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 <obd.h>
  42 #include <obd_class.h>
  43 #include <obd_cksum.h>
  44 #include <lustre_lfsck.h>
  45 #include <lustre_nodemap.h>
  46 #include <lustre_acl.h>
  47
  48 #include "tgt_internal.h"
  49
  50 char *tgt_name(struct lu_target *tgt)
  51 {
  52         LASSERT(tgt->lut_obd != NULL);
  53         return tgt->lut_obd->obd_name;
  54 }
  55 EXPORT_SYMBOL(tgt_name);
  56
  57 /*
  58  * Generic code handling requests that have struct mdt_body passed in:
  59  *
  60  *  - extract mdt_body from request and save it in @tsi, if present;
  61  *
  62  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  63  *  @tsi;
  64  *
  65  *  - if HAS_BODY flag is set for this request type check whether object
  66  *  actually exists on storage (lu_object_exists()).
  67  *
  68  */
  69 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  70 {
  71         const struct mdt_body   *body;
  72         struct lu_object        *obj;
  73         struct req_capsule      *pill = tsi->tsi_pill;
  74         int                      rc;
  75
  76         ENTRY;
  77
  78         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  79         if (body == NULL)
  80                 RETURN(-EFAULT);
  81
  82         tsi->tsi_mdt_body = body;
  83
  84         if (!(body->mbo_valid & OBD_MD_FLID))
  85                 RETURN(0);
  86
  87         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  88          * in any case in pre-2.5 clients. Fix that here if needed */
  89         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  90                 RETURN(0);
  91
  92         if (!fid_is_sane(&body->mbo_fid1)) {
  93                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  94                        PFID(&body->mbo_fid1));
  95                 RETURN(-EINVAL);
  96         }
  97
  98         obj = lu_object_find(tsi->tsi_env,
  99                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 100                              &body->mbo_fid1, NULL);
 101         if (!IS_ERR(obj)) {
 102                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 103                         lu_object_put(tsi->tsi_env, obj);
 104                         rc = -ENOENT;
 105                 } else {
 106                         tsi->tsi_corpus = obj;
 107                         rc = 0;
 108                 }
 109         } else {
 110                 rc = PTR_ERR(obj);
 111         }
 112
 113         tsi->tsi_fid = body->mbo_fid1;
 114
 115         RETURN(rc);
 116 }
 117
 118 /**
 119  * Validate oa from client.
 120  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 121  * req are valid.
 122  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 123  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 124  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 125  *       whether this is oi_fid or real ostid. So it will check
 126  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 127  *    c. Old FID-disable osc will send IDIF.
 128  *    d. new FID-enable osc/osp will send normal FID.
 129  *
 130  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 131  * be used for LAST_ID file, and only being accessed inside OST now.
 132  */
 133 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 134 {
 135         struct ost_id   *oi     = &oa->o_oi;
 136         u64              seq    = ostid_seq(oi);
 137         u64              id     = ostid_id(oi);
 138         int              rc;
 139         ENTRY;
 140
 141         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 142                      fid_seq_is_echo(seq))) {
 143                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 144                  * during create, and we will reset this to 1, since this
 145                  * oi_id is basically useless in the following create process,
 146                  * but oi_id == 0 will make it difficult to tell whether it is
 147                  * real FID or ost_id. */
 148                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 149                 oi->oi_fid.f_oid = id ?: 1;
 150                 oi->oi_fid.f_ver = 0;
 151         } else {
 152                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 153
 154                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 155                         GOTO(out, rc = -EPROTO);
 156
 157                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 158                  * all of the requests are required to setup FLGROUP */
 159                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 160                         ostid_set_seq_mdt0(oi);
 161                         oa->o_valid |= OBD_MD_FLGROUP;
 162                         seq = ostid_seq(oi);
 163                 }
 164
 165                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 166                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 167                         GOTO(out, rc = -EPROTO);
