lustre/target/tgt_handler.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2013, 2017, Intel Corporation.
  25  */
  26 /*
  27  * lustre/target/tgt_handler.c
  28  *
  29  * Lustre Unified Target request handler code
  30  *
  31  * Author: Brian Behlendorf <behlendorf1@llnl.gov>
  32  * Author: Mikhail Pershin <mike.pershin@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_CLASS
  36
  37 #include <linux/user_namespace.h>
  38 #include <linux/delay.h>
  39 #include <linux/uidgid.h>
  40
  41 #include <libcfs/linux/linux-mem.h>
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47 #include <lustre_acl.h>
  48
  49 #include "tgt_internal.h"
  50
  51 char *tgt_name(struct lu_target *tgt)
  52 {
  53         LASSERT(tgt->lut_obd != NULL);
  54         return tgt->lut_obd->obd_name;
  55 }
  56 EXPORT_SYMBOL(tgt_name);
  57
  58 /*
  59  * Generic code handling requests that have struct mdt_body passed in:
  60  *
  61  *  - extract mdt_body from request and save it in @tsi, if present;
  62  *
  63  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  64  *  @tsi;
  65  *
  66  *  - if HAS_BODY flag is set for this request type check whether object
  67  *  actually exists on storage (lu_object_exists()).
  68  *
  69  */
  70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  71 {
  72         const struct mdt_body   *body;
  73         struct lu_object        *obj;
  74         struct req_capsule      *pill = tsi->tsi_pill;
  75         int                      rc;
  76
  77         ENTRY;
  78
  79         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EFAULT);
  82
  83         tsi->tsi_mdt_body = body;
  84
  85         if (!(body->mbo_valid & OBD_MD_FLID))
  86                 RETURN(0);
  87
  88         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  89          * in any case in pre-2.5 clients. Fix that here if needed */
  90         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  91                 RETURN(0);
  92
  93         if (!fid_is_sane(&body->mbo_fid1)) {
  94                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  95                        PFID(&body->mbo_fid1));
  96                 RETURN(-EINVAL);
  97         }
  98
  99         obj = lu_object_find(tsi->tsi_env,
 100                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 101                              &body->mbo_fid1, NULL);
 102         if (!IS_ERR(obj)) {
 103                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 104                         lu_object_put(tsi->tsi_env, obj);
 105                         rc = -ENOENT;
 106                 } else {
 107                         tsi->tsi_corpus = obj;
 108                         rc = 0;
 109                 }
 110         } else {
 111                 rc = PTR_ERR(obj);
 112         }
 113
 114         tsi->tsi_fid = body->mbo_fid1;
 115
 116         RETURN(rc);
 117 }
 118
 119 /**
 120  * Validate oa from client.
 121  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 122  * req are valid.
 123  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 124  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 125  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 126  *       whether this is oi_fid or real ostid. So it will check
 127  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 128  *    c. Old FID-disable osc will send IDIF.
 129  *    d. new FID-enable osc/osp will send normal FID.
 130  *
 131  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 132  * be used for LAST_ID file, and only being accessed inside OST now.
 133  */
 134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 135 {
 136         struct ost_id   *oi     = &oa->o_oi;
 137         u64              seq    = ostid_seq(oi);
 138         u64              id     = ostid_id(oi);
 139         int              rc;
 140         ENTRY;
 141
 142         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 143                      fid_seq_is_echo(seq))) {
 144                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 145                  * during create, and we will reset this to 1, since this
 146                  * oi_id is basically useless in the following create process,
 147                  * but oi_id == 0 will make it difficult to tell whether it is
 148                  * real FID or ost_id. */
 149                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 150                 oi->oi_fid.f_oid = id ?: 1;
 151                 oi->oi_fid.f_ver = 0;
 152         } else {
 153                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 154
 155                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 156                         GOTO(out, rc = -EPROTO);
 157
 158                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 159                  * all of the requests are required to setup FLGROUP */
 160                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 161                         ostid_set_seq_mdt0(oi);
 162                         oa->o_valid |= OBD_MD_FLGROUP;
 163                         seq = ostid_seq(oi);
 164                 }
 165
 166                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 167                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 168                         GOTO(out, rc = -EPROTO);
 169
 170                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 171                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 172                 if (unlikely(rc != 0))
 173                         GOTO(out, rc);
 174
 175                 oi->oi_fid = tti->tti_fid1;
 176         }
 177
 178         RETURN(0);
 179
 180 out:
 181         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 182                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 183                seq, id, rc);
 184         return rc;
 185 }
 186 EXPORT_SYMBOL(tgt_validate_obdo);
 187
 188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 189 {
 190         unsigned                 max_brw;
 191         struct niobuf_remote    *rnb;
 192         struct obd_ioobj        *ioo;
 193         int                      obj_count;
 194
 195         ENTRY;
 196
 197         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 198         if (ioo == NULL)
 199                 RETURN(-EPROTO);
 200
 201         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 202         if (rnb == NULL)
 203                 RETURN(-EPROTO);
 204
 205         max_brw = ioobj_max_brw_get(ioo);
 206         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 207                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 208                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 209                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 210                        POSTID(oi), -EPROTO);
 211                 RETURN(-EPROTO);
 212         }
 213         ioo->ioo_oid = *oi;
 214
 215         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 216                                         RCL_CLIENT) / sizeof(*ioo);
 217         if (obj_count == 0) {
 218                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 219                 RETURN(-EPROTO);
 220         } else if (obj_count > 1) {
 221                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 222                        obj_count);
 223                 RETURN(-EPROTO);
 224         }
 225
 226         if (ioo->ioo_bufcnt == 0) {
 227                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 228                 RETURN(-EPROTO);
 229         }
 230
 231         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 232                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 233                           "bulk has too many pages (%d)",
 234                           ioo->ioo_bufcnt);
 235                 RETURN(-EPROTO);
 236         }
 237
 238         RETURN(0);
 239 }
 240
 241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 242 {
 243         struct ost_body         *body;
 244         struct req_capsule      *pill = tsi->tsi_pill;
 245         struct lu_nodemap       *nodemap;
 246         int                      rc;
 247
 248         ENTRY;
 249
 250         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 251         if (body == NULL)
 252                 RETURN(-EFAULT);
 253
 254         rc = tgt_validate_obdo(tsi, &body->oa);
 255         if (rc)
 256                 RETURN(rc);
 257
 258         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 259         if (IS_ERR(nodemap))
 260                 RETURN(PTR_ERR(nodemap));
 261
 262         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_uid);
 265         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 266                                         NODEMAP_CLIENT_TO_FS,
 267                                         body->oa.o_gid);
 268         nodemap_putref(nodemap);
 269
 270         tsi->tsi_ost_body = body;
 271         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 272
 273         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 274                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 275                 if (rc < 0)
 276                         RETURN(rc);
 277         }
 278
 279         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 280                 if (flags & HAS_BODY) {
 281                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 282                                tgt_name(tsi->tsi_tgt));
 283                         RETURN(-EPROTO);
 284                 } else {
 285                         RETURN(0);
 286                 }
 287         }
 288
 289         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 290
 291         /*
 292          * OST doesn't get object in advance for further use to prevent
 293          * situations with nested object_find which is potential deadlock.
 294          */
 295         tsi->tsi_corpus = NULL;
 296         RETURN(rc);
 297 }
 298
 299 /*
 300  * Do necessary preprocessing according to handler ->th_flags.
 301  */
 302 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 303                                   struct tgt_handler *h,
 304                                   struct ptlrpc_request *req)
 305 {
 306         struct req_capsule      *pill = tsi->tsi_pill;
 307         __u32                    flags = h->th_flags;
 308         int                      rc = 0;
 309
 310         ENTRY;
 311
 312         if (tsi->tsi_preprocessed)
 313                 RETURN(0);
 314
 315         LASSERT(h->th_act != NULL);
 316         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 317         LASSERT(current->journal_info == NULL);
 318
 319         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 320                      h->th_fmt != NULL));
 321         if (h->th_fmt != NULL) {
 322                 req_capsule_set(pill, h->th_fmt);
 323                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT) &&
 324                     req_capsule_field_present(pill, &RMF_MDT_BODY,
 325                                               RCL_CLIENT)) {
 326                         rc = tgt_mdt_body_unpack(tsi, flags);
 327                         if (rc < 0)
 328                                 RETURN(rc);
 329                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 330                                                  RCL_CLIENT) &&
 331                            req_capsule_field_present(pill, &RMF_OST_BODY,
 332                                                  RCL_CLIENT)) {
 333                         rc = tgt_ost_body_unpack(tsi, flags);
 334                         if (rc < 0)
 335                                 RETURN(rc);
 336                 }
 337         }
 338
 339         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 340                 RETURN(-EROFS);
 341
 342         if (flags & HAS_KEY) {
 343                 struct ldlm_request *dlm_req;
 344
 345                 LASSERT(h->th_fmt != NULL);
 346
 347                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 348                 if (dlm_req != NULL) {
 349                         union ldlm_wire_policy_data *policy =
 350                                         &dlm_req->lock_desc.l_policy_data;
 351
 352                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 353                                      LDLM_IBITS &&
 354                                      (policy->l_inodebits.bits |
 355                                       policy->l_inodebits.try_bits) == 0)) {
 356                                 /*
 357                                  * Lock without inodebits makes no sense and
 358                                  * will oops later in ldlm. If client miss to
 359                                  * set such bits, do not trigger ASSERTION.
 360                                  *
 361                                  * For liblustre flock case, it maybe zero.
 362                                  */
 363                                 rc = -EPROTO;
 364                         } else {
 365                                 tsi->tsi_dlm_req = dlm_req;
 366                         }
 367                 } else {
 368                         rc = -EFAULT;
 369                 }
 370         }
 371         tsi->tsi_preprocessed = 1;
 372         RETURN(rc);
 373 }
 374
 375 /*
 376  * Invoke handler for this request opc. Also do necessary preprocessing
 377  * (according to handler ->th_flags), and post-processing (setting of
 378  * ->last_{xid,committed}).
 379  */
 380 static int tgt_handle_request0(struct tgt_session_info *tsi,
 381                                struct tgt_handler *h,
 382                                struct ptlrpc_request *req)
 383 {
 384         int      serious = 0;
 385         int      rc;
 386         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 387
 388         ENTRY;
 389
 390
 391         /* When dealing with sec context requests, no export is associated yet,
 392          * because these requests are sent before *_CONNECT requests.
 393          * A NULL req->rq_export means the normal *_common_slice handlers will
 394          * not be called, because there is no reference to the target.
 395          * So deal with them by hand and jump directly to target_send_reply().
