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                         oa->o_valid |= OBD_MD_FLGROUP;
 162
 163                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 164                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 165                         GOTO(out, rc = -EPROTO);
 166
 167                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 168                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 169                 if (unlikely(rc != 0))
 170                         GOTO(out, rc);
 171
 172                 oi->oi_fid = tti->tti_fid1;
 173         }
 174
 175         RETURN(0);
 176
 177 out:
 178         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 179                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 180                seq, id, rc);
 181         return rc;
 182 }
 183 EXPORT_SYMBOL(tgt_validate_obdo);
 184
 185 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 186 {
 187         unsigned                 max_brw;
 188         struct niobuf_remote    *rnb;
 189         struct obd_ioobj        *ioo;
 190         int                      obj_count;
 191
 192         ENTRY;
 193
 194         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 195         if (ioo == NULL)
 196                 RETURN(-EPROTO);
 197
 198         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 199         if (rnb == NULL)
 200                 RETURN(-EPROTO);
 201
 202         max_brw = ioobj_max_brw_get(ioo);
 203         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 204                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 205                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 206                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 207                        POSTID(oi), -EPROTO);
 208                 RETURN(-EPROTO);
 209         }
 210         ioo->ioo_oid = *oi;
 211
 212         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 213                                         RCL_CLIENT) / sizeof(*ioo);
 214         if (obj_count == 0) {
 215                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 216                 RETURN(-EPROTO);
 217         } else if (obj_count > 1) {
 218                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 219                        obj_count);
 220                 RETURN(-EPROTO);
 221         }
 222
 223         if (ioo->ioo_bufcnt == 0) {
 224                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 225                 RETURN(-EPROTO);
 226         }
 227
 228         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 229                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 230                           "bulk has too many pages (%d)",
 231                           ioo->ioo_bufcnt);
 232                 RETURN(-EPROTO);
 233         }
 234
 235         RETURN(0);
 236 }
 237
 238 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 239 {
 240         struct ost_body         *body;
 241         struct req_capsule      *pill = tsi->tsi_pill;
 242         struct lu_nodemap       *nodemap;
 243         int                      rc;
 244
 245         ENTRY;
 246
 247         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 248         if (body == NULL)
 249                 RETURN(-EFAULT);
 250
 251         rc = tgt_validate_obdo(tsi, &body->oa);
 252         if (rc)
 253                 RETURN(rc);
 254
 255         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 256         if (IS_ERR(nodemap))
 257                 RETURN(PTR_ERR(nodemap));
 258
 259         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 260                                         NODEMAP_CLIENT_TO_FS,
 261                                         body->oa.o_uid);
 262         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_gid);
 265         body->oa.o_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
 266                                            NODEMAP_CLIENT_TO_FS,
 267                                            body->oa.o_projid);
 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 (CFS_FAIL_CHECK_ORSET(h->th_fail_id, CFS_FAIL_ONCE))
 417                 RETURN(0);
 418         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 419                      CFS_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                      CFS_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                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_FILE_ENCCTX,
 459                                           RCL_SERVER))
 460                         req_capsule_set_size(tsi->tsi_pill, &RMF_FILE_ENCCTX,
 461                                              RCL_SERVER, 0);
 462
 463                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_OBD_QUOTA_ITER,
 464                                           RCL_SERVER)) {
 465                         req_capsule_set_size(tsi->tsi_pill,
 466                                             &RMF_OBD_QUOTA_ITER, RCL_SERVER, 0);
 467                 }
 468
 469                 rc = req_capsule_server_pack(tsi->tsi_pill);
 470         }
 471
 472         if (likely(rc == 0)) {
 473                 /*
 474                  * Process request, there can be two types of rc:
 475                  * 1) errors with msg unpack/pack, other failures outside the
 476                  * operation itself. This is counted as serious errors;
 477                  * 2) errors during fs operation, should be placed in rq_status
 478                  * only
 479                  */
 480                 rc = h->th_act(tsi);
 481                 if (!is_serious(rc) &&
 482                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 483                         DEBUG_REQ(D_ERROR, req,
 484                                   "%s: %s handler did not pack reply but returned no error",
 485                                   tgt_name(tsi->tsi_tgt), h->th_name);
 486                         LBUG();
 487                 }
 488                 serious = is_serious(rc);
 489                 rc = clear_serious(rc);
 490         } else {
 491                 serious = 1;
 492         }
 493
 494         req->rq_status = rc;
 495
 496         /*
 497          * ELDLM_* codes which > 0 should be in rq_status only as well as
 498          * all non-serious errors.
 499          */
 500         if (rc > 0 || !serious)
 501                 rc = 0;
 502
 503         LASSERT(current->journal_info == NULL);
 504
 505         if (likely(rc == 0 && req->rq_export))
 506                 target_committed_to_req(req);
 507
 508 out:
 509         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 510         RETURN(0);
 511 }
 512
 513 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 514                                        struct obd_device *obd, int *process)
 515 {
 516         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 517         case MDS_DISCONNECT:
 518         case OST_DISCONNECT:
 519         case OBD_IDX_READ:
 520                 *process = 1;
 521                 RETURN(0);
 522         case MDS_CLOSE:
 523         case MDS_SYNC: /* used in unmounting */
 524         case OBD_PING:
 525         case MDS_REINT:
 526         case OUT_UPDATE:
 527         case MDS_BATCH:
 528         case SEQ_QUERY:
 529         case FLD_QUERY:
 530         case FLD_READ:
 531         case LDLM_ENQUEUE:
 532         case OST_CREATE:
 533         case OST_DESTROY:
 534         case OST_PUNCH:
 535         case OST_SETATTR:
 536         case OST_SYNC:
 537         case OST_WRITE:
 538         case MDS_HSM_PROGRESS:
 539         case MDS_HSM_STATE_SET:
 540         case MDS_HSM_REQUEST:
 541         case OST_FALLOCATE:
 542                 *process = target_queue_recovery_request(req, obd);
 543                 RETURN(0);
 544
 545         default:
 546                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 547                 *process = -EAGAIN;
 548                 RETURN(0);
 549         }
 550 }
 551
 552 /*
 553  * Handle recovery. Return:
 554  *        +1: continue request processing;
 555  *       -ve: abort immediately with the given error code;
 556  *         0: send reply with error code in req->rq_status;
 557  */
 558 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 559 {
 560         ENTRY;
 561
 562         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 563         case MDS_CONNECT:
 564         case OST_CONNECT:
 565         case MGS_CONNECT:
 566         case SEC_CTX_INIT:
 567         case SEC_CTX_INIT_CONT:
 568         case SEC_CTX_FINI:
 569                 RETURN(+1);
 570         }
 571
 572         if (!req->rq_export->exp_obd->obd_replayable)
 573                 RETURN(+1);
 574
 575         /* sanity check: if the xid matches, the request must be marked as a
 576          * resent or replayed */
 577         if (req_can_reconstruct(req, NULL) == 1) {
 578                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 579                       (MSG_RESENT | MSG_REPLAY))) {
 580                         DEBUG_REQ(D_WARNING, req,
 581                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 582                                   req->rq_xid,
 583                                   lustre_msg_get_flags(req->rq_reqmsg));
 584                         req->rq_status = -ENOTCONN;
 585                         RETURN(-ENOTCONN);
 586                 }
 587         }
 588         /* else: note the opposite is not always true; a RESENT req after a
 589          * failover will usually not match the last_xid, since it was likely
 590          * never committed. A REPLAYed request will almost never match the
 591          * last xid, however it could for a committed, but still retained,
 592          * open. */
 593
 594         /* Check for aborted recovery... */
 595         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 596                 int rc;
 597                 int should_process;
 598
 599                 DEBUG_REQ(D_INFO, req, "Got new replay");
 600                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 601                                                  &should_process);
 602                 if (rc != 0 || !should_process)
 603                         RETURN(rc);
 604                 else if (should_process < 0) {
 605                         req->rq_status = should_process;
 606                         rc = ptlrpc_error(req);
 607                         RETURN(rc);
 608                 }
 609         }
 610         RETURN(+1);
 611 }
 612
 613 /* Initial check for request, it is validation mostly */
 614 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 615 {
 616         struct tgt_handler      *h;
 617         struct tgt_opc_slice    *s;
 618         struct lu_target        *tgt;
 619         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 620         /* don't spew error messages for unhandled RPCs */
 621         static bool              printed;
 622
 623         ENTRY;
 624
 625         tgt = class_exp2tgt(req->rq_export);
 626         if (unlikely(tgt == NULL)) {
 627                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 628                           class_exp2obd(req->rq_export)->obd_name);
 629                 RETURN(ERR_PTR(-EINVAL));
 630         }
 631
 632         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 633                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 634                         break;
 635
 636         /* opcode was not found in slice */
 637         if (unlikely(s->tos_hs == NULL)) {
 638                 if (!printed) {
 639                         CERROR("%s: no handler for opcode 0x%x from %s\n",
 640                                tgt_name(tgt), opc, libcfs_idstr(&req->rq_peer));
 641                         printed = true;
 642                 }
 643                 goto err_unsupported;
 644         }
 645
 646         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 647         h = s->tos_hs + (opc - s->tos_opc_start);
 648         if (unlikely(h->th_opc == 0)) {
 649                 if (!printed) {
 650                         CERROR("%s: unsupported opcode 0x%x\n",
 651                                tgt_name(tgt), opc);
 652                         printed = true;
 653                 }
 654                 goto err_unsupported;
 655         }
 656
 657         if (CFS_FAIL_CHECK(OBD_FAIL_OST_OPCODE) && opc == cfs_fail_val)
 658                 goto err_unsupported;
 659
 660         RETURN(h);
 661 err_unsupported:
 662         /*
 663          * Unknown opcode does not necessarily means insane client. A new
 664          * client might send RPCs with new opcodes to an old server. The
 665          * client might desperately stuck there waiting for a reply. So,
 666          * send an error back here.
 667          *
 668          * An old client might also send RPCs with deprecated opcodes (e.g.
 669          * OST_QUOTACHECK).
 670          *
 671          * Error in ptlrpc_send_error() is ignored.
 672          */
 673         req->rq_status = -EOPNOTSUPP;
 674         ptlrpc_send_error(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
 675         RETURN(ERR_PTR(-EOPNOTSUPP));
 676 }
 677
 678 static int process_req_last_xid(struct ptlrpc_request *req)
 679 {
 680         __u64   last_xid;
 681         int rc = 0;
 682         struct obd_export *exp = req->rq_export;
 683         struct tg_export_data *ted = &exp->exp_target_data;
 684         bool need_lock = tgt_is_multimodrpcs_client(exp);
 685         ENTRY;
 686
 687         if (need_lock)
 688                 mutex_lock(&ted->ted_lcd_lock);
 689         /* check request's xid is consistent with export's last_xid */
 690         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 691         if (last_xid > exp->exp_last_xid)
 692                 exp->exp_last_xid = last_xid;
 693
 694         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 695                 /* Some request is allowed to be sent during replay,
 696                  * such as OUT update requests, FLD requests, so it
 697                  * is possible that replay requests has smaller XID
 698                  * than the exp_last_xid.