 168
 169                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 170                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 171                 if (unlikely(rc != 0))
 172                         GOTO(out, rc);
 173
 174                 oi->oi_fid = tti->tti_fid1;
 175         }
 176
 177         RETURN(0);
 178
 179 out:
 180         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 181                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 182                seq, id, rc);
 183         return rc;
 184 }
 185 EXPORT_SYMBOL(tgt_validate_obdo);
 186
 187 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 188 {
 189         unsigned                 max_brw;
 190         struct niobuf_remote    *rnb;
 191         struct obd_ioobj        *ioo;
 192         int                      obj_count;
 193
 194         ENTRY;
 195
 196         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 197         if (ioo == NULL)
 198                 RETURN(-EPROTO);
 199
 200         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 201         if (rnb == NULL)
 202                 RETURN(-EPROTO);
 203
 204         max_brw = ioobj_max_brw_get(ioo);
 205         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 206                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 207                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 208                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 209                        POSTID(oi), -EPROTO);
 210                 RETURN(-EPROTO);
 211         }
 212         ioo->ioo_oid = *oi;
 213
 214         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 215                                         RCL_CLIENT) / sizeof(*ioo);
 216         if (obj_count == 0) {
 217                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 218                 RETURN(-EPROTO);
 219         } else if (obj_count > 1) {
 220                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 221                        obj_count);
 222                 RETURN(-EPROTO);
 223         }
 224
 225         if (ioo->ioo_bufcnt == 0) {
 226                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 227                 RETURN(-EPROTO);
 228         }
 229
 230         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 231                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 232                           "bulk has too many pages (%d)",
 233                           ioo->ioo_bufcnt);
 234                 RETURN(-EPROTO);
 235         }
 236
 237         RETURN(0);
 238 }
 239
 240 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 241 {
 242         struct ost_body         *body;
 243         struct req_capsule      *pill = tsi->tsi_pill;
 244         struct lu_nodemap       *nodemap;
 245         int                      rc;
 246
 247         ENTRY;
 248
 249         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 250         if (body == NULL)
 251                 RETURN(-EFAULT);
 252
 253         rc = tgt_validate_obdo(tsi, &body->oa);
 254         if (rc)
 255                 RETURN(rc);
 256
 257         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 258         if (IS_ERR(nodemap))
 259                 RETURN(PTR_ERR(nodemap));
 260
 261         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 262                                         NODEMAP_CLIENT_TO_FS,
 263                                         body->oa.o_uid);
 264         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 265                                         NODEMAP_CLIENT_TO_FS,
 266                                         body->oa.o_gid);
 267         nodemap_putref(nodemap);
 268
 269         tsi->tsi_ost_body = body;
 270         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 271
 272         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 273                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 274                 if (rc < 0)
 275                         RETURN(rc);
 276         }
 277
 278         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 279                 if (flags & HAS_BODY) {
 280                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 281                                tgt_name(tsi->tsi_tgt));
 282                         RETURN(-EPROTO);
 283                 } else {
 284                         RETURN(0);
 285                 }
 286         }
 287
 288         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 289
 290         /*
 291          * OST doesn't get object in advance for further use to prevent
 292          * situations with nested object_find which is potential deadlock.
 293          */
 294         tsi->tsi_corpus = NULL;
 295         RETURN(rc);
 296 }
 297
 298 /*
 299  * Do necessary preprocessing according to handler ->th_flags.