 396          */
 397         switch (opc) {
 398         case SEC_CTX_INIT:
 399         case SEC_CTX_INIT_CONT:
 400         case SEC_CTX_FINI:
 401                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 402                 GOTO(out, rc = 0);
 403         }
 404
 405         /*
 406          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 407          * to put same checks into handlers like mdt_close(), mdt_reint(),
 408          * etc., without talking to mdt authors first. Checking same thing
 409          * there again is useless and returning 0 error without packing reply
 410          * is buggy! Handlers either pack reply or return error.
 411          *
 412          * We return 0 here and do not send any reply in order to emulate
 413          * network failure. Do not send any reply in case any of NET related
 414          * fail_id has occured.
 415          */
 416         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 417                 RETURN(0);
 418         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 419                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 420                 RETURN(0);
 421
 422         /* drop OUT_UPDATE rpc */
 423         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == OUT_UPDATE &&
 424                      OBD_FAIL_CHECK(OBD_FAIL_OUT_UPDATE_DROP)))
 425                 RETURN(0);
 426
 427         rc = tgt_request_preprocess(tsi, h, req);
 428         /* pack reply if reply format is fixed */
 429         if (rc == 0 && h->th_flags & HAS_REPLY) {
 430                 /* Pack reply */
 431                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 432                                           RCL_SERVER))
 433                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 434                                              RCL_SERVER,
 435                                              tsi->tsi_mdt_body->mbo_eadatasize);
 436                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 437                                           RCL_SERVER))
 438                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 439                                              RCL_SERVER, 0);
 440                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 441                         req_capsule_set_size(tsi->tsi_pill,
 442                                              &RMF_ACL, RCL_SERVER,
 443                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 444
 445                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 446                                           RCL_SERVER)) {
 447                         struct niobuf_remote *remote_nb =
 448                                 req_capsule_client_get(tsi->tsi_pill,
 449                                                        &RMF_NIOBUF_REMOTE);
 450                         struct ost_body *body = tsi->tsi_ost_body;
 451
 452                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 453                                          RCL_SERVER,
 454                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 455                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 456                                          remote_nb[0].rnb_len : 0);
 457                 }
 458
 459                 rc = req_capsule_server_pack(tsi->tsi_pill);
 460         }
 461
 462         if (likely(rc == 0)) {
 463                 /*
 464                  * Process request, there can be two types of rc:
 465                  * 1) errors with msg unpack/pack, other failures outside the
 466                  * operation itself. This is counted as serious errors;
 467                  * 2) errors during fs operation, should be placed in rq_status
 468                  * only
 469                  */
 470                 rc = h->th_act(tsi);
 471                 if (!is_serious(rc) &&
 472                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 473                         DEBUG_REQ(D_ERROR, req,
 474                                   "%s: %s handler did not pack reply but returned no error",
 475                                   tgt_name(tsi->tsi_tgt), h->th_name);
 476                         LBUG();
 477                 }
 478                 serious = is_serious(rc);
 479                 rc = clear_serious(rc);
 480         } else {
 481                 serious = 1;
 482         }
 483
 484         req->rq_status = rc;
 485
 486         /*
 487          * ELDLM_* codes which > 0 should be in rq_status only as well as
 488          * all non-serious errors.
 489          */
 490         if (rc > 0 || !serious)
 491                 rc = 0;
 492
 493         LASSERT(current->journal_info == NULL);
 494
 495         if (likely(rc == 0 && req->rq_export))
 496                 target_committed_to_req(req);
 497
 498 out:
 499         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 500         RETURN(0);
 501 }
 502
 503 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 504                                        struct obd_device *obd, int *process)
 505 {
 506         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 507         case MDS_DISCONNECT:
 508         case OST_DISCONNECT:
 509         case OBD_IDX_READ:
 510                 *process = 1;
 511                 RETURN(0);
 512         case MDS_CLOSE:
 513         case MDS_SYNC: /* used in unmounting */
 514         case OBD_PING:
 515         case MDS_REINT:
 516         case OUT_UPDATE:
 517         case SEQ_QUERY:
 518         case FLD_QUERY:
 519         case FLD_READ:
 520         case LDLM_ENQUEUE:
 521         case OST_CREATE:
 522         case OST_DESTROY:
 523         case OST_PUNCH:
 524         case OST_SETATTR:
 525         case OST_SYNC:
 526         case OST_WRITE:
 527         case MDS_HSM_PROGRESS:
 528         case MDS_HSM_STATE_SET:
 529         case MDS_HSM_REQUEST:
 530         case OST_FALLOCATE:
 531                 *process = target_queue_recovery_request(req, obd);
 532                 RETURN(0);
 533
 534         default:
 535                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 536                 *process = -EAGAIN;
 537                 RETURN(0);
 538         }
 539 }
 540
 541 /*
 542  * Handle recovery. Return:
 543  *        +1: continue request processing;
 544  *       -ve: abort immediately with the given error code;
 545  *         0: send reply with error code in req->rq_status;
 546  */
 547 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 548 {
 549         ENTRY;
 550
 551         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 552         case MDS_CONNECT:
 553         case OST_CONNECT:
 554         case MGS_CONNECT:
 555         case SEC_CTX_INIT:
 556         case SEC_CTX_INIT_CONT:
 557         case SEC_CTX_FINI:
 558                 RETURN(+1);
 559         }
 560
 561         if (!req->rq_export->exp_obd->obd_replayable)
 562                 RETURN(+1);
 563
 564         /* sanity check: if the xid matches, the request must be marked as a
 565          * resent or replayed */
 566         if (req_can_reconstruct(req, NULL) == 1) {
 567                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 568                       (MSG_RESENT | MSG_REPLAY))) {
 569                         DEBUG_REQ(D_WARNING, req,
 570                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 571                                   req->rq_xid,
 572                                   lustre_msg_get_flags(req->rq_reqmsg));
 573                         req->rq_status = -ENOTCONN;
 574                         RETURN(-ENOTCONN);
 575                 }
 576         }
 577         /* else: note the opposite is not always true; a RESENT req after a
 578          * failover will usually not match the last_xid, since it was likely
 579          * never committed. A REPLAYed request will almost never match the
 580          * last xid, however it could for a committed, but still retained,
 581          * open. */
 582
 583         /* Check for aborted recovery... */
 584         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 585                 int rc;
 586                 int should_process;
 587
 588                 DEBUG_REQ(D_INFO, req, "Got new replay");
 589                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 590                                                  &should_process);
 591                 if (rc != 0 || !should_process)
 592                         RETURN(rc);
 593                 else if (should_process < 0) {
 594                         req->rq_status = should_process;
 595                         rc = ptlrpc_error(req);
 596                         RETURN(rc);
 597                 }
 598         }
 599         RETURN(+1);
 600 }
 601
 602 /* Initial check for request, it is validation mostly */
 603 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 604 {
 605         struct tgt_handler      *h;
 606         struct tgt_opc_slice    *s;
 607         struct lu_target        *tgt;
 608         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 609
 610         ENTRY;
 611
 612         tgt = class_exp2tgt(req->rq_export);
 613         if (unlikely(tgt == NULL)) {
 614                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 615                           class_exp2obd(req->rq_export)->obd_name);
 616                 RETURN(ERR_PTR(-EINVAL));
 617         }
 618
 619         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 620                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 621                         break;
 622
 623         /* opcode was not found in slice */
 624         if (unlikely(s->tos_hs == NULL)) {
 625                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 626                        opc);
 627                 RETURN(ERR_PTR(-ENOTSUPP));
 628         }
 629
 630         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 631         h = s->tos_hs + (opc - s->tos_opc_start);
 632         if (unlikely(h->th_opc == 0)) {
 633                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 634                 RETURN(ERR_PTR(-ENOTSUPP));
 635         }
 636
 637         RETURN(h);
 638 }
 639
 640 static int process_req_last_xid(struct ptlrpc_request *req)
 641 {
 642         __u64   last_xid;
 643         int rc = 0;
 644         struct obd_export *exp = req->rq_export;
 645         struct tg_export_data *ted = &exp->exp_target_data;
 646         bool need_lock = tgt_is_multimodrpcs_client(exp);
 647         ENTRY;
 648
 649         if (need_lock)
 650                 mutex_lock(&ted->ted_lcd_lock);
 651         /* check request's xid is consistent with export's last_xid */
 652         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 653         if (last_xid > exp->exp_last_xid)
 654                 exp->exp_last_xid = last_xid;
 655
 656         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 657                 /* Some request is allowed to be sent during replay,
 658                  * such as OUT update requests, FLD requests, so it
 659                  * is possible that replay requests has smaller XID
 660                  * than the exp_last_xid.
 661                  *
 662                  * Some non-replay requests may have smaller XID as
 663                  * well:
 664                  *
 665                  * - Client send a no_resend RPC, like statfs;
 666                  * - The RPC timedout (or some other error) on client,
 667                  *   then it's removed from the unreplied list;
 668                  * - Client send some other request to bump the
 669                  *   exp_last_xid on server;
 670                  * - The former RPC got chance to be processed;
 671                  */
 672                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 673                         rc = -EPROTO;
 674
 675                 DEBUG_REQ(D_WARNING, req,
 676                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 677                           req->rq_xid, exp->exp_last_xid, rc);
 678                 if (rc)
 679                         GOTO(out, rc);
 680         }
 681
 682         /* The "last_xid" is the minimum xid among unreplied requests,
 683          * if the request is from the previous connection, its xid can
 684          * still be larger than "exp_last_xid", then the above check of
 685          * xid is not enough to determine whether the request is delayed.