 699                  *
 700                  * Some non-replay requests may have smaller XID as
 701                  * well:
 702                  *
 703                  * - Client send a no_resend RPC, like statfs;
 704                  * - The RPC timedout (or some other error) on client,
 705                  *   then it's removed from the unreplied list;
 706                  * - Client send some other request to bump the
 707                  *   exp_last_xid on server;
 708                  * - The former RPC got chance to be processed;
 709                  */
 710                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 711                         rc = -EPROTO;
 712
 713                 DEBUG_REQ(D_WARNING, req,
 714                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 715                           req->rq_xid, exp->exp_last_xid, rc);
 716                 if (rc)
 717                         GOTO(out, rc);
 718         }
 719
 720         /* The "last_xid" is the minimum xid among unreplied requests,
 721          * if the request is from the previous connection, its xid can
 722          * still be larger than "exp_last_xid", then the above check of
 723          * xid is not enough to determine whether the request is delayed.
 724          *
 725          * For example, if some replay request was delayed and caused
 726          * timeout at client and the replay is restarted, the delayed
 727          * replay request will have the larger xid than "exp_last_xid"
 728          */
 729         if (req->rq_export->exp_conn_cnt >
 730             lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 731                 CDEBUG(D_RPCTRACE,
 732                        "Dropping request %llu from an old epoch %u/%u\n",
 733                        req->rq_xid,
 734                        lustre_msg_get_conn_cnt(req->rq_reqmsg),
 735                        req->rq_export->exp_conn_cnt);
 736                 req->rq_no_reply = 1;
 737                 GOTO(out, rc = -ESTALE);
 738         }
 739
 740         /* try to release in-memory reply data */
 741         if (tgt_is_multimodrpcs_client(exp)) {
 742                 tgt_handle_received_xid(exp, last_xid);
 743                 rc = tgt_handle_tag(req);
 744         }
 745
 746 out:
 747         if (need_lock)
 748                 mutex_unlock(&ted->ted_lcd_lock);
 749
 750         RETURN(rc);
 751 }
 752
 753 int tgt_request_handle(struct ptlrpc_request *req)
 754 {
 755         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 756
 757         struct lustre_msg       *msg = req->rq_reqmsg;
 758         struct tgt_handler      *h;
 759         struct lu_target        *tgt;
 760         int                      request_fail_id = 0;
 761         __u32                    opc = lustre_msg_get_opc(msg);
 762         struct obd_device       *obd;
 763         int                      rc;
 764         bool                     is_connect = false;
 765         ENTRY;
 766
 767         if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 768                 if (cfs_fail_val == 0 &&
 769                     lustre_msg_get_opc(msg) != OBD_PING &&
 770                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 771                         cfs_fail_val = 1;
 772                         cfs_race_state = 0;
 773                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 774                 }
 775         }
 776
 777         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 778         tsi->tsi_pill = &req->rq_pill;
 779         tsi->tsi_env = req->rq_svc_thread->t_env;
 780
 781         /* if request has export then get handlers slice from corresponding
 782          * target, otherwise that should be connect operation */
 783         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 784             opc == MGS_CONNECT) {
 785                 is_connect = true;
 786                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 787                 rc = target_handle_connect(req);
 788                 if (rc != 0) {
 789                         rc = ptlrpc_error(req);
 790                         GOTO(out, rc);
 791                 }
 792                 /* recovery-small test 18c asks to drop connect reply */
 793                 if (unlikely(opc == OST_CONNECT &&
 794                              CFS_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 795                         GOTO(out, rc = 0);
 796         }
 797
 798         if (unlikely(!class_connected_export(req->rq_export))) {
 799                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 800                     opc == SEC_CTX_FINI) {
 801                         /* sec context initialization has to be handled
 802                          * by hand in tgt_handle_request0() */
 803                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 804                         h = NULL;
 805                         GOTO(handle_recov, rc = 0);
 806                 }
 807                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 808                        opc, libcfs_idstr(&req->rq_peer));
 809                 req->rq_status = -ENOTCONN;
 810                 rc = ptlrpc_error(req);
 811                 GOTO(out, rc);
 812         }
 813
 814         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 815         tsi->tsi_exp = req->rq_export;
 816         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 817                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 818         else
 819                 tsi->tsi_jobid = NULL;
 820
 821         if (tgt == NULL) {
 822                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 823                           class_exp2obd(req->rq_export)->obd_name);
 824                 req->rq_status = -EINVAL;
 825                 rc = ptlrpc_error(req);
 826                 GOTO(out, rc);
 827         }
 828
 829         /* Skip last_xid processing for the recovery thread, otherwise, the
 830          * last_xid on same request could be processed twice: first time when
 831          * processing the incoming request, second time when the request is
 832          * being processed by recovery thread. */
 833         obd = class_exp2obd(req->rq_export);
 834         if (is_connect) {
 835                 /* reset the exp_last_xid on each connection. */
 836                 req->rq_export->exp_last_xid = 0;
 837         } else if (obd->obd_recovery_data.trd_processing_task !=
 838                    current->pid) {
 839                 rc = process_req_last_xid(req);
 840                 if (rc) {
 841                         req->rq_status = rc;
 842                         rc = ptlrpc_error(req);
 843                         GOTO(out, rc);
 844                 }
 845         }
 846
 847         request_fail_id = tgt->lut_request_fail_id;
 848         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 849
 850         h = tgt_handler_find_check(req);
 851         if (IS_ERR(h)) {
 852                 req->rq_status = PTR_ERR(h);
 853                 rc = ptlrpc_error(req);
 854                 GOTO(out, rc);
 855         }
 856
 857         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 858                  h->th_opc, opc);
 859
 860         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 861              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 862                 GOTO(out, rc = 0);
 863
 864         rc = lustre_msg_check_version(msg, h->th_version);
 865         if (unlikely(rc)) {
 866                 DEBUG_REQ(D_ERROR, req,
 867                           "%s: drop malformed request version=%08x expect=%08x",
 868                           tgt_name(tgt), lustre_msg_get_version(msg),
 869                           h->th_version);
 870                 req->rq_status = -EINVAL;
 871                 rc = ptlrpc_error(req);
 872                 GOTO(out, rc);
 873         }
 874
 875 handle_recov:
 876         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 877         if (likely(rc == 1)) {
 878                 rc = tgt_handle_request0(tsi, h, req);
 879                 if (rc)
 880                         GOTO(out, rc);
 881         }
 882         EXIT;
 883 out:
 884         req_capsule_fini(tsi->tsi_pill);
 885         if (tsi->tsi_corpus != NULL) {
 886                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 887                 tsi->tsi_corpus = NULL;
 888         }
 889         return rc;
 890 }
 891 EXPORT_SYMBOL(tgt_request_handle);
 892
 893 /** Assign high priority operations to the request if needed. */
 894 int tgt_hpreq_handler(struct ptlrpc_request *req)
 895 {
 896         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 897         struct tgt_handler      *h;
 898         int                      rc;
 899
 900         ENTRY;
 901
 902         if (req->rq_export == NULL)
 903                 RETURN(0);
 904
 905         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 906         tsi->tsi_pill = &req->rq_pill;
 907         tsi->tsi_env = req->rq_svc_thread->t_env;
 908         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 909         tsi->tsi_exp = req->rq_export;
 910
 911         h = tgt_handler_find_check(req);
 912         if (IS_ERR(h)) {
 913                 rc = PTR_ERR(h);
 914                 RETURN(rc);
 915         }
 916
 917         rc = tgt_request_preprocess(tsi, h, req);
 918         if (unlikely(rc != 0))
 919                 RETURN(rc);
 920
 921         if (h->th_hp != NULL)
 922                 h->th_hp(tsi);
 923         RETURN(0);
 924 }
 925 EXPORT_SYMBOL(tgt_hpreq_handler);
 926
 927 void tgt_counter_incr(struct obd_export *exp, int opcode)
 928 {
 929         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 930         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 931                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 932 }
 933 EXPORT_SYMBOL(tgt_counter_incr);
 934
 935 /*
 936  * Unified target generic handlers.
 937  */
 938
 939 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 940 {
 941         struct lu_target *tgt = class_exp2tgt(exp);
 942         struct sptlrpc_flavor flvr;
 943         int rc = 0;
 944
 945         LASSERT(tgt);
 946         LASSERT(tgt->lut_obd);
 947         LASSERT(tgt->lut_slice);
 948
 949         /* always allow ECHO client */
 950         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 951                             LUSTRE_ECHO_NAME) == 0)) {
 952                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 953                 return 0;
 954         }
 955
 956         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 957                 read_lock(&tgt->lut_sptlrpc_lock);
 958                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 959                                              req->rq_sp_from,
 960                                              &req->rq_peer.nid,
 961                                              &flvr);
 962                 read_unlock(&tgt->lut_sptlrpc_lock);
 963
 964                 spin_lock(&exp->exp_lock);
 965                 exp->exp_sp_peer = req->rq_sp_from;
 966                 exp->exp_flvr = flvr;
 967
 968                 /* when on mgs, if no restriction is set, or if the client
 969                  * NID is on the local node, allow any flavor
 970                  */
 971                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 972                             LUSTRE_MGS_NAME) == 0) &&
 973                     (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 974                      LNetIsPeerLocal(&exp->exp_connection->c_peer.nid)))
 975                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 976
 977                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 978                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 979                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 980                                "expect %x\n", tgt_name(tgt),
 981                                req->rq_flvr.sf_rpc,
 982                                libcfs_nidstr(&req->rq_peer.nid),
 983                                exp->exp_flvr.sf_rpc);
 984                         rc = -EACCES;
 985                 }
 986                 spin_unlock(&exp->exp_lock);
 987         } else {
 988                 if (exp->exp_sp_peer != req->rq_sp_from) {
 989                         CERROR("%s: RPC source %s doesn't match %s\n",
 990                                tgt_name(tgt),
 991                                sptlrpc_part2name(req->rq_sp_from),
 992                                sptlrpc_part2name(exp->exp_sp_peer));
 993                         rc = -EACCES;
 994                 } else {
 995                         rc = sptlrpc_target_export_check(exp, req);
 996                 }
 997         }
 998
 999         return rc;
1000 }
1001
1002 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
1003 {
1004         struct sptlrpc_rule_set  tmp_rset;
1005         int                      rc;
1006
1007         if (unlikely(tgt == NULL)) {
1008                 CERROR("No target passed\n");
1009                 return -EINVAL;
1010         }
1011
1012         sptlrpc_rule_set_init(&tmp_rset);
1013         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
1014         if (rc) {
1015                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
1016                        tgt_name(tgt), rc);
1017                 return rc;
1018         }
1019
1020         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
1021
1022         write_lock(&tgt->lut_sptlrpc_lock);
1023         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
1024         tgt->lut_sptlrpc_rset = tmp_rset;
1025         write_unlock(&tgt->lut_sptlrpc_lock);
1026
1027         return 0;
1028 }
1029 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
1030
1031 int tgt_connect(struct tgt_session_info *tsi)
1032 {
1033         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1034         struct obd_connect_data *reply;
1035         int                      rc;
1036
1037         ENTRY;
1038
1039         /* XXX: better to call this check right after getting new export but
1040          * before last_rcvd slot allocation to avoid server load upon insecure
1041          * connects. This is to be fixed after unifiyng all targets.