 300  */
 301 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 302                                   struct tgt_handler *h,
 303                                   struct ptlrpc_request *req)
 304 {
 305         struct req_capsule      *pill = tsi->tsi_pill;
 306         __u32                    flags = h->th_flags;
 307         int                      rc = 0;
 308
 309         ENTRY;
 310
 311         if (tsi->tsi_preprocessed)
 312                 RETURN(0);
 313
 314         LASSERT(h->th_act != NULL);
 315         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 316         LASSERT(current->journal_info == NULL);
 317
 318         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 319                      h->th_fmt != NULL));
 320         if (h->th_fmt != NULL) {
 321                 req_capsule_set(pill, h->th_fmt);
 322                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT) &&
 323                     req_capsule_field_present(pill, &RMF_MDT_BODY,
 324                                               RCL_CLIENT)) {
 325                         rc = tgt_mdt_body_unpack(tsi, flags);
 326                         if (rc < 0)
 327                                 RETURN(rc);
 328                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 329                                                  RCL_CLIENT) &&
 330                            req_capsule_field_present(pill, &RMF_OST_BODY,
 331                                                  RCL_CLIENT)) {
 332                         rc = tgt_ost_body_unpack(tsi, flags);
 333                         if (rc < 0)
 334                                 RETURN(rc);
 335                 }
 336         }
 337
 338         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 339                 RETURN(-EROFS);
 340
 341         if (flags & HAS_KEY) {
 342                 struct ldlm_request *dlm_req;
 343
 344                 LASSERT(h->th_fmt != NULL);
 345
 346                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 347                 if (dlm_req != NULL) {
 348                         union ldlm_wire_policy_data *policy =
 349                                         &dlm_req->lock_desc.l_policy_data;
 350
 351                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 352                                      LDLM_IBITS &&
 353                                      (policy->l_inodebits.bits |
 354                                       policy->l_inodebits.try_bits) == 0)) {
 355                                 /*
 356                                  * Lock without inodebits makes no sense and
 357                                  * will oops later in ldlm. If client miss to
 358                                  * set such bits, do not trigger ASSERTION.
 359                                  *
 360                                  * For liblustre flock case, it maybe zero.
 361                                  */
 362                                 rc = -EPROTO;
 363                         } else {
 364                                 tsi->tsi_dlm_req = dlm_req;
 365                         }
 366                 } else {
 367                         rc = -EFAULT;
 368                 }
 369         }
 370         tsi->tsi_preprocessed = 1;
 371         RETURN(rc);
 372 }
 373
 374 /*
 375  * Invoke handler for this request opc. Also do necessary preprocessing
 376  * (according to handler ->th_flags), and post-processing (setting of
 377  * ->last_{xid,committed}).
 378  */
 379 static int tgt_handle_request0(struct tgt_session_info *tsi,
 380                                struct tgt_handler *h,
 381                                struct ptlrpc_request *req)
 382 {
 383         int      serious = 0;
 384         int      rc;
 385         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 386
 387         ENTRY;
 388
 389
 390         /* When dealing with sec context requests, no export is associated yet,
 391          * because these requests are sent before *_CONNECT requests.
 392          * A NULL req->rq_export means the normal *_common_slice handlers will
 393          * not be called, because there is no reference to the target.
 394          * So deal with them by hand and jump directly to target_send_reply().
 395          */
 396         switch (opc) {
 397         case SEC_CTX_INIT:
 398         case SEC_CTX_INIT_CONT:
 399         case SEC_CTX_FINI:
 400                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 401                 GOTO(out, rc = 0);
 402         }
 403
 404         /*
 405          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 406          * to put same checks into handlers like mdt_close(), mdt_reint(),
 407          * etc., without talking to mdt authors first. Checking same thing
 408          * there again is useless and returning 0 error without packing reply
 409          * is buggy! Handlers either pack reply or return error.
 410          *
 411          * We return 0 here and do not send any reply in order to emulate
 412          * network failure. Do not send any reply in case any of NET related
 413          * fail_id has occured.
