the sector size is 512B rather than 4 KB, or the page size
2066                  * is larger than 4KB, this might drop some useful guards for
2067                  * partial page write, but it will only add minimal extra time
2068                  * of checksum calculation.
2069                  */
2070                 if (t10_cksum_type && opc == OST_WRITE &&
2071                     local_nb[i].lnb_len == PAGE_SIZE) {
2072                         local_nb[i].lnb_guard_rpc = 1;
2073                         memcpy(local_nb[i].lnb_guards,
2074                                guard_start + used_number,
2075                                used * sizeof(*local_nb[i].lnb_guards));
2076                 }
2077
2078                 used_number += used;
2079                 if (used_number == guard_number) {
2080                         cfs_crypto_hash_update_page(req, __page, 0,
2081                                 used_number * sizeof(*guard_start));
2082                         used_number = 0;
2083                 }
2084
2085                  /* corrupt the data after we compute the checksum, to
2086                  * simulate an OST->client data error */
2087                 if (unlikely(i == 0 && opc == OST_READ &&
2088                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2089                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2090                         int len = local_nb[i].lnb_len;
2091                         struct page *np = tgt_page_to_corrupt;
2092
2093                         if (np) {
2094                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2095                                                         KM_USER0);
2096                                 char *ptr2 = page_address(np);
2097
2098                                 memcpy(ptr2 + off, ptr + off, len);
2099                                 memcpy(ptr2 + off, "bad4", min(4, len));
2100                                 ll_kunmap_atomic(ptr, KM_USER0);
2101
2102                                 /* LU-8376 to preserve original index for
2103                                  * display in dump_all_bulk_pages() */
2104                                 np->index = i;
2105
2106                                 cfs_crypto_hash_update_page(req, np, off,
2107                                                             len);
2108                                 continue;
2109                         } else {
2110                                 CERROR("%s: can't alloc page for corruption\n",
2111                                        tgt_name(tgt));
2112                         }
2113                 }
2114         }
2115         kunmap(__page);
2116         if (rc)
2117                 GOTO(out, rc);
2118
2119         if (used_number != 0)
2120                 cfs_crypto_hash_update_page(req, __page, 0,
2121                         used_number * sizeof(*guard_start));
2122
2123         bufsize = sizeof(cksum);
2124         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2125
2126         if (rc == 0)
2127                 *check_sum = cksum;
2128 out:
2129         __free_page(__page);
2130         return rc;
2131 }
2132
2133 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2134                                   enum cksum_types cksum_type,
2135                                   struct niobuf_local *local_nb,
2136                                   int npages, int opc, u32 *check_sum)
2137 {
2138         obd_dif_csum_fn *fn = NULL;
2139         int sector_size = 0;
2140         int rc;
2141
2142         ENTRY;
2143         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2144
2145         if (fn)
2146                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2147                                                opc, fn, sector_size,
2148                                                check_sum);
2149         else
2150                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2151                                          cksum_type, check_sum);
2152         RETURN(rc);
2153 }
2154
2155 int tgt_brw_read(struct tgt_session_info *tsi)
2156 {
2157         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2158         struct ptlrpc_bulk_desc *desc = NULL;
2159         struct obd_export       *exp = tsi->tsi_exp;
2160         struct niobuf_remote    *remote_nb;
2161         struct niobuf_local     *local_nb;
2162         struct obd_ioobj        *ioo;
2163         struct ost_body         *body, *repbody;
2164         struct lustre_handle     lockh = { 0 };
2165         int                      npages, nob = 0, rc, i, no_reply = 0,
2166                                  npages_read;
2167         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2168         const char *obd_name = exp->exp_obd->obd_name;
2169
2170         ENTRY;
2171
2172         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2173             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2174                 CERROR("%s: deny read request from %s to portal %u\n",
2175                        tgt_name(tsi->tsi_tgt),
2176                        obd_export_nid2str(req->rq_export),
2177                        ptlrpc_req2svc(req)->srv_req_portal);
2178                 RETURN(-EPROTO);
2179         }
2180
2181         req->rq_bulk_read = 1;
2182
2183         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2184                 RETURN(-EIO);
2185
2186         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2187                          cfs_fail_val : (obd_timeout + 1) / 4);
2188
2189         /* Check if there is eviction in progress, and if so, wait for it to
2190          * finish */
2191         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2192                 /* We do not care how long it takes */
2193                 wait_event_idle(
2194                         exp->exp_obd->obd_evict_inprogress_waitq,
2195                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2196         }