1042          */
1043         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1044         if (rc)
1045                 GOTO(out, rc);
1046
1047         /* To avoid exposing partially initialized connection flags, changes up
1048          * to this point have been staged in reply->ocd_connect_flags. Now that
1049          * connection handling has completed successfully, atomically update
1050          * the connect flags in the shared export data structure. LU-1623 */
1051         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1052         spin_lock(&tsi->tsi_exp->exp_lock);
1053         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1054         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1055                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1056                         reply->ocd_connect_flags2;
1057         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1058         spin_unlock(&tsi->tsi_exp->exp_lock);
1059
1060         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1061                    LUSTRE_MDT_NAME) == 0) {
1062                 struct lu_nodemap *nm = NULL;
1063
1064                 rc = req_check_sepol(tsi->tsi_pill);
1065                 if (rc)
1066                         GOTO(out, rc);
1067
1068                 if (tsi->tsi_pill->rc_req->rq_export)
1069                         nm =
1070                          nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1071
1072                 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1073                     reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT) {
1074                         bool forbid_encrypt = true;
1075
1076                         if (!nm)
1077                                 /* nodemap_get_from_exp returns NULL in case
1078                                  * nodemap is not active, so we do not forbid
1079                                  */
1080                                 forbid_encrypt = false;
1081                         else if (!IS_ERR(nm))
1082                                 forbid_encrypt = nm->nmf_forbid_encryption;
1083                         if (forbid_encrypt)
1084                                 GOTO(put_nm, rc = -EACCES);
1085                 }
1086
1087                 if (!(reply->ocd_connect_flags & OBD_CONNECT_RDONLY)) {
1088                         bool readonly = false;
1089
1090                         if (!IS_ERR_OR_NULL(nm))
1091                                 readonly = nm->nmf_readonly_mount;
1092                         if (unlikely(readonly))
1093                                 GOTO(put_nm, rc = -EROFS);
1094                 }
1095
1096 put_nm:
1097                 if (!IS_ERR_OR_NULL(nm))
1098                         nodemap_putref(nm);
1099                 if (rc)
1100                         GOTO(out, rc);
1101         }
1102
1103         RETURN(0);
1104 out:
1105         obd_disconnect(class_export_get(tsi->tsi_exp));
1106         return rc;
1107 }
1108 EXPORT_SYMBOL(tgt_connect);
1109
1110 int tgt_disconnect(struct tgt_session_info *tsi)
1111 {
1112         int rc;
1113
1114         ENTRY;
1115
1116         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1117
1118         rc = target_handle_disconnect(tgt_ses_req(tsi));
1119         if (rc)
1120                 RETURN(err_serious(rc));
1121
1122         RETURN(rc);
1123 }
1124 EXPORT_SYMBOL(tgt_disconnect);
1125
1126 /*
1127  * Unified target OBD handlers
1128  */
1129 int tgt_obd_ping(struct tgt_session_info *tsi)
1130 {
1131         int rc;
1132
1133         ENTRY;
1134
1135         /* The target-specific part of OBD_PING request handling.
1136          * It controls Filter Modification Data (FMD) expiration each time
1137          * PING is received.
1138          *
1139          * Valid only for replayable targets, e.g. MDT and OFD
1140          */
1141         if (tsi->tsi_exp->exp_obd->obd_replayable)
1142                 tgt_fmd_expire(tsi->tsi_exp);
1143
1144         rc = req_capsule_server_pack(tsi->tsi_pill);
1145         if (rc)
1146                 RETURN(err_serious(rc));
1147
1148         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_VS_EVICT)) {
1149                 if (strstr(tsi->tsi_exp->exp_obd->obd_name, "MDT0000") &&
1150                     (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_MDS_MDS))
1151                         tsi->tsi_pill->rc_req->rq_no_reply = 1;
1152         }
1153
1154         RETURN(rc);
1155 }
1156 EXPORT_SYMBOL(tgt_obd_ping);
1157
1158 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1159 {
1160         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1161         struct obd_export       *exp = req->rq_export;
1162         struct ptlrpc_bulk_desc *desc;
1163         int                      i;
1164         int                      rc;
1165         int                      pages = 0;
1166
1167         ENTRY;
1168
1169         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1170                 unsigned int offset;
1171
1172                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1173                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1174                                       PAGE_SIZE);
1175         }
1176
1177         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1178                                   PTLRPC_BULK_PUT_SOURCE,
1179                                     MDS_BULK_PORTAL,
1180                                     &ptlrpc_bulk_kiov_nopin_ops);
1181         if (desc == NULL)
1182                 RETURN(-ENOMEM);
1183
1184         for (i = 0; i < rdbuf->rb_nbufs; i++)
1185                 desc->bd_frag_ops->add_iov_frag(desc,
1186                                         rdbuf->rb_bufs[i].lb_buf,
1187                                         rdbuf->rb_bufs[i].lb_len);
1188
1189         rc = target_bulk_io(exp, desc);
1190         ptlrpc_free_bulk(desc);
1191         RETURN(rc);
1192 }
1193 EXPORT_SYMBOL(tgt_send_buffer);
1194
1195 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1196 {
1197         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1198         struct obd_export       *exp = req->rq_export;
1199         struct ptlrpc_bulk_desc *desc;
1200         int                      tmpcount;
1201         int                      tmpsize;
1202         int                      i;
1203         int                      rc;
1204
1205         ENTRY;
1206
1207         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1208                                     PTLRPC_BULK_PUT_SOURCE,
1209                                     MDS_BULK_PORTAL,
1210                                     &ptlrpc_bulk_kiov_pin_ops);
1211         if (desc == NULL)
1212                 RETURN(-ENOMEM);
1213
1214         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1215                 /* old client requires reply size in it's PAGE_SIZE,
1216                  * which is rdpg->rp_count */
1217                 nob = rdpg->rp_count;
1218
1219         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1220              i++, tmpcount -= tmpsize) {
1221                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1222                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1223                                                  tmpsize);
1224         }
1225
1226         LASSERT(desc->bd_nob == nob);
1227         rc = target_bulk_io(exp, desc);
1228         ptlrpc_free_bulk(desc);
1229         RETURN(rc);
1230 }
1231 EXPORT_SYMBOL(tgt_sendpage);
1232
1233 /*
1234  * OBD_IDX_READ handler
1235  */
1236 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1237 {
1238         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1239         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1240         struct idx_info         *req_ii, *rep_ii;
1241         int                      rc, i;
1242
1243         ENTRY;
1244
1245         memset(rdpg, 0, sizeof(*rdpg));
1246         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1247
1248         /* extract idx_info buffer from request & reply */
1249         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1250         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1251                 RETURN(err_serious(-EPROTO));
1252
1253         rc = req_capsule_server_pack(tsi->tsi_pill);
1254         if (rc)
1255                 RETURN(err_serious(rc));
1256
1257         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1258         if (rep_ii == NULL)
1259                 RETURN(err_serious(-EFAULT));
1260         rep_ii->ii_magic = IDX_INFO_MAGIC;
1261
1262         /* extract hash to start with */
1263         rdpg->rp_hash = req_ii->ii_hash_start;
1264
1265         /* extract requested attributes */
1266         rdpg->rp_attrs = req_ii->ii_attrs;
1267
1268         /* check that fid packed in request is valid and supported */
1269         if (!fid_is_sane(&req_ii->ii_fid))
1270                 RETURN(-EINVAL);
1271         rep_ii->ii_fid = req_ii->ii_fid;
1272
1273         /* copy flags */
1274         rep_ii->ii_flags = req_ii->ii_flags;
1275
1276         /* compute number of pages to allocate, ii_count is the number of 4KB
1277          * containers */
1278         if (req_ii->ii_count <= 0)
1279                 GOTO(out, rc = -EFAULT);
1280         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1281                                exp_max_brw_size(tsi->tsi_exp));
1282         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1283
1284         /* allocate pages to store the containers */
1285         OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1286         if (rdpg->rp_pages == NULL)
1287                 GOTO(out, rc = -ENOMEM);
1288         for (i = 0; i < rdpg->rp_npages; i++) {
1289                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1290                 if (rdpg->rp_pages[i] == NULL)
1291                         GOTO(out, rc = -ENOMEM);
1292         }
1293
1294         /* populate pages with key/record pairs */
1295         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1296         if (rc < 0)
1297                 GOTO(out, rc);
1298
1299         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1300                  "asked %d > %d\n", rc, rdpg->rp_count);
1301
1302         /* send pages to client */
1303         rc = tgt_sendpage(tsi, rdpg, rc);
1304         if (rc)
1305                 GOTO(out, rc);
1306         EXIT;
1307 out:
1308         if (rdpg->rp_pages) {
1309                 for (i = 0; i < rdpg->rp_npages; i++)
1310                         if (rdpg->rp_pages[i])
1311                                 __free_page(rdpg->rp_pages[i]);
1312                 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1313         }
1314         return rc;
1315 }
1316
1317 struct tgt_handler tgt_obd_handlers[] = {
1318 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1319 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1320 };
1321 EXPORT_SYMBOL(tgt_obd_handlers);
1322
1323 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1324              struct dt_object *obj, __u64 start, __u64 end)
1325 {
1326         int rc = 0;
1327
1328         ENTRY;
1329
1330         /* if no objid is specified, it means "sync whole filesystem" */
1331         if (obj == NULL) {
1332                 rc = dt_sync(env, tgt->lut_bottom);
1333         } else if (dt_version_get(env, obj) >
1334                    tgt->lut_obd->obd_last_committed) {
1335                 rc = dt_object_sync(env, obj, start, end);
1336         }
1337         atomic_inc(&tgt->lut_sync_count);
1338
1339         RETURN(rc);
1340 }
1341 EXPORT_SYMBOL(tgt_sync);
1342 /*
1343  * Unified target DLM handlers.
1344  */
1345
1346 /**
1347  * Unified target BAST
1348  *
1349  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1350  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1351  * but if it's cross-MDT lock, because remote object version is not set, a
1352  * filesystem sync is needed.