 686          *
 687          * For example, if some replay request was delayed and caused
 688          * timeout at client and the replay is restarted, the delayed
 689          * replay request will have the larger xid than "exp_last_xid"
 690          */
 691         if (req->rq_export->exp_conn_cnt >
 692             lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 693                 CDEBUG(D_RPCTRACE,
 694                        "Dropping request %llu from an old epoch %u/%u\n",
 695                        req->rq_xid,
 696                        lustre_msg_get_conn_cnt(req->rq_reqmsg),
 697                        req->rq_export->exp_conn_cnt);
 698                 req->rq_no_reply = 1;
 699                 GOTO(out, rc = -ESTALE);
 700         }
 701
 702         /* try to release in-memory reply data */
 703         if (tgt_is_multimodrpcs_client(exp)) {
 704                 tgt_handle_received_xid(exp, last_xid);
 705                 rc = tgt_handle_tag(req);
 706         }
 707
 708 out:
 709         if (need_lock)
 710                 mutex_unlock(&ted->ted_lcd_lock);
 711
 712         RETURN(rc);
 713 }
 714
 715 int tgt_request_handle(struct ptlrpc_request *req)
 716 {
 717         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 718
 719         struct lustre_msg       *msg = req->rq_reqmsg;
 720         struct tgt_handler      *h;
 721         struct lu_target        *tgt;
 722         int                      request_fail_id = 0;
 723         __u32                    opc = lustre_msg_get_opc(msg);
 724         struct obd_device       *obd;
 725         int                      rc;
 726         bool                     is_connect = false;
 727         ENTRY;
 728
 729         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 730                 if (cfs_fail_val == 0 &&
 731                     lustre_msg_get_opc(msg) != OBD_PING &&
 732                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 733                         cfs_fail_val = 1;
 734                         cfs_race_state = 0;
 735                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 736                 }
 737         }
 738
 739         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 740         tsi->tsi_pill = &req->rq_pill;
 741         tsi->tsi_env = req->rq_svc_thread->t_env;
 742
 743         /* if request has export then get handlers slice from corresponding
 744          * target, otherwise that should be connect operation */
 745         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 746             opc == MGS_CONNECT) {
 747                 is_connect = true;
 748                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 749                 rc = target_handle_connect(req);
 750                 if (rc != 0) {
 751                         rc = ptlrpc_error(req);
 752                         GOTO(out, rc);
 753                 }
 754                 /* recovery-small test 18c asks to drop connect reply */
 755                 if (unlikely(opc == OST_CONNECT &&
 756                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 757                         GOTO(out, rc = 0);
 758         }
 759
 760         if (unlikely(!class_connected_export(req->rq_export))) {
 761                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 762                     opc == SEC_CTX_FINI) {
 763                         /* sec context initialization has to be handled
 764                          * by hand in tgt_handle_request0() */
 765                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 766                         h = NULL;
 767                         GOTO(handle_recov, rc = 0);
 768                 }
 769                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 770                        opc, libcfs_id2str(req->rq_peer));
 771                 req->rq_status = -ENOTCONN;
 772                 rc = ptlrpc_error(req);
 773                 GOTO(out, rc);
 774         }
 775
 776         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 777         tsi->tsi_exp = req->rq_export;
 778         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 779                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 780         else
 781                 tsi->tsi_jobid = NULL;
 782
 783         if (tgt == NULL) {
 784                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 785                           class_exp2obd(req->rq_export)->obd_name);
 786                 req->rq_status = -EINVAL;
 787                 rc = ptlrpc_error(req);
 788                 GOTO(out, rc);
 789         }
 790
 791         /* Skip last_xid processing for the recovery thread, otherwise, the
 792          * last_xid on same request could be processed twice: first time when
 793          * processing the incoming request, second time when the request is
 794          * being processed by recovery thread. */
 795         obd = class_exp2obd(req->rq_export);
 796         if (is_connect) {
 797                 /* reset the exp_last_xid on each connection. */
 798                 req->rq_export->exp_last_xid = 0;
 799         } else if (obd->obd_recovery_data.trd_processing_task !=
 800                    current->pid) {
 801                 rc = process_req_last_xid(req);
 802                 if (rc) {
 803                         req->rq_status = rc;
 804                         rc = ptlrpc_error(req);
 805                         GOTO(out, rc);
 806                 }
 807         }
 808
 809         request_fail_id = tgt->lut_request_fail_id;
 810         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 811
 812         h = tgt_handler_find_check(req);
 813         if (IS_ERR(h)) {
 814                 req->rq_status = PTR_ERR(h);
 815                 rc = ptlrpc_error(req);
 816                 GOTO(out, rc);
 817         }
 818
 819         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 820                  h->th_opc, opc);
 821
 822         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 823              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 824                 GOTO(out, rc = 0);
 825
 826         rc = lustre_msg_check_version(msg, h->th_version);
 827         if (unlikely(rc)) {
 828                 DEBUG_REQ(D_ERROR, req,
 829                           "%s: drop malformed request version=%08x expect=%08x",
 830                           tgt_name(tgt), lustre_msg_get_version(msg),
 831                           h->th_version);
 832                 req->rq_status = -EINVAL;
 833                 rc = ptlrpc_error(req);
 834                 GOTO(out, rc);
 835         }
 836
 837 handle_recov:
 838         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 839         if (likely(rc == 1)) {
 840                 rc = tgt_handle_request0(tsi, h, req);
 841                 if (rc)
 842                         GOTO(out, rc);
 843         }
 844         EXIT;
 845 out:
 846         req_capsule_fini(tsi->tsi_pill);
 847         if (tsi->tsi_corpus != NULL) {
 848                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 849                 tsi->tsi_corpus = NULL;
 850         }
 851         return rc;
 852 }
 853 EXPORT_SYMBOL(tgt_request_handle);
 854
 855 /** Assign high priority operations to the request if needed. */
 856 int tgt_hpreq_handler(struct ptlrpc_request *req)
 857 {
 858         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 859         struct tgt_handler      *h;
 860         int                      rc;
 861
 862         ENTRY;
 863
 864         if (req->rq_export == NULL)
 865                 RETURN(0);
 866
 867         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 868         tsi->tsi_pill = &req->rq_pill;
 869         tsi->tsi_env = req->rq_svc_thread->t_env;
 870         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 871         tsi->tsi_exp = req->rq_export;
 872
 873         h = tgt_handler_find_check(req);
 874         if (IS_ERR(h)) {
 875                 rc = PTR_ERR(h);
 876                 RETURN(rc);
 877         }
 878
 879         rc = tgt_request_preprocess(tsi, h, req);
 880         if (unlikely(rc != 0))
 881                 RETURN(rc);
 882
 883         if (h->th_hp != NULL)
 884                 h->th_hp(tsi);
 885         RETURN(0);
 886 }
 887 EXPORT_SYMBOL(tgt_hpreq_handler);
 888
 889 void tgt_counter_incr(struct obd_export *exp, int opcode)
 890 {
 891         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 892         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 893                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 894 }
 895 EXPORT_SYMBOL(tgt_counter_incr);
 896
 897 /*
 898  * Unified target generic handlers.
 899  */
 900
 901 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 902 {
 903         struct lu_target *tgt = class_exp2tgt(exp);
 904         struct sptlrpc_flavor flvr;
 905         int rc = 0;
 906
 907         LASSERT(tgt);
 908         LASSERT(tgt->lut_obd);
 909         LASSERT(tgt->lut_slice);
 910
 911         /* always allow ECHO client */
 912         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 913                             LUSTRE_ECHO_NAME) == 0)) {
 914                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 915                 return 0;
 916         }
 917
 918         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 919                 read_lock(&tgt->lut_sptlrpc_lock);
 920                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 921                                              req->rq_sp_from,
 922                                              req->rq_peer.nid,
 923                                              &flvr);
 924                 read_unlock(&tgt->lut_sptlrpc_lock);
 925
 926                 spin_lock(&exp->exp_lock);
 927                 exp->exp_sp_peer = req->rq_sp_from;
 928                 exp->exp_flvr = flvr;
 929
 930                 /* when on mgs, if no restriction is set, or if the client
 931                  * NID is on the local node, allow any flavor
 932                  */
 933                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 934                            LUSTRE_MGS_NAME) == 0) &&
 935                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 936                       LNetIsPeerLocal(exp->exp_connection->c_peer.nid)))
 937                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 938
 939                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 940                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 941                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 942                                "expect %x\n", tgt_name(tgt),
 943                                req->rq_flvr.sf_rpc,
 944                                libcfs_nid2str(req->rq_peer.nid),
 945                                exp->exp_flvr.sf_rpc);
 946                         rc = -EACCES;
 947                 }
 948                 spin_unlock(&exp->exp_lock);
 949         } else {
 950                 if (exp->exp_sp_peer != req->rq_sp_from) {
 951                         CERROR("%s: RPC source %s doesn't match %s\n",
 952                                tgt_name(tgt),
 953                                sptlrpc_part2name(req->rq_sp_from),
 954                                sptlrpc_part2name(exp->exp_sp_peer));
 955                         rc = -EACCES;
 956                 } else {
 957                         rc = sptlrpc_target_export_check(exp, req);
 958                 }
 959         }
 960
 961         return rc;
 962 }
 963
 964 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 965 {
 966         struct sptlrpc_rule_set  tmp_rset;
 967         int                      rc;
 968
 969         if (unlikely(tgt == NULL)) {
 970                 CERROR("No target passed");
 971                 return -EINVAL;
 972         }
 973
 974         sptlrpc_rule_set_init(&tmp_rset);
 975         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 976         if (rc) {
 977                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 978                        tgt_name(tgt), rc);
 979                 return rc;
 980         }
 981
 982         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 983
 984         write_lock(&tgt->lut_sptlrpc_lock);
 985         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 986         tgt->lut_sptlrpc_rset = tmp_rset;
 987         write_unlock(&tgt->lut_sptlrpc_lock);
 988
 989         return 0;
 990 }
 991 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 992
 993 int tgt_connect(struct tgt_session_info *tsi)
 994 {
 995         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 996         struct obd_connect_data *reply;
 997         int                      rc;
 998
 999         ENTRY;
1000
1001         /* XXX: better to call this check right after getting new export but
1002          * before last_rcvd slot allocation to avoid server load upon insecure
1003          * connects. This is to be fixed after unifiyng all targets.
1004          */
1005         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1006         if (rc)
1007                 GOTO(out, rc);
1008
1009         /* To avoid exposing partially initialized connection flags, changes up
1010          * to this point have been staged in reply->ocd_connect_flags. Now that
1011          * connection handling has completed successfully, atomically update
1012          * the connect flags in the shared export data structure. LU-1623 */
1013         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1014         spin_lock(&tsi->tsi_exp->exp_lock);
1015         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1016         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1017                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1018                         reply->ocd_connect_flags2;
1019         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1020         spin_unlock(&tsi->tsi_exp->exp_lock);
1021
1022         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1023                    LUSTRE_MDT_NAME) == 0) {
1024                 rc = req_check_sepol(tsi->tsi_pill);
1025                 if (rc)
1026                         GOTO(out, rc);
1027
1028                 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1029                     reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT &&
1030                     tsi->tsi_pill->rc_req->rq_export) {
1031                         bool forbid_encrypt = true;
1032                         struct lu_nodemap *nm =
1033                          nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1034
1035                         if (!nm) {
1036                                 /* nodemap_get_from_exp returns NULL in case
1037                                  * nodemap is not active, so we do not forbid
1038                                  */
1039                                 forbid_encrypt = false;
1040                         } else if (!IS_ERR(nm)) {
1041                                 forbid_encrypt = nm->nmf_forbid_encryption;
1042                                 nodemap_putref(nm);
1043                         }
1044
1045                         if (forbid_encrypt)
1046                                 GOTO(out, rc = -EACCES);
1047                 }
1048         }
1049
1050         RETURN(0);
1051 out:
1052         obd_disconnect(class_export_get(tsi->tsi_exp));
1053         return rc;
1054 }
1055 EXPORT_SYMBOL(tgt_connect);
1056
1057 int tgt_disconnect(struct tgt_session_info *tsi)
1058 {
1059         int rc;
1060
1061         ENTRY;
1062
1063         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1064
1065         rc = target_handle_disconnect(tgt_ses_req(tsi));
1066         if (rc)
1067                 RETURN(err_serious(rc));
1068
1069         RETURN(rc);
1070 }
1071 EXPORT_SYMBOL(tgt_disconnect);
1072
1073 /*
1074  * Unified target OBD handlers
1075  */
1076 int tgt_obd_ping(struct tgt_session_info *tsi)
1077 {
1078         int rc;
1079
1080         ENTRY;
1081
1082         /* The target-specific part of OBD_PING request handling.