2197
2198         /* There must be big cache in current thread to process this request
2199          * if it is NULL then something went wrong and it wasn't allocated,
2200          * report -ENOMEM in that case */
2201         if (tbc == NULL)
2202                 RETURN(-ENOMEM);
2203
2204         body = tsi->tsi_ost_body;
2205         LASSERT(body != NULL);
2206
2207         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2208         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2209
2210         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2211         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2212
2213         local_nb = tbc->local;
2214
2215         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2216                           &lockh, LCK_PR);
2217         if (rc != 0)
2218                 RETURN(rc);
2219
2220         /*
2221          * If getting the lock took more time than
2222          * client was willing to wait, drop it. b=11330
2223          */
2224         if (ktime_get_real_seconds() > req->rq_deadline ||
2225             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2226                 no_reply = 1;
2227                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2228                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2229                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2230                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2231                 GOTO(out_lock, rc = -ETIMEDOUT);
2232         }
2233
2234         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2235         repbody->oa = body->oa;
2236
2237         npages = PTLRPC_MAX_BRW_PAGES;
2238         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2239                         ioo, remote_nb, &npages, local_nb);
2240         if (rc != 0)
2241                 GOTO(out_lock, rc);
2242
2243         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2244             body->oa.o_flags & OBD_FL_SHORT_IO) {
2245                 desc = NULL;
2246         } else {
2247                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2248                                             PTLRPC_BULK_PUT_SOURCE |
2249                                                 PTLRPC_BULK_BUF_KIOV,
2250                                             OST_BULK_PORTAL,
2251                                             &ptlrpc_bulk_kiov_nopin_ops);
2252                 if (desc == NULL)
2253                         GOTO(out_commitrw, rc = -ENOMEM);
2254         }
2255
2256         nob = 0;
2257         npages_read = npages;
2258         for (i = 0; i < npages; i++) {
2259                 int page_rc = local_nb[i].lnb_rc;
2260
2261                 if (page_rc < 0) {
2262                         rc = page_rc;
2263                         npages_read = i;
2264                         break;
2265                 }
2266
2267                 nob += page_rc;
2268                 if (page_rc != 0 && desc != NULL) { /* some data! */
2269                         LASSERT(local_nb[i].lnb_page != NULL);
2270                         desc->bd_frag_ops->add_kiov_frag
2271                           (desc, local_nb[i].lnb_page,
2272                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2273                            page_rc);
2274                 }
2275
2276                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2277                         local_nb[i].lnb_len = page_rc;
2278                         npages_read = i + (page_rc != 0 ? 1 : 0);
2279                         /* All subsequent pages should be 0 */
2280                         while (++i < npages)
2281                                 LASSERT(local_nb[i].lnb_rc == 0);
2282                         break;
2283                 }
2284         }
2285         if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2286             nob != cfs_fail_val)
2287                 rc = -E2BIG;
2288
2289         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2290                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2291                            body->oa.o_flags : 0;
2292                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2293
2294                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2295                                                           cksum_type);
2296                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2297
2298                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2299                                             local_nb, npages_read, OST_READ,
2300                                             &repbody->oa.o_cksum);
2301                 if (rc < 0)
2302                         GOTO(out_commitrw, rc);
2303                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2304                        repbody->oa.o_cksum);
2305
2306                 /* if a resend it could be for a cksum error, so check Server
2307                  * cksum with returned Client cksum (this should even cover
2308                  * zero-cksum case) */
2309                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2310                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2311                         check_read_checksum(local_nb, npages_read, exp,
2312                                             &body->oa, &req->rq_peer,
2313                                             body->oa.o_cksum,
2314                                             repbody->oa.o_cksum, cksum_type);
2315         } else {
2316                 repbody->oa.o_valid = 0;
2317         }
2318         /* We're finishing using body->oa as an input variable */
2319
2320         /* Check if client was evicted while we were doing i/o before touching
2321          * network */
2322         if (rc == 0) {