1353  *
1354  * \param lock server side lock
1355  * \param desc lock desc
1356  * \param data ldlm_cb_set_arg
1357  * \param flag  indicates whether this cancelling or blocking callback
1358  * \retval      0 on success
1359  * \retval      negative number on error
1360  */
1361 int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1362                      void *data, int flag)
1363 {
1364         struct lu_env            env;
1365         struct lu_target        *tgt;
1366         struct dt_object        *obj = NULL;
1367         struct lu_fid            fid;
1368         int                      rc = 0;
1369
1370         ENTRY;
1371
1372         tgt = class_exp2tgt(lock->l_export);
1373
1374         if (unlikely(tgt == NULL)) {
1375                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1376                        class_exp2obd(lock->l_export)->obd_name);
1377                 RETURN(-EINVAL);
1378         }
1379
1380         if (flag == LDLM_CB_CANCELING &&
1381             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1382             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1383              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1384               ldlm_is_cbpending(lock))) &&
1385             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1386              lock->l_resource->lr_type == LDLM_EXTENT)) {
1387                 __u64 start = 0;
1388                 __u64 end = OBD_OBJECT_EOF;
1389
1390                 rc = lu_env_init(&env, LCT_DT_THREAD);
1391                 if (unlikely(rc != 0))
1392                         GOTO(err, rc);
1393
1394                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1395                                    tgt->lut_lsd.lsd_osd_index);
1396
1397                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1398                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1399                         if (IS_ERR(obj))
1400                                 GOTO(err_env, rc = PTR_ERR(obj));
1401
1402                         if (!dt_object_exists(obj))
1403                                 GOTO(err_put, rc = -ENOENT);
1404
1405                         start = lock->l_policy_data.l_extent.start;
1406                         end = lock->l_policy_data.l_extent.end;
1407                 }
1408
1409                 rc = tgt_sync(&env, tgt, obj, start, end);
1410                 if (rc < 0) {
1411                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1412                                "cancel: rc = %d\n",
1413                                tgt_name(tgt), PFID(&fid),
1414                                lock->l_policy_data.l_extent.start,
1415                                lock->l_policy_data.l_extent.end, rc);
1416                 }
1417 err_put:
1418                 if (obj != NULL)
1419                         dt_object_put(&env, obj);
1420 err_env:
1421                 lu_env_fini(&env);
1422         }
1423 err:
1424         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1425         RETURN(rc);
1426 }
1427 EXPORT_SYMBOL(tgt_blocking_ast);
1428
1429 static struct ldlm_callback_suite tgt_dlm_cbs = {
1430         .lcs_completion = ldlm_server_completion_ast,
1431         .lcs_blocking   = tgt_blocking_ast,
1432         .lcs_glimpse    = ldlm_server_glimpse_ast
1433 };
1434
1435 int tgt_enqueue(struct tgt_session_info *tsi)
1436 {
1437         struct ptlrpc_request *req = tgt_ses_req(tsi);
1438         int rc;
1439
1440         ENTRY;
1441         /*
1442          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1443          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1444          */
1445         LASSERT(tsi->tsi_dlm_req != NULL);
1446         rc = ldlm_handle_enqueue(tsi->tsi_exp->exp_obd->obd_namespace,
1447                                  &req->rq_pill, tsi->tsi_dlm_req, &tgt_dlm_cbs);
1448         if (rc)
1449                 RETURN(err_serious(rc));
1450
1451         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1452         case LUT_FAIL_MDT:
1453                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1454                 break;
1455         case LUT_FAIL_OST:
1456                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1457                 break;
1458         case LUT_FAIL_MGT:
1459                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1460                 break;
1461         default:
1462                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1463                 break;
1464         }
1465         RETURN(req->rq_status);
1466 }
1467 EXPORT_SYMBOL(tgt_enqueue);
1468
1469 int tgt_convert(struct tgt_session_info *tsi)
1470 {
1471         struct ptlrpc_request *req = tgt_ses_req(tsi);
1472         int rc;
1473
1474         ENTRY;
1475         LASSERT(tsi->tsi_dlm_req);
1476         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1477         if (rc)
1478                 RETURN(err_serious(rc));
1479
1480         RETURN(req->rq_status);
1481 }
1482
1483 int tgt_bl_callback(struct tgt_session_info *tsi)
1484 {
1485         return err_serious(-EOPNOTSUPP);
1486 }
1487
1488 int tgt_cp_callback(struct tgt_session_info *tsi)
1489 {
1490         return err_serious(-EOPNOTSUPP);
1491 }
1492
1493 /* generic LDLM target handler */
1494 struct tgt_handler tgt_dlm_handlers[] = {
1495 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1496 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1497 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1498 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1499 };
1500 EXPORT_SYMBOL(tgt_dlm_handlers);
1501
1502 /*
1503  * Unified target LLOG handlers.
1504  */
1505 int tgt_llog_open(struct tgt_session_info *tsi)
1506 {
1507         int rc;
1508
1509         ENTRY;
1510
1511         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1512
1513         RETURN(rc);
1514 }
1515 EXPORT_SYMBOL(tgt_llog_open);
1516
1517 int tgt_llog_read_header(struct tgt_session_info *tsi)
1518 {
1519         int rc;
1520
1521         ENTRY;
1522
1523         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1524
1525         RETURN(rc);
1526 }
1527 EXPORT_SYMBOL(tgt_llog_read_header);
1528
1529 int tgt_llog_next_block(struct tgt_session_info *tsi)
1530 {
1531         int rc;
1532
1533         ENTRY;
1534
1535         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1536
1537         RETURN(rc);
1538 }
1539 EXPORT_SYMBOL(tgt_llog_next_block);
1540
1541 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1542 {
1543         int rc;
1544
1545         ENTRY;
1546
1547         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1548
1549         RETURN(rc);
1550 }
1551 EXPORT_SYMBOL(tgt_llog_prev_block);
1552
1553 /* generic llog target handler */
1554 struct tgt_handler tgt_llog_handlers[] = {
1555 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1556 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1557 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1558 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1559 };
1560 EXPORT_SYMBOL(tgt_llog_handlers);
1561
1562 /*
1563  * sec context handlers
1564  */
1565 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1566 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1567 {
1568         return 0;
1569 }
1570
1571 struct tgt_handler tgt_sec_ctx_handlers[] = {
1572 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1573 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1574 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1575 };
1576 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1577
1578 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1579                                  struct dt_device *key,
1580                                  struct lfsck_req_local *lrl,
1581                                  struct thandle *th) = NULL;
1582
1583 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1584                                                       struct dt_device *,
1585                                                       struct lfsck_req_local *,
1586                                                       struct thandle *))
1587 {
1588         tgt_lfsck_in_notify_local = notify;
1589 }
1590 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1591
1592 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1593                            struct dt_device *key,
1594                            struct lfsck_request *lr) = NULL;
1595
1596 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1597                                                 struct dt_device *,
1598                                                 struct lfsck_request *))
1599 {
1600         tgt_lfsck_in_notify = notify;
1601 }
1602 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1603
1604 static int (*tgt_lfsck_query)(const struct lu_env *env,
1605                               struct dt_device *key,
1606                               struct lfsck_request *req,
1607                               struct lfsck_reply *rep,
1608                               struct lfsck_query *que) = NULL;
1609
1610 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1611                                            struct dt_device *,
1612                                            struct lfsck_request *,
1613                                            struct lfsck_reply *,
1614                                            struct lfsck_query *))
1615 {
1616         tgt_lfsck_query = query;
1617 }
1618 EXPORT_SYMBOL(tgt_register_lfsck_query);
1619
1620 /* LFSCK request handlers */
1621 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1622 {
1623         const struct lu_env     *env = tsi->tsi_env;
1624         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1625         struct lfsck_request    *lr;
1626         int                      rc;
1627         ENTRY;
1628
1629         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1630         if (lr == NULL)
1631                 RETURN(-EPROTO);
1632
1633         rc = tgt_lfsck_in_notify(env, key, lr);
1634
1635         RETURN(rc);
1636 }
1637
1638 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1639 {
1640         struct lfsck_request    *request;
1641         struct lfsck_reply      *reply;
1642         int                      rc;
1643         ENTRY;
1644
1645         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1646         if (request == NULL)
1647                 RETURN(-EPROTO);
1648
1649         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1650         if (reply == NULL)
1651                 RETURN(-ENOMEM);
1652
1653         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1654                              request, reply, NULL);
1655
1656         RETURN(rc < 0 ? rc : 0);
1657 }
1658
1659 struct tgt_handler tgt_lfsck_handlers[] = {
1660 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1661 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1662 };
1663 EXPORT_SYMBOL(tgt_lfsck_handlers);
1664
1665 /*
1666  * initialize per-thread page pool (bug 5137).
1667  */
1668 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1669 {
1670         struct tgt_thread_big_cache *tbc;
1671
1672         ENTRY;
1673
1674         LASSERT(thread != NULL);
1675         LASSERT(thread->t_data == NULL);
1676
1677         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1678         if (tbc == NULL)
1679                 RETURN(-ENOMEM);
1680         thread->t_data = tbc;
1681         RETURN(0);
1682 }
1683 EXPORT_SYMBOL(tgt_io_thread_init);
1684
1685 /*
1686  * free per-thread pool created by tgt_thread_init().
1687  */
1688 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1689 {
1690         struct tgt_thread_big_cache *tbc;
1691
1692         ENTRY;
1693
1694         LASSERT(thread != NULL);
1695
1696         /*
1697          * be prepared to handle partially-initialized pools (because this is
1698          * called from ost_io_thread_init() for cleanup.
1699          */
1700         tbc = thread->t_data;
1701         if (tbc != NULL) {
1702                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1703                 thread->t_data = NULL;
1704         }
1705         EXIT;
1706 }
1707 EXPORT_SYMBOL(tgt_io_thread_done);
1708
1709 /**
1710  * Helper function for getting Data-on-MDT file server DLM lock
1711  * if asked by client.
1712  */
1713 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1714                       struct lustre_handle *lh, int mode, __u64 *flags)
1715 {
1716         union ldlm_policy_data policy = {
1717                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1718         };
1719         int rc;
1720
1721         ENTRY;
1722
1723         LASSERT(lh != NULL);
1724         LASSERT(ns != NULL);
1725         LASSERT(!lustre_handle_is_used(lh));
1726
1727         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1728                                     flags, ldlm_blocking_ast,
1729                                     ldlm_completion_ast, ldlm_glimpse_ast,
1730                                     NULL, 0, LVB_T_NONE, NULL, lh);
1731
1732         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1733 }
1734 EXPORT_SYMBOL(tgt_mdt_data_lock);
1735
1736 /**
1737  * Helper function for getting server side [start, start+count] DLM lock
1738  * if asked by client.
1739  */
1740 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1741                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1742                     struct lustre_handle *lh, int mode, __u64 *flags)
1743 {
1744         union ldlm_policy_data policy;
1745         int rc;
1746
1747         ENTRY;
1748
1749         LASSERT(lh != NULL);
1750         LASSERT(ns != NULL);
1751         LASSERT(!lustre_handle_is_used(lh));
1752
1753         policy.l_extent.gid = 0;
1754         policy.l_extent.start = start & PAGE_MASK;
1755
1756         /*
1757          * If ->o_blocks is EOF it means "lock till the end of the file".
1758          * Otherwise, it's size of an extent or hole being punched (in bytes).