1083          * It controls Filter Modification Data (FMD) expiration each time
1084          * PING is received.
1085          *
1086          * Valid only for replayable targets, e.g. MDT and OFD
1087          */
1088         if (tsi->tsi_exp->exp_obd->obd_replayable)
1089                 tgt_fmd_expire(tsi->tsi_exp);
1090
1091         rc = req_capsule_server_pack(tsi->tsi_pill);
1092         if (rc)
1093                 RETURN(err_serious(rc));
1094
1095         RETURN(rc);
1096 }
1097 EXPORT_SYMBOL(tgt_obd_ping);
1098
1099 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1100 {
1101         return err_serious(-EOPNOTSUPP);
1102 }
1103
1104 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1105 {
1106         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1107         struct obd_export       *exp = req->rq_export;
1108         struct ptlrpc_bulk_desc *desc;
1109         int                      i;
1110         int                      rc;
1111         int                      pages = 0;
1112
1113         ENTRY;
1114
1115         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1116                 unsigned int offset;
1117
1118                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1119                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1120                                       PAGE_SIZE);
1121         }
1122
1123         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1124                                   PTLRPC_BULK_PUT_SOURCE,
1125                                     MDS_BULK_PORTAL,
1126                                     &ptlrpc_bulk_kiov_nopin_ops);
1127         if (desc == NULL)
1128                 RETURN(-ENOMEM);
1129
1130         for (i = 0; i < rdbuf->rb_nbufs; i++)
1131                 desc->bd_frag_ops->add_iov_frag(desc,
1132                                         rdbuf->rb_bufs[i].lb_buf,
1133                                         rdbuf->rb_bufs[i].lb_len);
1134
1135         rc = target_bulk_io(exp, desc);
1136         ptlrpc_free_bulk(desc);
1137         RETURN(rc);
1138 }
1139 EXPORT_SYMBOL(tgt_send_buffer);
1140
1141 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1142 {
1143         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1144         struct obd_export       *exp = req->rq_export;
1145         struct ptlrpc_bulk_desc *desc;
1146         int                      tmpcount;
1147         int                      tmpsize;
1148         int                      i;
1149         int                      rc;
1150
1151         ENTRY;
1152
1153         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1154                                     PTLRPC_BULK_PUT_SOURCE,
1155                                     MDS_BULK_PORTAL,
1156                                     &ptlrpc_bulk_kiov_pin_ops);
1157         if (desc == NULL)
1158                 RETURN(-ENOMEM);
1159
1160         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1161                 /* old client requires reply size in it's PAGE_SIZE,
1162                  * which is rdpg->rp_count */
1163                 nob = rdpg->rp_count;
1164
1165         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1166              i++, tmpcount -= tmpsize) {
1167                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1168                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1169                                                  tmpsize);
1170         }
1171
1172         LASSERT(desc->bd_nob == nob);
1173         rc = target_bulk_io(exp, desc);
1174         ptlrpc_free_bulk(desc);
1175         RETURN(rc);
1176 }
1177 EXPORT_SYMBOL(tgt_sendpage);
1178
1179 /*
1180  * OBD_IDX_READ handler
1181  */
1182 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1183 {
1184         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1185         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1186         struct idx_info         *req_ii, *rep_ii;
1187         int                      rc, i;
1188
1189         ENTRY;
1190
1191         memset(rdpg, 0, sizeof(*rdpg));
1192         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1193
1194         /* extract idx_info buffer from request & reply */
1195         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1196         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1197                 RETURN(err_serious(-EPROTO));
1198
1199         rc = req_capsule_server_pack(tsi->tsi_pill);
1200         if (rc)
1201                 RETURN(err_serious(rc));
1202
1203         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1204         if (rep_ii == NULL)
1205                 RETURN(err_serious(-EFAULT));
1206         rep_ii->ii_magic = IDX_INFO_MAGIC;
1207
1208         /* extract hash to start with */
1209         rdpg->rp_hash = req_ii->ii_hash_start;
1210
1211         /* extract requested attributes */
1212         rdpg->rp_attrs = req_ii->ii_attrs;
1213
1214         /* check that fid packed in request is valid and supported */
1215         if (!fid_is_sane(&req_ii->ii_fid))
1216                 RETURN(-EINVAL);
1217         rep_ii->ii_fid = req_ii->ii_fid;
1218
1219         /* copy flags */
1220         rep_ii->ii_flags = req_ii->ii_flags;
1221
1222         /* compute number of pages to allocate, ii_count is the number of 4KB
1223          * containers */
1224         if (req_ii->ii_count <= 0)
1225                 GOTO(out, rc = -EFAULT);
1226         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1227                                exp_max_brw_size(tsi->tsi_exp));
1228         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1229
1230         /* allocate pages to store the containers */
1231         OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1232         if (rdpg->rp_pages == NULL)
1233                 GOTO(out, rc = -ENOMEM);
1234         for (i = 0; i < rdpg->rp_npages; i++) {
1235                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1236                 if (rdpg->rp_pages[i] == NULL)
1237                         GOTO(out, rc = -ENOMEM);
1238         }
1239
1240         /* populate pages with key/record pairs */
1241         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1242         if (rc < 0)
1243                 GOTO(out, rc);
1244
1245         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1246                  "asked %d > %d\n", rc, rdpg->rp_count);
1247
1248         /* send pages to client */
1249         rc = tgt_sendpage(tsi, rdpg, rc);
1250         if (rc)
1251                 GOTO(out, rc);
1252         EXIT;
1253 out:
1254         if (rdpg->rp_pages) {
1255                 for (i = 0; i < rdpg->rp_npages; i++)
1256                         if (rdpg->rp_pages[i])
1257                                 __free_page(rdpg->rp_pages[i]);
1258                 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1259         }
1260         return rc;
1261 }
1262
1263 struct tgt_handler tgt_obd_handlers[] = {
1264 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1265 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1266 };
1267 EXPORT_SYMBOL(tgt_obd_handlers);
1268
1269 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1270              struct dt_object *obj, __u64 start, __u64 end)
1271 {
1272         int rc = 0;
1273
1274         ENTRY;
1275
1276         /* if no objid is specified, it means "sync whole filesystem" */
1277         if (obj == NULL) {
1278                 rc = dt_sync(env, tgt->lut_bottom);
1279         } else if (dt_version_get(env, obj) >
1280                    tgt->lut_obd->obd_last_committed) {
1281                 rc = dt_object_sync(env, obj, start, end);
1282         }
1283         atomic_inc(&tgt->lut_sync_count);
1284
1285         RETURN(rc);
1286 }
1287 EXPORT_SYMBOL(tgt_sync);
1288 /*
1289  * Unified target DLM handlers.
1290  */
1291
1292 /**
1293  * Unified target BAST
1294  *
1295  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1296  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1297  * but if it's cross-MDT lock, because remote object version is not set, a
1298  * filesystem sync is needed.
1299  *
1300  * \param lock server side lock
1301  * \param desc lock desc
1302  * \param data ldlm_cb_set_arg
1303  * \param flag  indicates whether this cancelling or blocking callback
1304  * \retval      0 on success
1305  * \retval      negative number on error
1306  */
1307 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1308                             void *data, int flag)
1309 {
1310         struct lu_env            env;
1311         struct lu_target        *tgt;
1312         struct dt_object        *obj = NULL;
1313         struct lu_fid            fid;
1314         int                      rc = 0;
1315
1316         ENTRY;
1317
1318         tgt = class_exp2tgt(lock->l_export);
1319
1320         if (unlikely(tgt == NULL)) {
1321                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1322                        class_exp2obd(lock->l_export)->obd_name);
1323                 RETURN(-EINVAL);
1324         }
1325
1326         if (flag == LDLM_CB_CANCELING &&
1327             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1328             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1329              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1330               ldlm_is_cbpending(lock))) &&
1331             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1332              lock->l_resource->lr_type == LDLM_EXTENT)) {
1333                 __u64 start = 0;
1334                 __u64 end = OBD_OBJECT_EOF;
1335
1336                 rc = lu_env_init(&env, LCT_DT_THREAD);
1337                 if (unlikely(rc != 0))
1338                         GOTO(err, rc);
1339
1340                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1341                                    tgt->lut_lsd.lsd_osd_index);
1342
1343                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1344                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1345                         if (IS_ERR(obj))
1346                                 GOTO(err_env, rc = PTR_ERR(obj));
1347
1348                         if (!dt_object_exists(obj))
1349                                 GOTO(err_put, rc = -ENOENT);
1350
1351                         start = lock->l_policy_data.l_extent.start;
1352                         end = lock->l_policy_data.l_extent.end;
1353                 }
1354
1355                 rc = tgt_sync(&env, tgt, obj, start, end);
1356                 if (rc < 0) {
1357                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1358                                "cancel: rc = %d\n",
1359                                tgt_name(tgt), PFID(&fid),
1360                                lock->l_policy_data.l_extent.start,
1361                                lock->l_policy_data.l_extent.end, rc);
1362                 }
1363 err_put:
1364                 if (obj != NULL)
1365                         dt_object_put(&env, obj);
1366 err_env:
1367                 lu_env_fini(&env);
1368         }
1369 err:
1370         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1371         RETURN(rc);
1372 }
1373
1374 static struct ldlm_callback_suite tgt_dlm_cbs = {
1375         .lcs_completion = ldlm_server_completion_ast,
1376         .lcs_blocking   = tgt_blocking_ast,
1377         .lcs_glimpse    = ldlm_server_glimpse_ast
1378 };
1379
1380 int tgt_enqueue(struct tgt_session_info *tsi)
1381 {
1382         struct ptlrpc_request *req = tgt_ses_req(tsi);
1383         int rc;
1384
1385         ENTRY;
1386         /*
1387          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1388          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1389          */
1390         LASSERT(tsi->tsi_dlm_req != NULL);
1391         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1392                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1393         if (rc)
1394                 RETURN(err_serious(rc));
1395
1396         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1397         case LUT_FAIL_MDT:
1398                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1399                 break;
1400         case LUT_FAIL_OST:
1401                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1402                 break;
1403         case LUT_FAIL_MGT:
1404                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1405                 break;
1406         default:
1407                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1408                 break;
1409         }
1410         RETURN(req->rq_status);
1411 }
1412 EXPORT_SYMBOL(tgt_enqueue);
1413
1414 int tgt_convert(struct tgt_session_info *tsi)
1415 {
1416         struct ptlrpc_request *req = tgt_ses_req(tsi);
1417         int rc;
1418
1419         ENTRY;
1420         LASSERT(tsi->tsi_dlm_req);
1421         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1422         if (rc)
1423                 RETURN(err_serious(rc));
1424
1425         RETURN(req->rq_status);
1426 }
1427
1428 int tgt_bl_callback(struct tgt_session_info *tsi)
1429 {
1430         return err_serious(-EOPNOTSUPP);
1431 }
1432
1433 int tgt_cp_callback(struct tgt_session_info *tsi)
1434 {
1435         return err_serious(-EOPNOTSUPP);
1436 }
1437
1438 /* generic LDLM target handler */
1439 struct tgt_handler tgt_dlm_handlers[] = {
1440 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1441 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1442 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1443 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1444 };
1445 EXPORT_SYMBOL(tgt_dlm_handlers);
1446
1447 /*
1448  * Unified target LLOG handlers.