 414          */
 415         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 416                 RETURN(0);
 417         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 418                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 419                 RETURN(0);
 420
 421         rc = tgt_request_preprocess(tsi, h, req);
 422         /* pack reply if reply format is fixed */
 423         if (rc == 0 && h->th_flags & HAS_REPLY) {
 424                 /* Pack reply */
 425                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 426                                           RCL_SERVER))
 427                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 428                                              RCL_SERVER,
 429                                              tsi->tsi_mdt_body->mbo_eadatasize);
 430                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 431                                           RCL_SERVER))
 432                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 433                                              RCL_SERVER, 0);
 434                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 435                         req_capsule_set_size(tsi->tsi_pill,
 436                                              &RMF_ACL, RCL_SERVER,
 437                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 438
 439                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 440                                           RCL_SERVER)) {
 441                         struct niobuf_remote *remote_nb =
 442                                 req_capsule_client_get(tsi->tsi_pill,
 443                                                        &RMF_NIOBUF_REMOTE);
 444                         struct ost_body *body = tsi->tsi_ost_body;
 445
 446                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 447                                          RCL_SERVER,
 448                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 449                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 450                                          remote_nb[0].rnb_len : 0);
 451                 }
 452
 453                 rc = req_capsule_server_pack(tsi->tsi_pill);
 454         }
 455
 456         if (likely(rc == 0)) {
 457                 /*
 458                  * Process request, there can be two types of rc:
 459                  * 1) errors with msg unpack/pack, other failures outside the
 460                  * operation itself. This is counted as serious errors;
 461                  * 2) errors during fs operation, should be placed in rq_status
 462                  * only
 463                  */
 464                 rc = h->th_act(tsi);
 465                 if (!is_serious(rc) &&
 466                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 467                         DEBUG_REQ(D_ERROR, req,
 468                                   "%s: %s handler did not pack reply but returned no error",
 469                                   tgt_name(tsi->tsi_tgt), h->th_name);
 470                         LBUG();
 471                 }
 472                 serious = is_serious(rc);
 473                 rc = clear_serious(rc);
 474         } else {
 475                 serious = 1;
 476         }
 477
 478         req->rq_status = rc;
 479
 480         /*
 481          * ELDLM_* codes which > 0 should be in rq_status only as well as
 482          * all non-serious errors.
 483          */
 484         if (rc > 0 || !serious)
 485                 rc = 0;
 486
 487         LASSERT(current->journal_info == NULL);
 488
 489         if (likely(rc == 0 && req->rq_export))
 490                 target_committed_to_req(req);
 491
 492 out:
 493         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 494         RETURN(0);
 495 }
 496
 497 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 498                                        struct obd_device *obd, int *process)
 499 {
 500         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 501         case MDS_DISCONNECT:
 502         case OST_DISCONNECT:
 503         case OBD_IDX_READ:
 504                 *process = 1;
 505                 RETURN(0);
 506         case MDS_CLOSE:
 507         case MDS_SYNC: /* used in unmounting */
 508         case OBD_PING:
 509         case MDS_REINT:
 510         case OUT_UPDATE:
 511         case SEQ_QUERY:
 512         case FLD_QUERY:
 513         case FLD_READ:
 514         case LDLM_ENQUEUE:
 515         case OST_CREATE:
 516         case OST_DESTROY:
 517         case OST_PUNCH:
 518         case OST_SETATTR:
 519         case OST_SYNC:
 520         case OST_WRITE:
 521         case MDS_HSM_PROGRESS:
 522         case MDS_HSM_STATE_SET:
 523         case MDS_HSM_REQUEST:
 524                 *process = target_queue_recovery_request(req, obd);
 525                 RETURN(0);
 526
 527         default:
 528                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 529                 *process = -EAGAIN;
 530                 RETURN(0);
 531         }
 532 }
 533
 534 /*
 535  * Handle recovery. Return:
 536  *        +1: continue request processing;
 537  *       -ve: abort immediately with the given error code;
 538  *         0: send reply with error code in req->rq_status;
 539  */
 540 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 541 {
 542         ENTRY;
 543
 544         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 545         case MDS_CONNECT:
 546         case OST_CONNECT:
 547         case MGS_CONNECT:
 548         case SEC_CTX_INIT:
 549         case SEC_CTX_INIT_CONT:
 550         case SEC_CTX_FINI:
 551                 RETURN(+1);
 552         }
 553
 554         if (!req->rq_export->exp_obd->obd_replayable)
 555                 RETURN(+1);
 556
 557         /* sanity check: if the xid matches, the request must be marked as a
 558          * resent or replayed */
 559         if (req_can_reconstruct(req, NULL) == 1) {
 560                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 561                       (MSG_RESENT | MSG_REPLAY))) {
 562                         DEBUG_REQ(D_WARNING, req,
 563                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 564                                   req->rq_xid,
 565                                   lustre_msg_get_flags(req->rq_reqmsg));
 566                         req->rq_status = -ENOTCONN;
 567                         RETURN(-ENOTCONN);
 568                 }
 569         }
 570         /* else: note the opposite is not always true; a RESENT req after a
 571          * failover will usually not match the last_xid, since it was likely
 572          * never committed. A REPLAYed request will almost never match the
 573          * last xid, however it could for a committed, but still retained,
 574          * open. */
 575
 576         /* Check for aborted recovery... */
 577         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 578                 int rc;
 579                 int should_process;
 580
 581                 DEBUG_REQ(D_INFO, req, "Got new replay");
 582                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 583                                                  &should_process);
 584                 if (rc != 0 || !should_process)
 585                         RETURN(rc);
 586                 else if (should_process < 0) {
 587                         req->rq_status = should_process;
 588                         rc = ptlrpc_error(req);
 589                         RETURN(rc);
 590                 }
 591         }
 592         RETURN(+1);
 593 }
 594
 595 /* Initial check for request, it is validation mostly */
 596 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 597 {
 598         struct tgt_handler      *h;
 599         struct tgt_opc_slice    *s;
 600         struct lu_target        *tgt;
 601         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 602
 603         ENTRY;
 604
 605         tgt = class_exp2tgt(req->rq_export);
 606         if (unlikely(tgt == NULL)) {
 607                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 608                           class_exp2obd(req->rq_export)->obd_name);
 609                 RETURN(ERR_PTR(-EINVAL));
 610         }
 611
 612         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 613                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 614                         break;
 615
 616         /* opcode was not found in slice */
 617         if (unlikely(s->tos_hs == NULL)) {
 618                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 619                        opc);
 620                 RETURN(ERR_PTR(-ENOTSUPP));
 621         }
 622
 623         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 624         h = s->tos_hs + (opc - s->tos_opc_start);
 625         if (unlikely(h->th_opc == 0)) {
 626                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 627                 RETURN(ERR_PTR(-ENOTSUPP));
 628         }
 629
 630         RETURN(h);
 631 }
 632
 633 static int process_req_last_xid(struct ptlrpc_request *req)
 634 {
 635         __u64   last_xid;
 636         int rc = 0;
 637         struct obd_export *exp = req->rq_export;
 638         struct tg_export_data *ted = &exp->exp_target_data;
 639         bool need_lock = tgt_is_multimodrpcs_client(exp);
 640         ENTRY;
 641
 642         if (need_lock)
 643                 mutex_lock(&ted->ted_lcd_lock);
 644         /* check request's xid is consistent with export's last_xid */
 645         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 646         if (last_xid > exp->exp_last_xid)
 647                 exp->exp_last_xid = last_xid;
 648
 649         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 650                 /* Some request is allowed to be sent during replay,
 651                  * such as OUT update requests, FLD requests, so it
 652                  * is possible that replay requests has smaller XID
 653                  * than the exp_last_xid.
 654                  *
 655                  * Some non-replay requests may have smaller XID as
 656                  * well:
 657                  *
 658                  * - Client send a no_resend RPC, like statfs;
 659                  * - The RPC timedout (or some other error) on client,
 660                  *   then it's removed from the unreplied list;
 661                  * - Client send some other request to bump the
 662                  *   exp_last_xid on server;
 663                  * - The former RPC got chance to be processed;
 664                  */
 665                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 666                         rc = -EPROTO;
 667
 668                 DEBUG_REQ(D_WARNING, req,
 669                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 670                           req->rq_xid, exp->exp_last_xid, rc);
 671                 if (rc)
 672                         GOTO(out, rc);
 673         }
 674
 675         /* The "last_xid" is the minimum xid among unreplied requests,
 676          * if the request is from the previous connection, its xid can
 677          * still be larger than "exp_last_xid", then the above check of
 678          * xid is not enough to determine whether the request is delayed.