1759          */
1760         if (end == OBD_OBJECT_EOF || end < start)
1761                 policy.l_extent.end = OBD_OBJECT_EOF;
1762         else
1763                 policy.l_extent.end = end | ~PAGE_MASK;
1764
1765         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1766                                     mode, flags, ldlm_blocking_ast,
1767                                     ldlm_completion_ast, ldlm_glimpse_ast,
1768                                     NULL, 0, LVB_T_NONE, NULL, lh);
1769         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1770 }
1771 EXPORT_SYMBOL(tgt_extent_lock);
1772
1773 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1774                          struct ldlm_res_id *res_id, __u64 start, __u64 end,
1775                          struct lustre_handle *lh, enum ldlm_mode mode)
1776 {
1777         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1778         __u64 flags = 0;
1779
1780         /* MDT IO for data-on-mdt */
1781         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1782                 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1783
1784         return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1785 }
1786
1787 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1788 {
1789         LASSERT(lustre_handle_is_used(lh));
1790         ldlm_lock_decref(lh, mode);
1791 }
1792 EXPORT_SYMBOL(tgt_data_unlock);
1793
1794 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1795                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1796                         struct niobuf_remote *nb, struct lustre_handle *lh,
1797                         enum ldlm_mode mode)
1798 {
1799         int nrbufs = obj->ioo_bufcnt;
1800         int i;
1801
1802         ENTRY;
1803
1804         LASSERT(mode == LCK_PR || mode == LCK_PW);
1805         LASSERT(!lustre_handle_is_used(lh));
1806
1807         if (exp->exp_obd->obd_recovering)
1808                 RETURN(0);
1809
1810         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1811                 RETURN(0);
1812
1813         for (i = 1; i < nrbufs; i++)
1814                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1815                         RETURN(-EFAULT);
1816
1817         return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1818                              nb[nrbufs - 1].rnb_offset +
1819                              nb[nrbufs - 1].rnb_len - 1, lh, mode);
1820 }
1821
1822 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1823                            struct niobuf_remote *niob,
1824                            struct lustre_handle *lh, enum ldlm_mode mode)
1825 {
1826         ENTRY;
1827
1828         LASSERT(mode == LCK_PR || mode == LCK_PW);
1829         LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1830                  niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1831                 lustre_handle_is_used(lh));
1832
1833         if (lustre_handle_is_used(lh))
1834                 tgt_data_unlock(lh, mode);
1835         EXIT;
1836 }
1837
1838 static int tgt_checksum_niobuf(struct lu_target *tgt,
1839                                  struct niobuf_local *local_nb, int npages,
1840                                  int opc, enum cksum_types cksum_type,
1841                                  __u32 *cksum)
1842 {
1843         struct ahash_request           *req;
1844         unsigned int                    bufsize;
1845         int                             i, err;
1846         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1847
1848         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1849         if (IS_ERR(req)) {
1850                 CERROR("%s: unable to initialize checksum hash %s\n",
1851                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1852                 return PTR_ERR(req);
1853         }
1854
1855         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1856         for (i = 0; i < npages; i++) {
1857                 /* corrupt the data before we compute the checksum, to
1858                  * simulate a client->OST data error */
1859                 if (i == 0 && opc == OST_WRITE &&
1860                     CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1861                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1862                         int len = local_nb[i].lnb_len;
1863                         struct page *np = tgt_page_to_corrupt;
1864
1865                         if (np) {
1866                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1867                                 char *ptr2 = page_address(np);
1868
1869                                 memcpy(ptr2 + off, ptr + off, len);
1870                                 memcpy(ptr2 + off, "bad3", min(4, len));
1871                                 kunmap_atomic(ptr);
1872
1873                                 /* LU-8376 to preserve original index for
1874                                  * display in dump_all_bulk_pages() */
1875                                 np->index = i;
1876
1877                                 cfs_crypto_hash_update_page(req, np, off,
1878                                                             len);
1879                                 continue;
1880                         } else {
1881                                 CERROR("%s: can't alloc page for corruption\n",
1882                                        tgt_name(tgt));
1883                         }
1884                 }
1885                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1886                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1887                                   local_nb[i].lnb_len);
1888
1889                  /* corrupt the data after we compute the checksum, to
1890                  * simulate an OST->client data error */
1891                 if (i == 0 && opc == OST_READ &&
1892                     CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1893                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1894                         int len = local_nb[i].lnb_len;
1895                         struct page *np = tgt_page_to_corrupt;
1896
1897                         if (np) {
1898                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1899                                 char *ptr2 = page_address(np);
1900
1901                                 memcpy(ptr2 + off, ptr + off, len);
1902                                 memcpy(ptr2 + off, "bad4", min(4, len));
1903                                 kunmap_atomic(ptr);
1904
1905                                 /* LU-8376 to preserve original index for
1906                                  * display in dump_all_bulk_pages() */
1907                                 np->index = i;
1908
1909                                 cfs_crypto_hash_update_page(req, np, off,
1910                                                             len);
1911                                 continue;
1912                         } else {
1913                                 CERROR("%s: can't alloc page for corruption\n",
1914                                        tgt_name(tgt));
1915                         }
1916                 }
1917         }
1918
1919         bufsize = sizeof(*cksum);
1920         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1921
1922         return 0;
1923 }
1924
1925 char dbgcksum_file_name[PATH_MAX];
1926
1927 static void dump_all_bulk_pages(struct obdo *oa, int count,
1928                                 struct niobuf_local *local_nb,
1929                                 __u32 server_cksum, __u32 client_cksum)
1930 {
1931         struct file *filp;
1932         int rc, i;
1933         unsigned int len;
1934         char *buf;
1935
1936         /* will only keep dump of pages on first error for the same range in
1937          * file/fid, not during the resends/retries. */
1938         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1939                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1940                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1941                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1942                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1943                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1944                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1945                  local_nb[0].lnb_file_offset,
1946                  local_nb[count-1].lnb_file_offset +
1947                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1948         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1949         filp = filp_open(dbgcksum_file_name,
1950                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1951         if (IS_ERR(filp)) {
1952                 rc = PTR_ERR(filp);
1953                 if (rc == -EEXIST)
1954                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1955                                "checksum error: rc = %d\n", dbgcksum_file_name,
1956                                rc);
1957                 else
1958                         CERROR("%s: can't open to dump pages with checksum "
1959                                "error: rc = %d\n", dbgcksum_file_name, rc);
1960                 return;
1961         }
1962
1963         for (i = 0; i < count; i++) {
1964                 len = local_nb[i].lnb_len;
1965                 buf = kmap(local_nb[i].lnb_page);
1966                 while (len != 0) {
1967                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1968                         if (rc < 0) {
1969                                 CERROR("%s: wanted to write %u but got %d "
1970                                        "error\n", dbgcksum_file_name, len, rc);
1971                                 break;
1972                         }
1973                         len -= rc;
1974                         buf += rc;
1975                 }
1976                 kunmap(local_nb[i].lnb_page);
1977         }
1978
1979         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1980         if (rc)
1981                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1982         filp_close(filp, NULL);
1983
1984         libcfs_debug_dumplog();
1985 }
1986
1987 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1988                                struct obd_export *exp, struct obdo *oa,
1989                                const struct lnet_processid *peer,
1990                                __u32 client_cksum, __u32 server_cksum,
1991                                enum cksum_types server_cksum_type)
1992 {
1993         char *msg;
1994         enum cksum_types cksum_type;
1995         loff_t start, end;
1996
1997         if (unlikely(npages <= 0))
1998                 return 0;
1999
2000         /* unlikely to happen and only if resend does not occur due to cksum
2001          * control failure on Client */
2002         if (unlikely(server_cksum == client_cksum)) {
2003                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
2004                        client_cksum);
2005                 return 0;
2006         }
2007
2008         if (exp->exp_obd->obd_checksum_dump)
2009                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
2010                                     client_cksum);
2011
2012         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
2013                                            oa->o_flags : 0);
2014
2015         if (cksum_type != server_cksum_type)
2016                 msg = "the server may have not used the checksum type specified"
2017                       " in the original request - likely a protocol problem";
2018         else
2019                 msg = "should have changed on the client or in transit";
2020
2021         start = local_nb[0].lnb_file_offset;
2022         end = local_nb[npages-1].lnb_file_offset +
2023                                         local_nb[npages-1].lnb_len - 1;
2024
2025         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
2026                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
2027                 " %x (type %x), server csum %x (type %x)\n",
2028                 exp->exp_obd->obd_name,
2029                 msg, libcfs_nidstr(&peer->nid),
2030                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
2031                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
2032                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
2033                 POSTID(&oa->o_oi),
2034                 start, end, client_cksum, cksum_type, server_cksum,
2035                 server_cksum_type);
2036
2037         return 1;
2038 }
2039
2040 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
2041                              unsigned char *buf, int size)
2042 {
2043         int     i, off, len, copied = size;
2044         char    *ptr;
2045
2046         for (i = 0; i < npages; i++) {
2047                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2048                 len = local[i].lnb_len;
2049
2050                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2051                        i, off, len, size);
2052                 if (len > size)
2053                         return -EINVAL;
2054
2055                 ptr = kmap_atomic(local[i].lnb_page);
2056                 memcpy(buf, ptr + off, len);
2057                 kunmap_atomic(ptr);
2058                 buf += len;
2059                 size -= len;
2060         }
2061         return copied - size;
2062 }
2063
2064 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2065                                      enum cksum_types cksum_type,