1449  */
1450 int tgt_llog_open(struct tgt_session_info *tsi)
1451 {
1452         int rc;
1453
1454         ENTRY;
1455
1456         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1457
1458         RETURN(rc);
1459 }
1460 EXPORT_SYMBOL(tgt_llog_open);
1461
1462 int tgt_llog_read_header(struct tgt_session_info *tsi)
1463 {
1464         int rc;
1465
1466         ENTRY;
1467
1468         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1469
1470         RETURN(rc);
1471 }
1472 EXPORT_SYMBOL(tgt_llog_read_header);
1473
1474 int tgt_llog_next_block(struct tgt_session_info *tsi)
1475 {
1476         int rc;
1477
1478         ENTRY;
1479
1480         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1481
1482         RETURN(rc);
1483 }
1484 EXPORT_SYMBOL(tgt_llog_next_block);
1485
1486 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1487 {
1488         int rc;
1489
1490         ENTRY;
1491
1492         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1493
1494         RETURN(rc);
1495 }
1496 EXPORT_SYMBOL(tgt_llog_prev_block);
1497
1498 /* generic llog target handler */
1499 struct tgt_handler tgt_llog_handlers[] = {
1500 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1501 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1502 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1503 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1504 };
1505 EXPORT_SYMBOL(tgt_llog_handlers);
1506
1507 /*
1508  * sec context handlers
1509  */
1510 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1511 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1512 {
1513         return 0;
1514 }
1515
1516 struct tgt_handler tgt_sec_ctx_handlers[] = {
1517 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1518 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1519 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1520 };
1521 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1522
1523 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1524                                  struct dt_device *key,
1525                                  struct lfsck_req_local *lrl,
1526                                  struct thandle *th) = NULL;
1527
1528 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1529                                                       struct dt_device *,
1530                                                       struct lfsck_req_local *,
1531                                                       struct thandle *))
1532 {
1533         tgt_lfsck_in_notify_local = notify;
1534 }
1535 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1536
1537 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1538                            struct dt_device *key,
1539                            struct lfsck_request *lr) = NULL;
1540
1541 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1542                                                 struct dt_device *,
1543                                                 struct lfsck_request *))
1544 {
1545         tgt_lfsck_in_notify = notify;
1546 }
1547 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1548
1549 static int (*tgt_lfsck_query)(const struct lu_env *env,
1550                               struct dt_device *key,
1551                               struct lfsck_request *req,
1552                               struct lfsck_reply *rep,
1553                               struct lfsck_query *que) = NULL;
1554
1555 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1556                                            struct dt_device *,
1557                                            struct lfsck_request *,
1558                                            struct lfsck_reply *,
1559                                            struct lfsck_query *))
1560 {
1561         tgt_lfsck_query = query;
1562 }
1563 EXPORT_SYMBOL(tgt_register_lfsck_query);
1564
1565 /* LFSCK request handlers */
1566 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1567 {
1568         const struct lu_env     *env = tsi->tsi_env;
1569         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1570         struct lfsck_request    *lr;
1571         int                      rc;
1572         ENTRY;
1573
1574         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1575         if (lr == NULL)
1576                 RETURN(-EPROTO);
1577
1578         rc = tgt_lfsck_in_notify(env, key, lr);
1579
1580         RETURN(rc);
1581 }
1582
1583 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1584 {
1585         struct lfsck_request    *request;
1586         struct lfsck_reply      *reply;
1587         int                      rc;
1588         ENTRY;
1589
1590         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1591         if (request == NULL)
1592                 RETURN(-EPROTO);
1593
1594         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1595         if (reply == NULL)
1596                 RETURN(-ENOMEM);
1597
1598         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1599                              request, reply, NULL);
1600
1601         RETURN(rc < 0 ? rc : 0);
1602 }
1603
1604 struct tgt_handler tgt_lfsck_handlers[] = {
1605 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1606 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1607 };
1608 EXPORT_SYMBOL(tgt_lfsck_handlers);
1609
1610 /*
1611  * initialize per-thread page pool (bug 5137).
1612  */
1613 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1614 {
1615         struct tgt_thread_big_cache *tbc;
1616
1617         ENTRY;
1618
1619         LASSERT(thread != NULL);
1620         LASSERT(thread->t_data == NULL);
1621
1622         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1623         if (tbc == NULL)
1624                 RETURN(-ENOMEM);
1625         thread->t_data = tbc;
1626         RETURN(0);
1627 }
1628 EXPORT_SYMBOL(tgt_io_thread_init);
1629
1630 /*
1631  * free per-thread pool created by tgt_thread_init().
1632  */
1633 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1634 {
1635         struct tgt_thread_big_cache *tbc;
1636
1637         ENTRY;
1638
1639         LASSERT(thread != NULL);
1640
1641         /*
1642          * be prepared to handle partially-initialized pools (because this is
1643          * called from ost_io_thread_init() for cleanup.
1644          */
1645         tbc = thread->t_data;
1646         if (tbc != NULL) {
1647                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1648                 thread->t_data = NULL;
1649         }
1650         EXIT;
1651 }
1652 EXPORT_SYMBOL(tgt_io_thread_done);
1653
1654 /**
1655  * Helper function for getting Data-on-MDT file server DLM lock
1656  * if asked by client.
1657  */
1658 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1659                       struct lustre_handle *lh, int mode, __u64 *flags)
1660 {
1661         union ldlm_policy_data policy = {
1662                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1663         };
1664         int rc;
1665
1666         ENTRY;
1667
1668         LASSERT(lh != NULL);
1669         LASSERT(ns != NULL);
1670         LASSERT(!lustre_handle_is_used(lh));
1671
1672         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1673                                     flags, ldlm_blocking_ast,
1674                                     ldlm_completion_ast, ldlm_glimpse_ast,
1675                                     NULL, 0, LVB_T_NONE, NULL, lh);
1676
1677         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1678 }
1679 EXPORT_SYMBOL(tgt_mdt_data_lock);
1680
1681 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1682 {
1683         LASSERT(lustre_handle_is_used(lh));
1684         ldlm_lock_decref(lh, mode);
1685 }
1686 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1687
1688 /**
1689  * Helper function for getting server side [start, start+count] DLM lock
1690  * if asked by client.
1691  */
1692 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1693                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1694                     struct lustre_handle *lh, int mode, __u64 *flags)
1695 {
1696         union ldlm_policy_data policy;
1697         int rc;
1698
1699         ENTRY;
1700
1701         LASSERT(lh != NULL);
1702         LASSERT(ns != NULL);
1703         LASSERT(!lustre_handle_is_used(lh));
1704
1705         policy.l_extent.gid = 0;
1706         policy.l_extent.start = start & PAGE_MASK;
1707
1708         /*
1709          * If ->o_blocks is EOF it means "lock till the end of the file".
1710          * Otherwise, it's size of an extent or hole being punched (in bytes).