2323                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2324                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2325                         unsigned char *short_io_buf;
2326                         int short_io_size;
2327
2328                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2329                                                               &RMF_SHORT_IO);
2330                         short_io_size = req_capsule_get_size(&req->rq_pill,
2331                                                              &RMF_SHORT_IO,
2332                                                              RCL_SERVER);
2333                         rc = tgt_pages2shortio(local_nb, npages_read,
2334                                                short_io_buf, short_io_size);
2335                         if (rc >= 0)
2336                                 req_capsule_shrink(&req->rq_pill,
2337                                                    &RMF_SHORT_IO, rc,
2338                                                    RCL_SERVER);
2339                         rc = rc > 0 ? 0 : rc;
2340                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2341                         rc = target_bulk_io(exp, desc);
2342                 }
2343                 no_reply = rc != 0;
2344         } else {
2345                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2346                     body->oa.o_flags & OBD_FL_SHORT_IO)
2347                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2348                                            RCL_SERVER);
2349         }
2350
2351 out_commitrw:
2352         /* Must commit after prep above in all cases */
2353         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2354                           remote_nb, npages, local_nb, rc);
2355 out_lock:
2356         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2357
2358         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2359                 ptlrpc_free_bulk(desc);
2360
2361         LASSERT(rc <= 0);
2362         if (rc == 0) {
2363                 rc = nob;
2364                 ptlrpc_lprocfs_brw(req, nob);
2365         } else if (no_reply) {
2366                 req->rq_no_reply = 1;
2367                 /* reply out callback would free */
2368                 ptlrpc_req_drop_rs(req);
2369                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2370                               "client will retry: rc %d\n",
2371                               obd_name,
2372                               obd_uuid2str(&exp->exp_client_uuid),
2373                               obd_export_nid2str(exp), rc);
2374         }
2375         /* send a bulk after reply to simulate a network delay or reordering
2376          * by a router - Note that !desc implies short io, so there is no bulk
2377          * to reorder. */
2378         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2379             desc) {
2380                 CDEBUG(D_INFO, "reorder BULK\n");
2381
2382                 ssleep(3);
2383                 target_bulk_io(exp, desc);
2384                 ptlrpc_free_bulk(desc);
2385         }
2386
2387         RETURN(rc);
2388 }
2389 EXPORT_SYMBOL(tgt_brw_read);
2390
2391 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2392                              unsigned char *buf, unsigned int size)
2393 {
2394         int     i, off, len;
2395         char    *ptr;
2396
2397         for (i = 0; i < npages; i++) {
2398                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2399                 len = local[i].lnb_len;
2400
2401                 if (len == 0)
2402                         continue;
2403
2404                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2405                        i, off, len, size);
2406                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2407                 if (ptr == NULL)
2408                         return -EINVAL;
2409                 memcpy(ptr + off, buf, len < size ? len : size);
2410                 ll_kunmap_atomic(ptr, KM_USER0);
2411                 buf += len;
2412                 size -= len;
2413         }
2414         return 0;
2415 }
2416
2417 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2418                               struct ptlrpc_bulk_desc *desc,
2419                               struct niobuf_local *local_nb, int npages,
2420                               u32 client_cksum, u32 server_cksum,
2421                               bool mmap)
2422 {
2423         struct obd_export *exp = req->rq_export;
2424         struct ost_body *body;
2425         char *router = "";
2426         char *via = "";
2427
2428         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2429         LASSERT(body != NULL);
2430
2431         if (desc && req->rq_peer.nid != desc->bd_sender) {
2432                 via = " via ";
2433                 router = libcfs_nid2str(desc->bd_sender);
2434         }
2435
2436         if (exp->exp_obd->obd_checksum_dump)
2437                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2438                                     client_cksum);
2439
2440         if (mmap) {
2441                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2442                              client_cksum, server_cksum);
2443                 return;
2444         }
2445
2446         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2447                            DFID" object "DOSTID" extent [%llu-%llu"
2448                            "]: client csum %x, server csum %x\n",
2449                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2450                            via, router,
2451                            body->oa.o_valid & OBD_MD_FLFID ?
2452                            body->oa.o_parent_seq : (__u64)0,
2453                            body->oa.o_valid & OBD_MD_FLFID ?
2454                            body->oa.o_parent_oid : 0,
2455                            body->oa.o_valid & OBD_MD_FLFID ?