2066                                      struct niobuf_local *local_nb, int npages,
2067                                      int opc, obd_dif_csum_fn *fn,
2068                                      int sector_size, u32 *check_sum,
2069                                      bool resend)
2070 {
2071         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2072         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2073         const char *obd_name = tgt->lut_obd->obd_name;
2074         struct ahash_request *req;
2075         unsigned char *buffer;
2076         struct page *__page;
2077         __be16 *guard_start;
2078         int guard_number;
2079         int used_number = 0;
2080         __u32 cksum;
2081         unsigned int bufsize = sizeof(cksum);
2082         int rc = 0, rc2;
2083         int used;
2084         int i;
2085
2086         __page = alloc_page(GFP_KERNEL);
2087         if (__page == NULL)
2088                 return -ENOMEM;
2089
2090         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2091         if (IS_ERR(req)) {
2092                 rc = PTR_ERR(req);
2093                 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
2094                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg), rc);
2095                 goto out;
2096         }
2097
2098         buffer = kmap(__page);
2099         guard_start = (__be16 *)buffer;
2100         guard_number = PAGE_SIZE / sizeof(*guard_start);
2101         if (unlikely(resend))
2102                 CDEBUG(D_PAGE | D_HA, "GRD tags per page = %u\n", guard_number);
2103         for (i = 0; i < npages; i++) {
2104                 bool use_t10_grd;
2105                 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2106                 int len = local_nb[i].lnb_len;
2107                 int guards_needed = DIV_ROUND_UP(off + len, sector_size) -
2108                                         (off / sector_size);
2109
2110                 if (guards_needed > guard_number - used_number) {
2111                         cfs_crypto_hash_update_page(req, __page, 0,
2112                                 used_number * sizeof(*guard_start));
2113                         used_number = 0;
2114                 }
2115
2116                 /* corrupt the data before we compute the checksum, to
2117                  * simulate a client->OST data error */
2118                 if (i == 0 && opc == OST_WRITE &&
2119                     CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2120                         struct page *np = tgt_page_to_corrupt;
2121
2122                         if (np) {
2123                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2124                                 char *ptr2 = page_address(np);
2125
2126                                 memcpy(ptr2 + off, ptr + off, len);
2127                                 memcpy(ptr2 + off, "bad3", min(4, len));
2128                                 kunmap_atomic(ptr);
2129
2130                                 /* LU-8376 to preserve original index for
2131                                  * display in dump_all_bulk_pages() */
2132                                 np->index = i;
2133
2134                                 cfs_crypto_hash_update_page(req, np, off,
2135                                                             len);
2136                                 continue;
2137                         } else {
2138                                 CERROR("%s: can't alloc page for corruption\n",
2139                                        tgt_name(tgt));
2140                         }
2141                 }
2142
2143                 /*
2144                  * The left guard number should be able to hold checksums of a
2145                  * whole page
2146                  */
2147                 use_t10_grd = t10_cksum_type && t10_cksum_type == cksum_type &&
2148                               opc == OST_READ &&
2149                               local_nb[i].lnb_len == PAGE_SIZE &&
2150                               local_nb[i].lnb_guard_disk;
2151                 if (use_t10_grd) {
2152                         used = guards_needed;
2153                         if (used > (guard_number - used_number)) {
2154                                 rc = -E2BIG;
2155                                 CDEBUG(D_PAGE | D_HA,
2156                                        "%s: used %u, guard %u/%u, data size %u+%u, sector_size %u: rc = %d\n",
2157                                        obd_name, used, guard_number,
2158                                        used_number, local_nb[i].lnb_page_offset,
2159                                        local_nb[i].lnb_len, sector_size, rc);
2160                                 break;
2161                         }
2162                         memcpy(guard_start + used_number,
2163                                local_nb[i].lnb_guards,
2164                                used * sizeof(*guard_start));
2165                         if (unlikely(resend))
2166                                 CDEBUG(D_PAGE | D_HA,
2167                                        "lnb[%u]: used %u off %u+%u lnb checksum: %*phN\n",
2168                                        i, used,
2169                                        local_nb[i].lnb_page_offset,
2170                                        local_nb[i].lnb_len,
2171                                        (int)(used * sizeof(*guard_start)),
2172                                        guard_start + used_number);
2173                 }
2174                 if (!use_t10_grd || unlikely(resend)) {
2175                         __be16 guard_tmp[MAX_GUARD_NUMBER];
2176                         __be16 *guards = guard_start + used_number;
2177                         int used_tmp = -1, *usedp = &used;
2178
2179                         if (unlikely(use_t10_grd)) {
2180                                 guards = guard_tmp;
2181                                 usedp = &used_tmp;
2182                         }
2183                         rc = obd_page_dif_generate_buffer(obd_name,
2184                                 local_nb[i].lnb_page,
2185                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2186                                 local_nb[i].lnb_len, guards,
2187                                 guard_number - used_number, usedp, sector_size,
2188                                 fn);
2189                         if (unlikely(resend)) {
2190                                 bool bad = use_t10_grd &&
2191                                         memcmp(guard_tmp,
2192                                                local_nb[i].lnb_guards,
2193                                                used_tmp * sizeof(*guard_tmp));
2194
2195                                 if (bad)
2196                                         CERROR("lnb[%u]: used %u/%u off %u+%u tmp checksum: %*phN\n",
2197                                                i, used, used_tmp,
2198                                                local_nb[i].lnb_page_offset,
2199                                                local_nb[i].lnb_len,
2200                                                (int)(used_tmp * sizeof(*guard_start)),
2201                                                guard_tmp);
2202                                 CDEBUG_LIMIT(D_PAGE | D_HA | (bad ? D_ERROR : 0),
2203                                        "lnb[%u]: used %u/%u off %u+%u gen checksum: %*phN\n",
2204                                        i, used, used_tmp,
2205                                        local_nb[i].lnb_page_offset,
2206                                        local_nb[i].lnb_len,
2207                                        (int)(used * sizeof(*guard_start)),
2208                                        guard_start + used_number);
2209                         }
2210                         if (rc)
2211                                 break;
2212                 }
2213
2214                 LASSERT(used <= MAX_GUARD_NUMBER);
2215                 /*
2216                  * If disk support T10PI checksum, copy guards to local_nb.
2217                  * If the write is partial page, do not use the guards for bio
2218                  * submission since the data might not be full-sector. The bio
2219                  * guards will be generated later based on the full sectors. If
2220                  * the sector size is 512B rather than 4 KB, or the page size
2221                  * is larger than 4KB, this might drop some useful guards for
2222                  * partial page write, but it will only add minimal extra time
2223                  * of checksum calculation.
2224                  */
2225                 if (t10_cksum_type && t10_cksum_type == cksum_type &&
2226                     opc == OST_WRITE &&
2227                     local_nb[i].lnb_len == PAGE_SIZE) {
2228                         local_nb[i].lnb_guard_rpc = 1;
2229                         memcpy(local_nb[i].lnb_guards,
2230                                guard_start + used_number,
2231                                used * sizeof(*local_nb[i].lnb_guards));
2232                 }
2233
2234                 used_number += used;
2235
2236                  /* corrupt the data after we compute the checksum, to
2237                  * simulate an OST->client data error */
2238                 if (unlikely(i == 0 && opc == OST_READ &&
2239                              CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2240                         struct page *np = tgt_page_to_corrupt;
2241
2242                         if (np) {
2243                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2244                                 char *ptr2 = page_address(np);
2245
2246                                 memcpy(ptr2 + off, ptr + off, len);
2247                                 memcpy(ptr2 + off, "bad4", min(4, len));
2248                                 kunmap_atomic(ptr);
2249
2250                                 /* LU-8376 to preserve original index for
2251                                  * display in dump_all_bulk_pages() */
2252                                 np->index = i;
2253
2254                                 cfs_crypto_hash_update_page(req, np, off,
2255                                                             len);
2256                                 continue;
2257                         } else {
2258                                 CERROR("%s: can't alloc page for corruption\n",
2259                                        tgt_name(tgt));
2260                         }
2261                 }
2262         }
2263         kunmap(__page);
2264         if (rc)
2265                 GOTO(out_hash, rc);
2266
2267         if (used_number != 0)
2268                 cfs_crypto_hash_update_page(req, __page, 0,
2269                                             used_number * sizeof(*guard_start));
2270 out_hash:
2271         rc2 = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2272         if (!rc)
2273                 rc = rc2;
2274         if (rc == 0)
2275                 *check_sum = cksum;
2276 out:
2277         __free_page(__page);
2278         return rc;
2279 }
2280
2281 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2282                                   enum cksum_types cksum_type,
2283                                   struct niobuf_local *local_nb,
2284                                   int npages, int opc, u32 *check_sum,
2285                                   bool resend)
2286 {
2287         obd_dif_csum_fn *fn = NULL;
2288         int sector_size = 0;
2289         int rc;
2290
2291         ENTRY;
2292         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2293
2294         if (fn)
2295                 rc = tgt_checksum_niobuf_t10pi(tgt, cksum_type,
2296                                                local_nb, npages,
2297                                                opc, fn, sector_size,
2298                                                check_sum, resend);
2299         else
2300                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2301                                          cksum_type, check_sum);
2302
2303         RETURN(rc);
2304 }
2305
2306 int tgt_brw_read(struct tgt_session_info *tsi)
2307 {
2308         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2309         struct ptlrpc_bulk_desc *desc = NULL;
2310         struct obd_export       *exp = tsi->tsi_exp;
2311         struct niobuf_remote    *remote_nb;
2312         struct niobuf_local     *local_nb;
2313         struct obd_ioobj        *ioo;
2314         struct ost_body         *body, *repbody;
2315         struct lustre_handle     lockh = { 0 };
2316         int                      npages, nob = 0, rc, i, no_reply = 0,
2317                                  npages_read;
2318         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2319         const char *obd_name = exp->exp_obd->obd_name;
2320         ktime_t kstart;
2321
2322         ENTRY;
2323
2324         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2325             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2326                 CERROR("%s: deny read request from %s to portal %u\n",
2327                        tgt_name(tsi->tsi_tgt),
2328                        obd_export_nid2str(req->rq_export),
2329                        ptlrpc_req2svc(req)->srv_req_portal);
2330                 RETURN(-EPROTO);
2331         }
2332
2333         req->rq_bulk_read = 1;
2334
2335         if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) {
2336                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2337                  * this server to fail requests.