1711          */
1712         if (end == OBD_OBJECT_EOF || end < start)
1713                 policy.l_extent.end = OBD_OBJECT_EOF;
1714         else
1715                 policy.l_extent.end = end | ~PAGE_MASK;
1716
1717         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1718                                     mode, flags, ldlm_blocking_ast,
1719                                     ldlm_completion_ast, ldlm_glimpse_ast,
1720                                     NULL, 0, LVB_T_NONE, NULL, lh);
1721         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1722 }
1723 EXPORT_SYMBOL(tgt_extent_lock);
1724
1725 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1726 {
1727         LASSERT(lustre_handle_is_used(lh));
1728         ldlm_lock_decref(lh, mode);
1729 }
1730 EXPORT_SYMBOL(tgt_extent_unlock);
1731
1732 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1733                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1734                         struct niobuf_remote *nb, struct lustre_handle *lh,
1735                         enum ldlm_mode mode)
1736 {
1737         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1738         __u64                    flags = 0;
1739         int                      nrbufs = obj->ioo_bufcnt;
1740         int                      i;
1741         int                      rc;
1742
1743         ENTRY;
1744
1745         LASSERT(mode == LCK_PR || mode == LCK_PW);
1746         LASSERT(!lustre_handle_is_used(lh));
1747
1748         if (ns->ns_obd->obd_recovering)
1749                 RETURN(0);
1750
1751         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1752                 RETURN(0);
1753
1754         for (i = 1; i < nrbufs; i++)
1755                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1756                         RETURN(-EFAULT);
1757
1758         /* MDT IO for data-on-mdt */
1759         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1760                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1761         else
1762                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1763                                      nb[nrbufs - 1].rnb_offset +
1764                                      nb[nrbufs - 1].rnb_len - 1,
1765                                      lh, mode, &flags);
1766         RETURN(rc);
1767 }
1768
1769 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1770                            struct niobuf_remote *niob,
1771                            struct lustre_handle *lh, enum ldlm_mode mode)
1772 {
1773         ENTRY;
1774
1775         LASSERT(mode == LCK_PR || mode == LCK_PW);
1776         LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1777                  niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1778                 lustre_handle_is_used(lh));
1779
1780         if (lustre_handle_is_used(lh))
1781                 tgt_extent_unlock(lh, mode);
1782         EXIT;
1783 }
1784
1785 static int tgt_checksum_niobuf(struct lu_target *tgt,
1786                                  struct niobuf_local *local_nb, int npages,
1787                                  int opc, enum cksum_types cksum_type,
1788                                  __u32 *cksum)
1789 {
1790         struct ahash_request           *req;
1791         unsigned int                    bufsize;
1792         int                             i, err;
1793         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1794
1795         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1796         if (IS_ERR(req)) {
1797                 CERROR("%s: unable to initialize checksum hash %s\n",
1798                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1799                 return PTR_ERR(req);
1800         }
1801
1802         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1803         for (i = 0; i < npages; i++) {
1804                 /* corrupt the data before we compute the checksum, to
1805                  * simulate a client->OST data error */
1806                 if (i == 0 && opc == OST_WRITE &&
1807                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1808                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1809                         int len = local_nb[i].lnb_len;
1810                         struct page *np = tgt_page_to_corrupt;
1811
1812                         if (np) {
1813                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1814                                 char *ptr2 = page_address(np);
1815
1816                                 memcpy(ptr2 + off, ptr + off, len);
1817                                 memcpy(ptr2 + off, "bad3", min(4, len));
1818                                 kunmap_atomic(ptr);
1819
1820                                 /* LU-8376 to preserve original index for
1821                                  * display in dump_all_bulk_pages() */
1822                                 np->index = i;
1823
1824                                 cfs_crypto_hash_update_page(req, np, off,
1825                                                             len);
1826                                 continue;
1827                         } else {
1828                                 CERROR("%s: can't alloc page for corruption\n",
1829                                        tgt_name(tgt));
1830                         }
1831                 }
1832                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1833                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1834                                   local_nb[i].lnb_len);
1835
1836                  /* corrupt the data after we compute the checksum, to
1837                  * simulate an OST->client data error */
1838                 if (i == 0 && opc == OST_READ &&
1839                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1840                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1841                         int len = local_nb[i].lnb_len;
1842                         struct page *np = tgt_page_to_corrupt;
1843
1844                         if (np) {
1845                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1846                                 char *ptr2 = page_address(np);
1847
1848                                 memcpy(ptr2 + off, ptr + off, len);
1849                                 memcpy(ptr2 + off, "bad4", min(4, len));
1850                                 kunmap_atomic(ptr);
1851
1852                                 /* LU-8376 to preserve original index for
1853                                  * display in dump_all_bulk_pages() */
1854                                 np->index = i;
1855
1856                                 cfs_crypto_hash_update_page(req, np, off,
1857                                                             len);
1858                                 continue;
1859                         } else {
1860                                 CERROR("%s: can't alloc page for corruption\n",
1861                                        tgt_name(tgt));
1862                         }
1863                 }
1864         }
1865
1866         bufsize = sizeof(*cksum);
1867         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1868
1869         return 0;
1870 }
1871
1872 char dbgcksum_file_name[PATH_MAX];
1873
1874 static void dump_all_bulk_pages(struct obdo *oa, int count,
1875                                 struct niobuf_local *local_nb,
1876                                 __u32 server_cksum, __u32 client_cksum)
1877 {
1878         struct file *filp;
1879         int rc, i;
1880         unsigned int len;
1881         char *buf;
1882
1883         /* will only keep dump of pages on first error for the same range in
1884          * file/fid, not during the resends/retries. */
1885         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1886                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1887                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1888                   libcfs_debug_file_path_arr :
1889                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1890                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1891                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1892                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1893                  local_nb[0].lnb_file_offset,
1894                  local_nb[count-1].lnb_file_offset +
1895                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1896         filp = filp_open(dbgcksum_file_name,
1897                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1898         if (IS_ERR(filp)) {
1899                 rc = PTR_ERR(filp);
1900                 if (rc == -EEXIST)
1901                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1902                                "checksum error: rc = %d\n", dbgcksum_file_name,
1903                                rc);
1904                 else
1905                         CERROR("%s: can't open to dump pages with checksum "
1906                                "error: rc = %d\n", dbgcksum_file_name, rc);
1907                 return;
1908         }
1909
1910         for (i = 0; i < count; i++) {
1911                 len = local_nb[i].lnb_len;
1912                 buf = kmap(local_nb[i].lnb_page);
1913                 while (len != 0) {
1914                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1915                         if (rc < 0) {
1916                                 CERROR("%s: wanted to write %u but got %d "
1917                                        "error\n", dbgcksum_file_name, len, rc);
1918                                 break;
1919                         }
1920                         len -= rc;
1921                         buf += rc;
1922                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1923                                dbgcksum_file_name, rc);
1924                 }
1925                 kunmap(local_nb[i].lnb_page);
1926         }
1927
1928         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1929         if (rc)
1930                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1931         filp_close(filp, NULL);
1932 }
1933
1934 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1935                                struct obd_export *exp, struct obdo *oa,
1936                                const struct lnet_process_id *peer,
1937                                __u32 client_cksum, __u32 server_cksum,
1938                                enum cksum_types server_cksum_type)
1939 {
1940         char *msg;
1941         enum cksum_types cksum_type;
1942         loff_t start, end;
1943
1944         /* unlikely to happen and only if resend does not occur due to cksum
1945          * control failure on Client */
1946         if (unlikely(server_cksum == client_cksum)) {
1947                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1948                        client_cksum);
1949                 return 0;
1950         }
1951
1952         if (exp->exp_obd->obd_checksum_dump)
1953                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1954                                     client_cksum);
1955
1956         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1957                                            oa->o_flags : 0);
1958
1959         if (cksum_type != server_cksum_type)
1960                 msg = "the server may have not used the checksum type specified"
1961                       " in the original request - likely a protocol problem";
1962         else
1963                 msg = "should have changed on the client or in transit";
1964
1965         start = local_nb[0].lnb_file_offset;
1966         end = local_nb[npages-1].lnb_file_offset +
1967                                         local_nb[npages-1].lnb_len - 1;
1968
1969         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1970                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1971                 " %x (type %x), server csum %x (type %x)\n",
1972                 exp->exp_obd->obd_name,
1973                 msg, libcfs_nid2str(peer->nid),
1974                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1975                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1976                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1977                 POSTID(&oa->o_oi),
1978                 start, end, client_cksum, cksum_type, server_cksum,
1979                 server_cksum_type);
1980
1981         return 1;
1982 }
1983
1984 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1985                              unsigned char *buf, int size)
1986 {
1987         int     i, off, len, copied = size;
1988         char    *ptr;
1989
1990         for (i = 0; i < npages; i++) {
1991                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1992                 len = local[i].lnb_len;
1993
1994                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1995                        i, off, len, size);
1996                 if (len > size)
1997                         return -EINVAL;
1998
1999                 ptr = kmap_atomic(local[i].lnb_page);
2000                 memcpy(buf, ptr + off, len);
2001                 kunmap_atomic(ptr);
2002                 buf += len;
2003                 size -= len;
2004         }
2005         return copied - size;
2006 }
2007
2008 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2009                                      struct niobuf_local *local_nb,
2010                                      int npages, int opc,
2011                                      obd_dif_csum_fn *fn,
2012                                      int sector_size,
2013                                      u32 *check_sum)
2014 {
2015         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2016         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2017         const char *obd_name = tgt->lut_obd->obd_name;
2018         struct ahash_request *req;
2019         unsigned int bufsize;
2020         unsigned char *buffer;
2021         struct page *__page;
2022         __u16 *guard_start;
2023         int guard_number;
2024         int used_number = 0;
2025         __u32 cksum;
2026         int rc = 0;
2027         int used;
2028         int i;
2029
2030         __page = alloc_page(GFP_KERNEL);
2031         if (__page == NULL)
2032                 return -ENOMEM;
2033
2034         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2035         if (IS_ERR(req)) {
2036                 CERROR("%s: unable to initialize checksum hash %s\n",
2037                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2038                 return PTR_ERR(req);
2039         }
2040
2041         buffer = kmap(__page);
2042         guard_start = (__u16 *)buffer;
2043         guard_number = PAGE_SIZE / sizeof(*guard_start);
2044         for (i = 0; i < npages; i++) {
2045                 /* corrupt the data before we compute the checksum, to
2046                  * simulate a client->OST data error */
2047                 if (i == 0 && opc == OST_WRITE &&
2048                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2049                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2050                         int len = local_nb[i].lnb_len;
2051                         struct page *np = tgt_page_to_corrupt;
2052
2053                         if (np) {
2054                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2055                                 char *ptr2 = page_address(np);
2056
2057                                 memcpy(ptr2 + off, ptr + off, len);
2058                                 memcpy(ptr2 + off, "bad3", min(4, len));
2059                                 kunmap_atomic(ptr);
2060
2061                                 /* LU-8376 to preserve original index for
2062                                  * display in dump_all_bulk_pages() */
2063                                 np->index = i;
2064
2065                                 cfs_crypto_hash_update_page(req, np, off,
2066                                                             len);
2067                                 continue;
2068                         } else {
2069                                 CERROR("%s: can't alloc page for corruption\n",
2070                                        tgt_name(tgt));
2071                         }
2072                 }
2073
2074                 /*
2075                  * The left guard number should be able to hold checksums of a
2076                  * whole page
2077                  */
2078                 if (t10_cksum_type && opc == OST_READ &&
2079                     local_nb[i].lnb_guard_disk) {
2080                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2081                         if (used > (guard_number - used_number)) {
2082                                 rc = -E2BIG;
2083                                 break;
2084                         }
2085                         memcpy(guard_start + used_number,
2086                                local_nb[i].lnb_guards,
2087                                used * sizeof(*local_nb[i].lnb_guards));
2088                 } else {
2089                         rc = obd_page_dif_generate_buffer(obd_name,
2090                                 local_nb[i].lnb_page,
2091                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2092                                 local_nb[i].lnb_len, guard_start + used_number,
2093                                 guard_number - used_number, &used, sector_size,
2094                                 fn);
2095                         if (rc)
2096                                 break;
2097                 }
2098
2099                 LASSERT(used <= MAX_GUARD_NUMBER);
2100                 /*
2101                  * If disk support T10PI checksum, copy guards to local_nb.
2102                  * If the write is partial page, do not use the guards for bio
2103                  * submission since the data might not be full-sector. The bio
2104                  * guards will be generated later based on the full sectors. If
2105                  * the sector size is 512B rather than 4 KB, or the page size
2106                  * is larger than 4KB, this might drop some useful guards for
2107                  * partial page write, but it will only add minimal extra time
2108                  * of checksum calculation.