2456                            body->oa.o_parent_ver : 0,
2457                            POSTID(&body->oa.o_oi),
2458                            local_nb[0].lnb_file_offset,
2459                            local_nb[npages-1].lnb_file_offset +
2460                            local_nb[npages - 1].lnb_len - 1,
2461                            client_cksum, server_cksum);
2462 }
2463
2464 int tgt_brw_write(struct tgt_session_info *tsi)
2465 {
2466         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2467         struct ptlrpc_bulk_desc *desc = NULL;
2468         struct obd_export       *exp = req->rq_export;
2469         struct niobuf_remote    *remote_nb;
2470         struct niobuf_local     *local_nb;
2471         struct obd_ioobj        *ioo;
2472         struct ost_body         *body, *repbody;
2473         struct lustre_handle     lockh = {0};
2474         __u32                   *rcs;
2475         int                      objcount, niocount, npages;
2476         int                      rc, i, j;
2477         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2478         bool                     no_reply = false, mmap;
2479         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2480         bool wait_sync = false;
2481         const char *obd_name = exp->exp_obd->obd_name;
2482
2483         ENTRY;
2484
2485         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2486             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2487                 CERROR("%s: deny write request from %s to portal %u\n",
2488                        tgt_name(tsi->tsi_tgt),
2489                        obd_export_nid2str(req->rq_export),
2490                        ptlrpc_req2svc(req)->srv_req_portal);
2491                 RETURN(err_serious(-EPROTO));
2492         }
2493
2494         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2495                 RETURN(err_serious(-ENOSPC));
2496         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2497                 RETURN(err_serious(-EROFS));
2498
2499         req->rq_bulk_write = 1;
2500
2501         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2502                 RETURN(err_serious(-EIO));
2503         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2504                 RETURN(err_serious(-EFAULT));
2505
2506         /* pause before transaction has been started */
2507         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2508                          cfs_fail_val : (obd_timeout + 1) / 4);
2509
2510         /* Delay write commit to show stale size information */
2511         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2512
2513         /* There must be big cache in current thread to process this request
2514          * if it is NULL then something went wrong and it wasn't allocated,
2515          * report -ENOMEM in that case */
2516         if (tbc == NULL)
2517                 RETURN(-ENOMEM);
2518
2519         body = tsi->tsi_ost_body;
2520         LASSERT(body != NULL);
2521
2522         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2523         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2524
2525         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2526                                         RCL_CLIENT) / sizeof(*ioo);
2527
2528         for (niocount = i = 0; i < objcount; i++)
2529                 niocount += ioo[i].ioo_bufcnt;
2530
2531         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2532         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2533         if (niocount != req_capsule_get_size(&req->rq_pill,
2534                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2535                         sizeof(*remote_nb))
2536                 RETURN(err_serious(-EPROTO));
2537
2538         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2539             ptlrpc_connection_is_local(exp->exp_connection))
2540                 memory_pressure_set();
2541
2542         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2543                              niocount * sizeof(*rcs));
2544         rc = req_capsule_server_pack(&req->rq_pill);
2545         if (rc != 0)
2546                 GOTO(out, rc = err_serious(rc));
2547
2548         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2549         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2550
2551         local_nb = tbc->local;
2552
2553         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2554                           &lockh, LCK_PW);
2555         if (rc != 0)
2556                 GOTO(out, rc);
2557
2558         /*
2559          * If getting the lock took more time than
2560          * client was willing to wait, drop it. b=11330
2561          */
2562         if (ktime_get_real_seconds() > req->rq_deadline ||
2563             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2564                 no_reply = true;
2565                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2566                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2567                        POSTID(&ioo->ioo_oid),
2568                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2569                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2570                 GOTO(out_lock, rc = -ETIMEDOUT);
2571         }
2572
2573         /* Because we already sync grant info with client when reconnect,
2574          * grant info will be cleared for resent req, then fed_grant and
2575          * total_grant will not be modified in following preprw_write */
2576         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2577                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2578                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2579         }
2580
2581         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2582         if (repbody == NULL)
2583                 GOTO(out_lock, rc = -ENOMEM);
2584         repbody->oa = body->oa;
2585
2586         npages = PTLRPC_MAX_BRW_PAGES;
2587         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2588                         objcount, ioo, remote_nb, &npages, local_nb);
2589         if (rc < 0)
2590                 GOTO(out_lock, rc);
2591         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2592             body->oa.o_flags & OBD_FL_SHORT_IO) {
2593                 unsigned int short_io_size;
2594                 unsigned char *short_io_buf;
2595
2596                 short_io_size = req_capsule_get_size(&req->rq_pill,
2597                                                      &RMF_SHORT_IO,
2598                                                      RCL_CLIENT);
2599                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2600                                                       &RMF_SHORT_IO);
2601                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2602                                " size = %d\n", short_io_size);
2603
2604                 /* Copy short io buf to pages */
2605                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2606                                        short_io_size);
2607                 desc = NULL;
2608         } else {
2609                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2610                                             PTLRPC_BULK_GET_SINK |
2611                                             PTLRPC_BULK_BUF_KIOV,