2338                  */
2339                 char fail_ost_name[MAX_OBD_NAME];
2340
2341                 if (cfs_fail_val > 0) {
2342                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2343                                  cfs_fail_val - 1);
2344
2345                         if (strstr(obd_name, fail_ost_name))
2346                                 RETURN(err_serious(-EIO));
2347                 } else {
2348                         RETURN(err_serious(-EIO));
2349                 }
2350         }
2351
2352         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2353                          cfs_fail_val : (obd_timeout + 1) / 4);
2354
2355         /* There must be big cache in current thread to process this request
2356          * if it is NULL then something went wrong and it wasn't allocated,
2357          * report -ENOMEM in that case */
2358         if (tbc == NULL)
2359                 RETURN(-ENOMEM);
2360
2361         body = tsi->tsi_ost_body;
2362         LASSERT(body != NULL);
2363
2364         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2365             body->oa.o_flags & OBD_FL_NORPC)
2366                 RETURN(0);
2367
2368         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2369         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2370
2371         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2372         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2373
2374         local_nb = tbc->local;
2375
2376         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2377                           &lockh, LCK_PR);
2378         if (rc != 0)
2379                 RETURN(rc);
2380
2381         /*
2382          * If getting the lock took more time than
2383          * client was willing to wait, drop it. b=11330
2384          */
2385         if (ktime_get_real_seconds() > req->rq_deadline ||
2386             CFS_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2387                 no_reply = 1;
2388                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2389                        libcfs_idstr(&req->rq_peer), POSTID(&ioo->ioo_oid),
2390                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2391                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2392                 GOTO(out_lock, rc = -ETIMEDOUT);
2393         }
2394
2395         /*
2396          * Because we already sync grant info with client when
2397          * reconnect, grant info will be cleared for resent req,
2398          * otherwise, outdated grant count in the rpc would de-sync
2399          * grant counters in case of shrink
2400          */
2401         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2402                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2403                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2404         }
2405
2406         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2407         repbody->oa = body->oa;
2408
2409         npages = PTLRPC_MAX_BRW_PAGES;
2410         kstart = ktime_get();
2411         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2412                         ioo, remote_nb, &npages, local_nb);
2413         if (rc != 0)
2414                 GOTO(out_lock, rc);
2415
2416         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2417             body->oa.o_flags & OBD_FL_SHORT_IO) {
2418                 desc = NULL;
2419         } else {
2420                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2421                                             PTLRPC_BULK_PUT_SOURCE,
2422                                             OST_BULK_PORTAL,
2423                                             &ptlrpc_bulk_kiov_nopin_ops);
2424                 if (desc == NULL)
2425                         GOTO(out_commitrw, rc = -ENOMEM);
2426         }
2427
2428         npages_read = npages;
2429         for (i = 0; i < npages; i++) {
2430                 int page_rc = local_nb[i].lnb_rc;
2431
2432                 if (page_rc < 0) {
2433                         rc = page_rc;
2434                         npages_read = i;
2435                         break;
2436                 }
2437
2438                 nob += page_rc;
2439                 if (page_rc != 0 && desc != NULL) { /* some data! */
2440                         LASSERT(local_nb[i].lnb_page != NULL);
2441                         desc->bd_frag_ops->add_kiov_frag
2442                           (desc, local_nb[i].lnb_page,
2443                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2444                            page_rc);
2445                 }
2446
2447                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2448                         local_nb[i].lnb_len = page_rc;
2449                         npages_read = i + (page_rc != 0 ? 1 : 0);
2450                         /* All subsequent pages should be 0 */
2451                         while (++i < npages)
2452                                 LASSERT(local_nb[i].lnb_rc == 0);
2453                         break;
2454                 }
2455         }
2456
2457         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2458                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2459                            body->oa.o_flags : 0;
2460                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2461                 bool resend = (body->oa.o_valid & OBD_MD_FLFLAGS) &&
2462                         (body->oa.o_flags & OBD_FL_RECOV_RESEND);
2463
2464                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2465                                                           cksum_type);
2466                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2467
2468                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2469                                             local_nb, npages_read, OST_READ,
2470                                             &repbody->oa.o_cksum, resend);
2471                 if (rc < 0)
2472                         GOTO(out_commitrw, rc);
2473                 CDEBUG(D_PAGE | (resend ? D_HA : 0),
2474                        "checksum at read origin: %x (%x)\n",
2475                        repbody->oa.o_cksum, cksum_type);
2476
2477                 /* if a resend it could be for a cksum error, so check Server
2478                  * cksum with returned Client cksum (this should even cover
2479                  * zero-cksum case) */
2480                 if (resend)
2481                         check_read_checksum(local_nb, npages_read, exp,
2482                                             &body->oa, &req->rq_peer,
2483                                             body->oa.o_cksum,
2484                                             repbody->oa.o_cksum, cksum_type);
2485         } else {
2486                 repbody->oa.o_valid = 0;
2487         }
2488         if (body->oa.o_valid & OBD_MD_FLGRANT)
2489                 repbody->oa.o_valid |= OBD_MD_FLGRANT;
2490         /* We're finishing using body->oa as an input variable */
2491
2492         /* Check if client was evicted while we were doing i/o before touching
2493          * network */
2494         if (rc == 0) {
2495                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2496                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2497                         unsigned char *short_io_buf;
2498                         int short_io_size;
2499
2500                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2501                                                               &RMF_SHORT_IO);
2502                         short_io_size = req_capsule_get_size(&req->rq_pill,
2503                                                              &RMF_SHORT_IO,
2504                                                              RCL_SERVER);
2505                         rc = tgt_pages2shortio(local_nb, npages_read,
2506                                                short_io_buf, short_io_size);
2507                         if (rc >= 0)
2508                                 req_capsule_shrink(&req->rq_pill,
2509                                                    &RMF_SHORT_IO, rc,
2510                                                    RCL_SERVER);
2511                         rc = rc > 0 ? 0 : rc;
2512                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2513                         rc = target_bulk_io(exp, desc);
2514                 }
2515                 no_reply = rc != 0;
2516         } else {
2517                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2518                     body->oa.o_flags & OBD_FL_SHORT_IO)
2519                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2520                                            RCL_SERVER);
2521         }
2522
2523 out_commitrw:
2524         /* Must commit after prep above in all cases */
2525         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2526                           remote_nb, npages, local_nb, rc, nob, kstart);
2527 out_lock:
2528         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2529
2530         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2531                 ptlrpc_free_bulk(desc);
2532
2533         LASSERT(rc <= 0);
2534         if (rc == 0) {
2535                 rc = nob;
2536                 ptlrpc_lprocfs_brw(req, nob);
2537         } else if (no_reply) {
2538                 req->rq_no_reply = 1;
2539                 /* reply out callback would free */
2540                 ptlrpc_req_drop_rs(req);
2541                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2542                               "client will retry: rc %d\n",
2543                               obd_name,
2544                               obd_uuid2str(&exp->exp_client_uuid),
2545                               obd_export_nid2str(exp), rc);
2546         }
2547         /* send a bulk after reply to simulate a network delay or reordering
2548          * by a router - Note that !desc implies short io, so there is no bulk
2549          * to reorder. */
2550         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2551             desc) {
2552                 /* Calculate checksum before request transfer, original
2553                  * it is done by target_bulk_io() */
2554                 rc = sptlrpc_svc_wrap_bulk(req, desc);
2555                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
2556                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2557                 else /* old version, bulk matchbits is rq_xid */
2558                         req->rq_mbits = req->rq_xid;
2559
2560                 req->rq_status = rc;
2561                 target_committed_to_req(req);
2562                 target_send_reply(req, 0, 0);
2563
2564                 CDEBUG(D_INFO, "reorder BULK\n");
2565                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
2566                                  cfs_fail_val ? : 3);
2567
2568                 target_bulk_io(exp, desc);
2569                 ptlrpc_free_bulk(desc);
2570         }
2571
2572         RETURN(rc);
2573 }
2574 EXPORT_SYMBOL(tgt_brw_read);
2575
2576 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2577                              unsigned char *buf, unsigned int size)
2578 {
2579         int     i, off, len;
2580         char    *ptr;
2581
2582         for (i = 0; i < npages; i++) {
2583                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2584                 len = local[i].lnb_len;
2585
2586                 if (len == 0)
2587                         continue;
2588
2589                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2590                        i, off, len, size);
2591                 ptr = kmap_atomic(local[i].lnb_page);
2592                 if (ptr == NULL)
2593                         return -EINVAL;
2594                 memcpy(ptr + off, buf, len < size ? len : size);
2595                 kunmap_atomic(ptr);
2596                 buf += len;
2597                 size -= len;
2598         }
2599         return 0;
2600 }
2601
2602 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2603                               struct ptlrpc_bulk_desc *desc,
2604                               struct niobuf_local *local_nb, int npages,
2605                               u32 client_cksum, u32 server_cksum,
2606                               bool mmap)
2607 {
2608         struct obd_export *exp = req->rq_export;
2609         struct ost_body *body;
2610         char *router = "";
2611         char *via = "";
2612
2613         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2614         LASSERT(body != NULL);
2615
2616         if (desc && !nid_same(&req->rq_peer.nid, &desc->bd_sender)) {
2617                 via = " via ";
2618                 router = libcfs_nidstr(&desc->bd_sender);
2619         }
2620
2621         if (exp->exp_obd->obd_checksum_dump)
2622                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2623                                     client_cksum);
2624
2625         if (mmap) {
2626                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2627                              client_cksum, server_cksum);
2628                 return;
2629         }
2630
2631         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2632                            DFID" object "DOSTID" extent [%llu-%llu"
2633                            "]: client csum %x, server csum %x\n",
2634                            exp->exp_obd->obd_name, libcfs_idstr(&req->rq_peer),
2635                            via, router,
2636                            body->oa.o_valid & OBD_MD_FLFID ?
2637                            body->oa.o_parent_seq : (__u64)0,
2638                            body->oa.o_valid & OBD_MD_FLFID ?
2639                            body->oa.o_parent_oid : 0,
2640                            body->oa.o_valid & OBD_MD_FLFID ?
2641                            body->oa.o_parent_ver : 0,
2642                            POSTID(&body->oa.o_oi),
2643                            local_nb[0].lnb_file_offset,
2644                            local_nb[npages-1].lnb_file_offset +
2645                            local_nb[npages - 1].lnb_len - 1,
2646                            client_cksum, server_cksum);
2647 }
2648
2649 int tgt_brw_write(struct tgt_session_info *tsi)
2650 {
2651         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2652         struct ptlrpc_bulk_desc *desc = NULL;
2653         struct obd_export       *exp = req->rq_export;
2654         struct niobuf_remote    *remote_nb;
2655         struct niobuf_local     *local_nb;
2656         struct obd_ioobj        *ioo;
2657         struct ost_body         *body, *repbody;
2658         struct lustre_handle     lockh = {0};
2659         __u32                   *rcs;
2660         int                      objcount, niocount, npages;
2661         int                      rc = 0;
2662         int                      i, j;
2663         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2664         bool                     no_reply = false, mmap;
2665         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2666         bool wait_sync = false;
2667         const char *obd_name = exp->exp_obd->obd_name;
2668         /* '1' for consistency with code that checks !mpflag to restore */
2669         unsigned int mpflags = 1;
2670         ktime_t kstart;
2671         int nob = 0;
2672
2673         ENTRY;
2674
2675         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2676             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2677                 CERROR("%s: deny write request from %s to portal %u\n",
2678                        tgt_name(tsi->tsi_tgt),
2679                        obd_export_nid2str(req->rq_export),
2680                        ptlrpc_req2svc(req)->srv_req_portal);
2681                 RETURN(err_serious(-EPROTO));
2682         }
2683
2684         if (CFS_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2685                 RETURN(err_serious(-ENOSPC));
2686         if (CFS_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2687                 RETURN(err_serious(cfs_fail_val ? -cfs_fail_val : -EROFS));
2688
2689         req->rq_bulk_write = 1;
2690
2691         if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2692                 rc = -EIO;
2693         if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2694                 rc = -EFAULT;
2695         if (rc < 0) {
2696                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2697                  * this server to fail requests.