2109                  */
2110                 if (t10_cksum_type && opc == OST_WRITE &&
2111                     local_nb[i].lnb_len == PAGE_SIZE) {
2112                         local_nb[i].lnb_guard_rpc = 1;
2113                         memcpy(local_nb[i].lnb_guards,
2114                                guard_start + used_number,
2115                                used * sizeof(*local_nb[i].lnb_guards));
2116                 }
2117
2118                 used_number += used;
2119                 if (used_number == guard_number) {
2120                         cfs_crypto_hash_update_page(req, __page, 0,
2121                                 used_number * sizeof(*guard_start));
2122                         used_number = 0;
2123                 }
2124
2125                  /* corrupt the data after we compute the checksum, to
2126                  * simulate an OST->client data error */
2127                 if (unlikely(i == 0 && opc == OST_READ &&
2128                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2129                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2130                         int len = local_nb[i].lnb_len;
2131                         struct page *np = tgt_page_to_corrupt;
2132
2133                         if (np) {
2134                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2135                                 char *ptr2 = page_address(np);
2136
2137                                 memcpy(ptr2 + off, ptr + off, len);
2138                                 memcpy(ptr2 + off, "bad4", min(4, len));
2139                                 kunmap_atomic(ptr);
2140
2141                                 /* LU-8376 to preserve original index for
2142                                  * display in dump_all_bulk_pages() */
2143                                 np->index = i;
2144
2145                                 cfs_crypto_hash_update_page(req, np, off,
2146                                                             len);
2147                                 continue;
2148                         } else {
2149                                 CERROR("%s: can't alloc page for corruption\n",
2150                                        tgt_name(tgt));
2151                         }
2152                 }
2153         }
2154         kunmap(__page);
2155         if (rc)
2156                 GOTO(out, rc);
2157
2158         if (used_number != 0)
2159                 cfs_crypto_hash_update_page(req, __page, 0,
2160                         used_number * sizeof(*guard_start));
2161
2162         bufsize = sizeof(cksum);
2163         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2164
2165         if (rc == 0)
2166                 *check_sum = cksum;
2167 out:
2168         __free_page(__page);
2169         return rc;
2170 }
2171
2172 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2173                                   enum cksum_types cksum_type,
2174                                   struct niobuf_local *local_nb,
2175                                   int npages, int opc, u32 *check_sum)
2176 {
2177         obd_dif_csum_fn *fn = NULL;
2178         int sector_size = 0;
2179         int rc;
2180
2181         ENTRY;
2182         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2183
2184         if (fn)
2185                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2186                                                opc, fn, sector_size,
2187                                                check_sum);
2188         else
2189                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2190                                          cksum_type, check_sum);
2191         RETURN(rc);
2192 }
2193
2194 int tgt_brw_read(struct tgt_session_info *tsi)
2195 {
2196         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2197         struct ptlrpc_bulk_desc *desc = NULL;
2198         struct obd_export       *exp = tsi->tsi_exp;
2199         struct niobuf_remote    *remote_nb;
2200         struct niobuf_local     *local_nb;
2201         struct obd_ioobj        *ioo;
2202         struct ost_body         *body, *repbody;
2203         struct lustre_handle     lockh = { 0 };
2204         int                      npages, nob = 0, rc, i, no_reply = 0,
2205                                  npages_read;
2206         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2207         const char *obd_name = exp->exp_obd->obd_name;
2208
2209         ENTRY;
2210
2211         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2212             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2213                 CERROR("%s: deny read request from %s to portal %u\n",
2214                        tgt_name(tsi->tsi_tgt),
2215                        obd_export_nid2str(req->rq_export),
2216                        ptlrpc_req2svc(req)->srv_req_portal);
2217                 RETURN(-EPROTO);
2218         }
2219
2220         req->rq_bulk_read = 1;
2221
2222         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2223                 RETURN(-EIO);
2224
2225         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2226                          cfs_fail_val : (obd_timeout + 1) / 4);
2227
2228         /* Check if there is eviction in progress, and if so, wait for it to
2229          * finish */
2230         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2231                 /* We do not care how long it takes */
2232                 wait_event_idle(
2233                         exp->exp_obd->obd_evict_inprogress_waitq,
2234                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2235         }
2236
2237         /* There must be big cache in current thread to process this request
2238          * if it is NULL then something went wrong and it wasn't allocated,
2239          * report -ENOMEM in that case */
2240         if (tbc == NULL)
2241                 RETURN(-ENOMEM);
2242
2243         body = tsi->tsi_ost_body;
2244         LASSERT(body != NULL);
2245
2246         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2247         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2248
2249         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2250         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2251
2252         local_nb = tbc->local;
2253
2254         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2255                           &lockh, LCK_PR);
2256         if (rc != 0)
2257                 RETURN(rc);
2258
2259         /*
2260          * If getting the lock took more time than
2261          * client was willing to wait, drop it. b=11330
2262          */
2263         if (ktime_get_real_seconds() > req->rq_deadline ||
2264             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2265                 no_reply = 1;
2266                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2267                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2268                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2269                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2270                 GOTO(out_lock, rc = -ETIMEDOUT);
2271         }
2272
2273         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2274         repbody->oa = body->oa;
2275
2276         npages = PTLRPC_MAX_BRW_PAGES;
2277         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2278                         ioo, remote_nb, &npages, local_nb);
2279         if (rc != 0)
2280                 GOTO(out_lock, rc);
2281
2282         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2283             body->oa.o_flags & OBD_FL_SHORT_IO) {
2284                 desc = NULL;
2285         } else {
2286                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2287                                             PTLRPC_BULK_PUT_SOURCE,
2288                                             OST_BULK_PORTAL,
2289                                             &ptlrpc_bulk_kiov_nopin_ops);
2290                 if (desc == NULL)
2291                         GOTO(out_commitrw, rc = -ENOMEM);
2292         }
2293
2294         nob = 0;
2295         npages_read = npages;
2296         for (i = 0; i < npages; i++) {
2297                 int page_rc = local_nb[i].lnb_rc;
2298
2299                 if (page_rc < 0) {
2300                         rc = page_rc;
2301                         npages_read = i;
2302                         break;
2303                 }
2304
2305                 nob += page_rc;
2306                 if (page_rc != 0 && desc != NULL) { /* some data! */
2307                         LASSERT(local_nb[i].lnb_page != NULL);
2308                         desc->bd_frag_ops->add_kiov_frag
2309                           (desc, local_nb[i].lnb_page,
2310                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2311                            page_rc);
2312                 }
2313
2314                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2315                         local_nb[i].lnb_len = page_rc;
2316                         npages_read = i + (page_rc != 0 ? 1 : 0);
2317                         /* All subsequent pages should be 0 */
2318                         while (++i < npages)
2319                                 LASSERT(local_nb[i].lnb_rc == 0);
2320                         break;
2321                 }
2322         }
2323
2324         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2325                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2326                            body->oa.o_flags : 0;
2327                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2328
2329                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2330                                                           cksum_type);
2331                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2332
2333                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2334                                             local_nb, npages_read, OST_READ,
2335                                             &repbody->oa.o_cksum);
2336                 if (rc < 0)
2337                         GOTO(out_commitrw, rc);
2338                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2339                        repbody->oa.o_cksum);
2340
2341                 /* if a resend it could be for a cksum error, so check Server
2342                  * cksum with returned Client cksum (this should even cover
2343                  * zero-cksum case) */
2344                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2345                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2346                         check_read_checksum(local_nb, npages_read, exp,
2347                                             &body->oa, &req->rq_peer,
2348                                             body->oa.o_cksum,
2349                                             repbody->oa.o_cksum, cksum_type);
2350         } else {
2351                 repbody->oa.o_valid = 0;
2352         }
2353         /* We're finishing using body->oa as an input variable */
2354
2355         /* Check if client was evicted while we were doing i/o before touching
2356          * network */
2357         if (rc == 0) {
2358                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2359                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2360                         unsigned char *short_io_buf;
2361                         int short_io_size;
2362
2363                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2364                                                               &RMF_SHORT_IO);
2365                         short_io_size = req_capsule_get_size(&req->rq_pill,
2366                                                              &RMF_SHORT_IO,
2367                                                              RCL_SERVER);
2368                         rc = tgt_pages2shortio(local_nb, npages_read,
2369                                                short_io_buf, short_io_size);
2370                         if (rc >= 0)
2371                                 req_capsule_shrink(&req->rq_pill,
2372                                                    &RMF_SHORT_IO, rc,
2373                                                    RCL_SERVER);
2374                         rc = rc > 0 ? 0 : rc;
2375                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2376                         rc = target_bulk_io(exp, desc);
2377                 }
2378                 no_reply = rc != 0;
2379         } else {
2380                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2381                     body->oa.o_flags & OBD_FL_SHORT_IO)
2382                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2383                                            RCL_SERVER);
2384         }
2385
2386 out_commitrw:
2387         /* Must commit after prep above in all cases */
2388         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2389                           remote_nb, npages, local_nb, rc);
2390 out_lock:
2391         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2392
2393         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2394                 ptlrpc_free_bulk(desc);
2395
2396         LASSERT(rc <= 0);
2397         if (rc == 0) {
2398                 rc = nob;
2399                 ptlrpc_lprocfs_brw(req, nob);
2400         } else if (no_reply) {
2401                 req->rq_no_reply = 1;
2402                 /* reply out callback would free */
2403                 ptlrpc_req_drop_rs(req);
2404                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2405                               "client will retry: rc %d\n",
2406                               obd_name,
2407                               obd_uuid2str(&exp->exp_client_uuid),
2408                               obd_export_nid2str(exp), rc);
2409         }
2410         /* send a bulk after reply to simulate a network delay or reordering
2411          * by a router - Note that !desc implies short io, so there is no bulk
2412          * to reorder. */
2413         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2414             desc) {
2415                 CDEBUG(D_INFO, "reorder BULK\n");
2416
2417                 ssleep(3);
2418                 target_bulk_io(exp, desc);
2419                 ptlrpc_free_bulk(desc);
2420         }
2421
2422         RETURN(rc);
2423 }
2424 EXPORT_SYMBOL(tgt_brw_read);
2425
2426 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2427                              unsigned char *buf, unsigned int size)
2428 {
2429         int     i, off, len;
2430         char    *ptr;
2431
2432         for (i = 0; i < npages; i++) {
2433                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2434                 len = local[i].lnb_len;
2435
2436                 if (len == 0)
2437                         continue;
2438
2439                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2440                        i, off, len, size);
2441                 ptr = kmap_atomic(local[i].lnb_page);
2442                 if (ptr == NULL)
2443                         return -EINVAL;
2444                 memcpy(ptr + off, buf, len < size ? len : size);
2445                 kunmap_atomic(ptr);
2446                 buf += len;
2447                 size -= len;
2448         }
2449         return 0;
2450 }
2451
2452 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2453                               struct ptlrpc_bulk_desc *desc,
2454                               struct niobuf_local *local_nb, int npages,
2455                               u32 client_cksum, u32 server_cksum,
2456                               bool mmap)
2457 {
2458         struct obd_export *exp = req->rq_export;
2459         struct ost_body *body;
2460         char *router = "";
2461         char *via = "";
2462
2463         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2464         LASSERT(body != NULL);
2465
2466         if (desc && req->rq_peer.nid != desc->bd_sender) {
2467                 via = " via ";
2468                 router = libcfs_nid2str(desc->bd_sender);
2469         }
2470
2471         if (exp->exp_obd->obd_checksum_dump)
2472                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2473                                     client_cksum);
2474
2475         if (mmap) {
2476                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2477                              client_cksum, server_cksum);
2478                 return;
2479         }
2480
2481         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2482                            DFID" object "DOSTID" extent [%llu-%llu"
2483                            "]: client csum %x, server csum %x\n",
2484                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2485                            via, router,
2486                            body->oa.o_valid & OBD_MD_FLFID ?