2612                                             OST_BULK_PORTAL,
2613                                             &ptlrpc_bulk_kiov_nopin_ops);
2614                 if (desc == NULL)
2615                         GOTO(skip_transfer, rc = -ENOMEM);
2616
2617                 /* NB Having prepped, we must commit... */
2618                 for (i = 0; i < npages; i++)
2619                         desc->bd_frag_ops->add_kiov_frag(desc,
2620                                         local_nb[i].lnb_page,
2621                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2622                                         local_nb[i].lnb_len);
2623
2624                 rc = sptlrpc_svc_prep_bulk(req, desc);
2625                 if (rc != 0)
2626                         GOTO(skip_transfer, rc);
2627
2628                 rc = target_bulk_io(exp, desc);
2629         }
2630
2631         no_reply = rc != 0;
2632
2633 skip_transfer:
2634         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2635                 static int cksum_counter;
2636
2637                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2638                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2639
2640                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2641                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2642                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2643                                                            cksum_type);
2644
2645                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2646                                             local_nb, npages, OST_WRITE,
2647                                             &repbody->oa.o_cksum);
2648                 if (rc < 0)
2649                         GOTO(out_commitrw, rc);
2650
2651                 cksum_counter++;
2652
2653                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2654                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2655                                 body->oa.o_flags & OBD_FL_MMAP);
2656
2657                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2658                                           body->oa.o_cksum,
2659                                           repbody->oa.o_cksum, mmap);
2660                         cksum_counter = 0;
2661                 } else if ((cksum_counter & (-cksum_counter)) ==
2662                            cksum_counter) {
2663                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2664                                cksum_counter, libcfs_id2str(req->rq_peer),
2665                                repbody->oa.o_cksum);
2666                 }
2667         }
2668
2669 out_commitrw:
2670         /* Must commit after prep above in all cases */
2671         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2672                           objcount, ioo, remote_nb, npages, local_nb, rc);
2673         if (rc == -ENOTCONN)
2674                 /* quota acquire process has been given up because
2675                  * either the client has been evicted or the client
2676                  * has timed out the request already */
2677                 no_reply = true;
2678
2679         for (i = 0; i < niocount; i++) {
2680                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2681                         wait_sync = true;
2682                         break;
2683                 }
2684         }
2685         /*
2686          * Disable sending mtime back to the client. If the client locked the
2687          * whole object, then it has already updated the mtime on its side,
2688          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2689          */
2690         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2691
2692         if (rc == 0) {
2693                 int nob = 0;
2694
2695                 /* set per-requested niobuf return codes */
2696                 for (i = j = 0; i < niocount; i++) {
2697                         int len = remote_nb[i].rnb_len;
2698
2699                         nob += len;
2700                         rcs[i] = 0;
2701                         do {
2702                                 LASSERT(j < npages);
2703                                 if (local_nb[j].lnb_rc < 0)
2704                                         rcs[i] = local_nb[j].lnb_rc;
2705                                 len -= local_nb[j].lnb_len;
2706                                 j++;
2707                         } while (len > 0);
2708                         LASSERT(len == 0);
2709                 }
2710                 LASSERT(j == npages);
2711                 ptlrpc_lprocfs_brw(req, nob);
2712         }
2713 out_lock:
2714         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2715         if (desc)
2716                 ptlrpc_free_bulk(desc);
2717 out:
2718         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2719                 req->rq_no_reply = 1;
2720                 /* reply out callback would free */
2721                 ptlrpc_req_drop_rs(req);
2722                 if (!exp->exp_obd->obd_no_transno)
2723                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2724                                       " client will retry: rc = %d\n",
2725                                       obd_name,
2726                                       obd_uuid2str(&exp->exp_client_uuid),
2727                                       obd_export_nid2str(exp), rc);
2728         }
2729         memory_pressure_clr();
2730         RETURN(rc);
2731 }
2732 EXPORT_SYMBOL(tgt_brw_write);
2733
2734 /* Check if request can be reconstructed from saved reply data
2735  * A copy of the reply data is returned in @trd if the pointer is not NULL
2736  */
2737 int req_can_reconstruct(struct ptlrpc_request *req,
2738                          struct tg_reply_data *trd)
2739 {
2740         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2741         struct lsd_client_data *lcd = ted->ted_lcd;
2742         int found;
2743
2744         if (tgt_is_multimodrpcs_client(req->rq_export))
2745                 return tgt_lookup_reply(req, trd);
2746
2747         mutex_lock(&ted->ted_lcd_lock);
2748         found = req->rq_xid == lcd->lcd_last_xid ||
2749                 req->rq_xid == lcd->lcd_last_close_xid;
2750
2751         if (found && trd != NULL) {
2752                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2753                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2754                         trd->trd_reply.lrd_transno =
2755                                                 lcd->lcd_last_close_transno;
2756                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2757                 } else {
2758                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2759                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2760                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2761                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2762                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2763                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2764                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2765                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2766                 }
2767         }
2768         mutex_unlock(&ted->ted_lcd_lock);
2769
2770         return found;
2771 }
2772 EXPORT_SYMBOL(req_can_reconstruct);
2773