2698                  */
2699                 char fail_ost_name[MAX_OBD_NAME];
2700
2701                 if (cfs_fail_val > 0) {
2702                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2703                                  cfs_fail_val - 1);
2704
2705                         if (strstr(obd_name, fail_ost_name))
2706                                 RETURN(err_serious(rc));
2707                 } else {
2708                         RETURN(err_serious(rc));
2709                 }
2710         }
2711
2712         /* pause before transaction has been started */
2713         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2714                          cfs_fail_val : (obd_timeout + 1) / 4);
2715
2716         /* Delay write commit to show stale size information */
2717         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2718
2719         /* There must be big cache in current thread to process this request
2720          * if it is NULL then something went wrong and it wasn't allocated,
2721          * report -ENOMEM in that case */
2722         if (tbc == NULL)
2723                 RETURN(-ENOMEM);
2724
2725         body = tsi->tsi_ost_body;
2726         LASSERT(body != NULL);
2727
2728         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2729             body->oa.o_flags & OBD_FL_NORPC)
2730                 RETURN(0);
2731
2732
2733         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2734         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2735
2736         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2737                                         RCL_CLIENT) / sizeof(*ioo);
2738
2739         for (niocount = i = 0; i < objcount; i++)
2740                 niocount += ioo[i].ioo_bufcnt;
2741
2742         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2743         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2744         if (niocount != req_capsule_get_size(&req->rq_pill,
2745                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2746                         sizeof(*remote_nb))
2747                 RETURN(err_serious(-EPROTO));
2748
2749         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2750             ptlrpc_connection_is_local(exp->exp_connection))
2751                 mpflags = memalloc_noreclaim_save();
2752
2753         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2754                              niocount * sizeof(*rcs));
2755         rc = req_capsule_server_pack(&req->rq_pill);
2756         if (rc != 0)
2757                 GOTO(out, rc = err_serious(rc));
2758
2759         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2760         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2761
2762         local_nb = tbc->local;
2763
2764         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2765                           &lockh, LCK_PW);
2766         if (rc != 0)
2767                 GOTO(out, rc);
2768
2769         /*
2770          * If getting the lock took more time than
2771          * client was willing to wait, drop it. b=11330
2772          */
2773         if (ktime_get_real_seconds() > req->rq_deadline ||
2774             CFS_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2775                 no_reply = true;
2776                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2777                        tgt_name(tsi->tsi_tgt), libcfs_idstr(&req->rq_peer),
2778                        POSTID(&ioo->ioo_oid),
2779                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2780                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2781                 GOTO(out_lock, rc = -ETIMEDOUT);
2782         }
2783
2784         /* Because we already sync grant info with client when reconnect,
2785          * grant info will be cleared for resent req, then fed_grant and
2786          * total_grant will not be modified in following preprw_write */
2787         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2788                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2789                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2790         }
2791
2792         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2793         if (repbody == NULL)
2794                 GOTO(out_lock, rc = -ENOMEM);
2795         repbody->oa = body->oa;
2796
2797         npages = PTLRPC_MAX_BRW_PAGES;
2798         kstart = ktime_get();
2799         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2800                         objcount, ioo, remote_nb, &npages, local_nb);
2801         if (rc < 0)
2802                 GOTO(out_lock, rc);
2803         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2804             body->oa.o_flags & OBD_FL_SHORT_IO) {
2805                 unsigned int short_io_size;
2806                 unsigned char *short_io_buf;
2807
2808                 short_io_size = req_capsule_get_size(&req->rq_pill,
2809                                                      &RMF_SHORT_IO,
2810                                                      RCL_CLIENT);
2811                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2812                                                       &RMF_SHORT_IO);
2813                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2814                                " size = %d\n", short_io_size);
2815
2816                 /* Copy short io buf to pages */
2817                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2818                                        short_io_size);
2819                 desc = NULL;
2820         } else {
2821                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2822                                             PTLRPC_BULK_GET_SINK,
2823                                             OST_BULK_PORTAL,
2824                                             &ptlrpc_bulk_kiov_nopin_ops);
2825                 if (desc == NULL)
2826                         GOTO(skip_transfer, rc = -ENOMEM);
2827
2828                 /* NB Having prepped, we must commit... */
2829                 for (i = 0; i < npages; i++)
2830                         desc->bd_frag_ops->add_kiov_frag(desc,
2831                                         local_nb[i].lnb_page,
2832                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2833                                         local_nb[i].lnb_len);
2834
2835                 rc = sptlrpc_svc_prep_bulk(req, desc);
2836                 if (rc != 0)
2837                         GOTO(skip_transfer, rc);
2838
2839                 rc = target_bulk_io(exp, desc);
2840         }
2841
2842         no_reply = rc != 0;
2843
2844 skip_transfer:
2845         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2846                 static int cksum_counter;
2847
2848                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2849                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2850
2851                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2852                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2853                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2854                                                            cksum_type);
2855
2856                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2857                                             local_nb, npages, OST_WRITE,
2858                                             &repbody->oa.o_cksum, false);
2859                 if (rc < 0)
2860                         GOTO(out_commitrw, rc);
2861
2862                 cksum_counter++;
2863
2864                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2865                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2866                                 body->oa.o_flags & OBD_FL_MMAP);
2867
2868                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2869                                           body->oa.o_cksum,
2870                                           repbody->oa.o_cksum, mmap);
2871                         cksum_counter = 0;
2872                 } else if ((cksum_counter & (-cksum_counter)) ==
2873                            cksum_counter) {
2874                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2875                                cksum_counter, libcfs_idstr(&req->rq_peer),
2876                                repbody->oa.o_cksum);
2877                 }
2878         }
2879
2880         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2881
2882 out_commitrw:
2883         /* calculate the expected actual write bytes (nob) for OFD stats.
2884          * Technically, if commit fails this would be wrong, but that should be
2885          * very rare
2886          */
2887         for (i = 0; i < niocount; i++) {
2888                 int len = remote_nb[i].rnb_len;
2889
2890                 nob += len;
2891         }
2892
2893         /* multiple transactions can be assigned during write commit */
2894         tsi->tsi_mult_trans = 1;
2895
2896         /* Must commit after prep above in all cases */
2897         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2898                           objcount, ioo, remote_nb, npages, local_nb, rc, nob,
2899                           kstart);
2900         if (rc == -ENOTCONN)
2901                 /* quota acquire process has been given up because
2902                  * either the client has been evicted or the client
2903                  * has timed out the request already
2904                  */
2905                 no_reply = true;
2906
2907         for (i = 0; i < niocount; i++) {
2908                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2909                         wait_sync = true;
2910                         break;
2911                 }
2912         }
2913         /*
2914          * Disable sending mtime back to the client. If the client locked the
2915          * whole object, then it has already updated the mtime on its side,
2916          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2917          */
2918         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2919
2920         if (rc == 0) {
2921                 /* set per-requested niobuf return codes */
2922                 for (i = j = 0; i < niocount; i++) {
2923                         int len = remote_nb[i].rnb_len;
2924
2925                         rcs[i] = 0;
2926                         do {
2927                                 LASSERT(j < npages);
2928                                 if (local_nb[j].lnb_rc < 0)
2929                                         rcs[i] = local_nb[j].lnb_rc;
2930                                 len -= local_nb[j].lnb_len;
2931                                 j++;
2932                         } while (len > 0);
2933                         LASSERT(len == 0);
2934                 }
2935                 LASSERT(j == npages);
2936                 ptlrpc_lprocfs_brw(req, nob);
2937         }
2938 out_lock:
2939         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2940         if (desc)
2941                 ptlrpc_free_bulk(desc);
2942 out:
2943         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2944                 req->rq_no_reply = 1;
2945                 /* reply out callback would free */
2946                 ptlrpc_req_drop_rs(req);
2947                 if (!exp->exp_obd->obd_no_transno)
2948                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2949                                       " client will retry: rc = %d\n",
2950                                       obd_name,
2951                                       obd_uuid2str(&exp->exp_client_uuid),
2952                                       obd_export_nid2str(exp), rc);
2953         }
2954
2955         if (mpflags)
2956                 memalloc_noreclaim_restore(mpflags);
2957
2958         RETURN(rc);
2959 }
2960 EXPORT_SYMBOL(tgt_brw_write);
2961
2962 /**
2963  * Common request handler for OST_SEEK RPC.
2964  *
2965  * Unified request handling for OST_SEEK RPC.
2966  * It takes object by its FID, does needed lseek and packs result
2967  * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2968  *
2969  * \param[in] tsi       target session environment for this request
2970  *
2971  * \retval              0 if successful
2972  * \retval              negative value on error
2973  */
2974 int tgt_lseek(struct tgt_session_info *tsi)
2975 {
2976         struct lustre_handle lh = { 0 };
2977         struct dt_object *dob;
2978         struct ost_body *repbody;
2979         loff_t offset = tsi->tsi_ost_body->oa.o_size;
2980         int whence = tsi->tsi_ost_body->oa.o_mode;
2981         bool srvlock;
2982         int rc = 0;
2983
2984         ENTRY;
2985
2986         if (whence != SEEK_HOLE && whence != SEEK_DATA)
2987                 RETURN(-EPROTO);
2988
2989         /* Negative offset is prohibited on wire and must be handled on client
2990          * prior sending RPC.
2991          */
2992         if (offset < 0)
2993                 RETURN(-EPROTO);
2994
2995         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2996         if (repbody == NULL)
2997                 RETURN(-ENOMEM);
2998         repbody->oa = tsi->tsi_ost_body->oa;
2999
3000         srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
3001                   tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
3002         if (srvlock) {
3003                 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
3004                                    offset, OBD_OBJECT_EOF, &lh, LCK_PR);
3005                 if (rc)
3006                         RETURN(rc);
3007         }
3008
3009         dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
3010         if (IS_ERR(dob))
3011                 GOTO(out, rc = PTR_ERR(dob));
3012
3013         if (!dt_object_exists(dob))
3014                 GOTO(obj_put, rc = -ENOENT);
3015
3016         repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
3017         rc = 0;
3018 obj_put:
3019         dt_object_put(tsi->tsi_env, dob);
3020 out:
3021         if (srvlock)
3022                 tgt_data_unlock(&lh, LCK_PR);
3023
3024         RETURN(rc);
3025 }
3026 EXPORT_SYMBOL(tgt_lseek);
3027
3028 /* Check if request can be reconstructed from saved reply data
3029  * A copy of the reply data is returned in @trd if the pointer is not NULL
3030  */
3031 int req_can_reconstruct(struct ptlrpc_request *req,
3032                          struct tg_reply_data *trd)
3033 {
3034         struct tg_export_data *ted = &req->rq_export->exp_target_data;
3035         struct lsd_client_data *lcd = ted->ted_lcd;
3036         int found;
3037
3038         if (tgt_is_multimodrpcs_client(req->rq_export))
3039                 return tgt_lookup_reply(req, trd);
3040
3041         mutex_lock(&ted->ted_lcd_lock);
3042         found = req->rq_xid == lcd->lcd_last_xid ||
3043                 req->rq_xid == lcd->lcd_last_close_xid;
3044
3045         if (found && trd != NULL) {
3046                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
3047                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
3048                         trd->trd_reply.lrd_transno =
3049                                                 lcd->lcd_last_close_transno;
3050                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
3051                 } else {
3052                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
3053                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
3054                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
3055                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
3056                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
3057                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
3058                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
3059                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
3060                 }
3061         }
3062         mutex_unlock(&ted->ted_lcd_lock);
3063
3064         return found;
3065 }
3066 EXPORT_SYMBOL(req_can_reconstruct);
3067
3068 bool tgt_check_resent(struct ptlrpc_request *req, struct tg_reply_data *trd)
3069 {
3070         struct lsd_reply_data *lrd;
3071         bool need_reconstruct = false;
3072
3073         ENTRY;
3074
3075         if (likely(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
3076                 return false;
3077
3078         if (req_can_reconstruct(req, trd)) {
3079                 lrd = &trd->trd_reply;
3080                 req->rq_transno = lrd->lrd_transno;
3081                 req->rq_status = lrd->lrd_result;
3082                 if (req->rq_status != 0)
3083                         req->rq_transno = 0;
3084                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
3085                 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
3086
3087                 DEBUG_REQ(D_HA, req, "reconstruct resent RPC");
3088                 need_reconstruct = true;
3089         } else {
3090                 DEBUG_REQ(D_HA, req, "no reply for RESENT req");
3091         }
3092
3093         return need_reconstruct;
3094 }
3095 EXPORT_SYMBOL(tgt_check_resent);