2487                            body->oa.o_parent_seq : (__u64)0,
2488                            body->oa.o_valid & OBD_MD_FLFID ?
2489                            body->oa.o_parent_oid : 0,
2490                            body->oa.o_valid & OBD_MD_FLFID ?
2491                            body->oa.o_parent_ver : 0,
2492                            POSTID(&body->oa.o_oi),
2493                            local_nb[0].lnb_file_offset,
2494                            local_nb[npages-1].lnb_file_offset +
2495                            local_nb[npages - 1].lnb_len - 1,
2496                            client_cksum, server_cksum);
2497 }
2498
2499 int tgt_brw_write(struct tgt_session_info *tsi)
2500 {
2501         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2502         struct ptlrpc_bulk_desc *desc = NULL;
2503         struct obd_export       *exp = req->rq_export;
2504         struct niobuf_remote    *remote_nb;
2505         struct niobuf_local     *local_nb;
2506         struct obd_ioobj        *ioo;
2507         struct ost_body         *body, *repbody;
2508         struct lustre_handle     lockh = {0};
2509         __u32                   *rcs;
2510         int                      objcount, niocount, npages;
2511         int                      rc, i, j;
2512         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2513         bool                     no_reply = false, mmap;
2514         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2515         bool wait_sync = false;
2516         const char *obd_name = exp->exp_obd->obd_name;
2517         /* '1' for consistency with code that checks !mpflag to restore */
2518         unsigned int mpflags = 1;
2519
2520         ENTRY;
2521
2522         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2523             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2524                 CERROR("%s: deny write request from %s to portal %u\n",
2525                        tgt_name(tsi->tsi_tgt),
2526                        obd_export_nid2str(req->rq_export),
2527                        ptlrpc_req2svc(req)->srv_req_portal);
2528                 RETURN(err_serious(-EPROTO));
2529         }
2530
2531         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2532                 RETURN(err_serious(-ENOSPC));
2533         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2534                 RETURN(err_serious(-EROFS));
2535
2536         req->rq_bulk_write = 1;
2537
2538         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2539                 RETURN(err_serious(-EIO));
2540         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2541                 RETURN(err_serious(-EFAULT));
2542
2543         /* pause before transaction has been started */
2544         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2545                          cfs_fail_val : (obd_timeout + 1) / 4);
2546
2547         /* Delay write commit to show stale size information */
2548         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2549
2550         /* There must be big cache in current thread to process this request
2551          * if it is NULL then something went wrong and it wasn't allocated,
2552          * report -ENOMEM in that case */
2553         if (tbc == NULL)
2554                 RETURN(-ENOMEM);
2555
2556         body = tsi->tsi_ost_body;
2557         LASSERT(body != NULL);
2558
2559         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2560         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2561
2562         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2563                                         RCL_CLIENT) / sizeof(*ioo);
2564
2565         for (niocount = i = 0; i < objcount; i++)
2566                 niocount += ioo[i].ioo_bufcnt;
2567
2568         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2569         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2570         if (niocount != req_capsule_get_size(&req->rq_pill,
2571                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2572                         sizeof(*remote_nb))
2573                 RETURN(err_serious(-EPROTO));
2574
2575         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2576             ptlrpc_connection_is_local(exp->exp_connection))
2577                 mpflags = memalloc_noreclaim_save();
2578
2579         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2580                              niocount * sizeof(*rcs));
2581         rc = req_capsule_server_pack(&req->rq_pill);
2582         if (rc != 0)
2583                 GOTO(out, rc = err_serious(rc));
2584
2585         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2586         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2587
2588         local_nb = tbc->local;
2589
2590         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2591                           &lockh, LCK_PW);
2592         if (rc != 0)
2593                 GOTO(out, rc);
2594
2595         /*
2596          * If getting the lock took more time than
2597          * client was willing to wait, drop it. b=11330
2598          */
2599         if (ktime_get_real_seconds() > req->rq_deadline ||
2600             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2601                 no_reply = true;
2602                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2603                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2604                        POSTID(&ioo->ioo_oid),
2605                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2606                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2607                 GOTO(out_lock, rc = -ETIMEDOUT);
2608         }
2609
2610         /* Because we already sync grant info with client when reconnect,
2611          * grant info will be cleared for resent req, then fed_grant and
2612          * total_grant will not be modified in following preprw_write */
2613         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2614                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2615                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2616         }
2617
2618         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2619         if (repbody == NULL)
2620                 GOTO(out_lock, rc = -ENOMEM);
2621         repbody->oa = body->oa;
2622
2623         npages = PTLRPC_MAX_BRW_PAGES;
2624         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2625                         objcount, ioo, remote_nb, &npages, local_nb);
2626         if (rc < 0)
2627                 GOTO(out_lock, rc);
2628         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2629             body->oa.o_flags & OBD_FL_SHORT_IO) {
2630                 unsigned int short_io_size;
2631                 unsigned char *short_io_buf;
2632
2633                 short_io_size = req_capsule_get_size(&req->rq_pill,
2634                                                      &RMF_SHORT_IO,
2635                                                      RCL_CLIENT);
2636                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2637                                                       &RMF_SHORT_IO);
2638                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2639                                " size = %d\n", short_io_size);
2640
2641                 /* Copy short io buf to pages */
2642                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2643                                        short_io_size);
2644                 desc = NULL;
2645         } else {
2646                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2647                                             PTLRPC_BULK_GET_SINK,
2648                                             OST_BULK_PORTAL,
2649                                             &ptlrpc_bulk_kiov_nopin_ops);
2650                 if (desc == NULL)
2651                         GOTO(skip_transfer, rc = -ENOMEM);
2652
2653                 /* NB Having prepped, we must commit... */
2654                 for (i = 0; i < npages; i++)
2655                         desc->bd_frag_ops->add_kiov_frag(desc,
2656                                         local_nb[i].lnb_page,
2657                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2658                                         local_nb[i].lnb_len);
2659
2660                 rc = sptlrpc_svc_prep_bulk(req, desc);
2661                 if (rc != 0)
2662                         GOTO(skip_transfer, rc);
2663
2664                 rc = target_bulk_io(exp, desc);
2665         }
2666
2667         no_reply = rc != 0;
2668
2669 skip_transfer:
2670         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2671                 static int cksum_counter;
2672
2673                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2674                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2675
2676                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2677                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2678                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2679                                                            cksum_type);
2680
2681                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2682                                             local_nb, npages, OST_WRITE,
2683                                             &repbody->oa.o_cksum);
2684                 if (rc < 0)
2685                         GOTO(out_commitrw, rc);
2686
2687                 cksum_counter++;
2688
2689                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2690                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2691                                 body->oa.o_flags & OBD_FL_MMAP);
2692
2693                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2694                                           body->oa.o_cksum,
2695                                           repbody->oa.o_cksum, mmap);
2696                         cksum_counter = 0;
2697                 } else if ((cksum_counter & (-cksum_counter)) ==
2698                            cksum_counter) {
2699                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2700                                cksum_counter, libcfs_id2str(req->rq_peer),
2701                                repbody->oa.o_cksum);
2702                 }
2703         }
2704
2705 out_commitrw:
2706         /* Must commit after prep above in all cases */
2707         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2708                           objcount, ioo, remote_nb, npages, local_nb, rc);
2709         if (rc == -ENOTCONN)
2710                 /* quota acquire process has been given up because
2711                  * either the client has been evicted or the client
2712                  * has timed out the request already */
2713                 no_reply = true;
2714
2715         for (i = 0; i < niocount; i++) {
2716                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2717                         wait_sync = true;
2718                         break;
2719                 }
2720         }
2721         /*
2722          * Disable sending mtime back to the client. If the client locked the
2723          * whole object, then it has already updated the mtime on its side,
2724          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2725          */
2726         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2727
2728         if (rc == 0) {
2729                 int nob = 0;
2730
2731                 /* set per-requested niobuf return codes */
2732                 for (i = j = 0; i < niocount; i++) {
2733                         int len = remote_nb[i].rnb_len;
2734
2735                         nob += len;
2736                         rcs[i] = 0;
2737                         do {
2738                                 LASSERT(j < npages);
2739                                 if (local_nb[j].lnb_rc < 0)
2740                                         rcs[i] = local_nb[j].lnb_rc;
2741                                 len -= local_nb[j].lnb_len;
2742                                 j++;
2743                         } while (len > 0);
2744                         LASSERT(len == 0);
2745                 }
2746                 LASSERT(j == npages);
2747                 ptlrpc_lprocfs_brw(req, nob);
2748         }
2749 out_lock:
2750         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2751         if (desc)
2752                 ptlrpc_free_bulk(desc);
2753 out:
2754         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2755                 req->rq_no_reply = 1;
2756                 /* reply out callback would free */
2757                 ptlrpc_req_drop_rs(req);
2758                 if (!exp->exp_obd->obd_no_transno)
2759                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2760                                       " client will retry: rc = %d\n",
2761                                       obd_name,
2762                                       obd_uuid2str(&exp->exp_client_uuid),
2763                                       obd_export_nid2str(exp), rc);
2764         }
2765
2766         if (mpflags)
2767                 memalloc_noreclaim_restore(mpflags);
2768
2769         RETURN(rc);
2770 }
2771 EXPORT_SYMBOL(tgt_brw_write);
2772
2773 /* Check if request can be reconstructed from saved reply data
2774  * A copy of the reply data is returned in @trd if the pointer is not NULL
2775  */
2776 int req_can_reconstruct(struct ptlrpc_request *req,
2777                          struct tg_reply_data *trd)
2778 {
2779         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2780         struct lsd_client_data *lcd = ted->ted_lcd;
2781         int found;
2782
2783         if (tgt_is_multimodrpcs_client(req->rq_export))
2784                 return tgt_lookup_reply(req, trd);
2785
2786         mutex_lock(&ted->ted_lcd_lock);
2787         found = req->rq_xid == lcd->lcd_last_xid ||
2788                 req->rq_xid == lcd->lcd_last_close_xid;
2789
2790         if (found && trd != NULL) {
2791                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2792                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2793                         trd->trd_reply.lrd_transno =
2794                                                 lcd->lcd_last_close_transno;
2795                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2796                 } else {
2797                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2798                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2799                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2800                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2801                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2802                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2803                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2804                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2805                 }
2806         }
2807         mutex_unlock(&ted->ted_lcd_lock);
2808
2809         return found;
2810 }
2811 EXPORT_SYMBOL(req_can_reconstruct);
2812