 679          *
 680          * For example, if some replay request was delayed and caused
 681          * timeout at client and the replay is restarted, the delayed
 682          * replay request will have the larger xid than "exp_last_xid"
 683          */
 684         if (req->rq_export->exp_conn_cnt >
 685             lustre_msg_get_conn_cnt(req->rq_reqmsg))
 686                 GOTO(out, rc = -ESTALE);
 687
 688         /* try to release in-memory reply data */
 689         if (tgt_is_multimodrpcs_client(exp)) {
 690                 tgt_handle_received_xid(exp, last_xid);
 691                 rc = tgt_handle_tag(req);
 692         }
 693
 694 out:
 695         if (need_lock)
 696                 mutex_unlock(&ted->ted_lcd_lock);
 697
 698         RETURN(rc);
 699 }
 700
 701 int tgt_request_handle(struct ptlrpc_request *req)
 702 {
 703         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 704
 705         struct lustre_msg       *msg = req->rq_reqmsg;
 706         struct tgt_handler      *h;
 707         struct lu_target        *tgt;
 708         int                      request_fail_id = 0;
 709         __u32                    opc = lustre_msg_get_opc(msg);
 710         struct obd_device       *obd;
 711         int                      rc;
 712         bool                     is_connect = false;
 713         ENTRY;
 714
 715         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 716                 if (cfs_fail_val == 0 &&
 717                     lustre_msg_get_opc(msg) != OBD_PING &&
 718                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 719                         cfs_fail_val = 1;
 720                         cfs_race_state = 0;
 721                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 722                 }
 723         }
 724
 725         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 726         tsi->tsi_pill = &req->rq_pill;
 727         tsi->tsi_env = req->rq_svc_thread->t_env;
 728
 729         /* if request has export then get handlers slice from corresponding
 730          * target, otherwise that should be connect operation */
 731         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 732             opc == MGS_CONNECT) {
 733                 is_connect = true;
 734                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 735                 rc = target_handle_connect(req);
 736                 if (rc != 0) {
 737                         rc = ptlrpc_error(req);
 738                         GOTO(out, rc);
 739                 }
 740                 /* recovery-small test 18c asks to drop connect reply */
 741                 if (unlikely(opc == OST_CONNECT &&
 742                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 743                         GOTO(out, rc = 0);
 744         }
 745
 746         if (unlikely(!class_connected_export(req->rq_export))) {
 747                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 748                     opc == SEC_CTX_FINI) {
 749                         /* sec context initialization has to be handled
 750                          * by hand in tgt_handle_request0() */
 751                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 752                         h = NULL;
 753                         GOTO(handle_recov, rc = 0);
 754                 }
 755                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 756                        opc, libcfs_id2str(req->rq_peer));
 757                 req->rq_status = -ENOTCONN;
 758                 rc = ptlrpc_error(req);
 759                 GOTO(out, rc);
 760         }
 761
 762         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 763         tsi->tsi_exp = req->rq_export;
 764         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 765                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 766         else
 767                 tsi->tsi_jobid = NULL;
 768
 769         if (tgt == NULL) {
 770                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 771                           class_exp2obd(req->rq_export)->obd_name);
 772                 req->rq_status = -EINVAL;
 773                 rc = ptlrpc_error(req);
 774                 GOTO(out, rc);
 775         }
 776
 777         /* Skip last_xid processing for the recovery thread, otherwise, the
 778          * last_xid on same request could be processed twice: first time when
 779          * processing the incoming request, second time when the request is
 780          * being processed by recovery thread. */
 781         obd = class_exp2obd(req->rq_export);
 782         if (is_connect) {
 783                 /* reset the exp_last_xid on each connection. */
 784                 req->rq_export->exp_last_xid = 0;
 785         } else if (obd->obd_recovery_data.trd_processing_task !=
 786                    current->pid) {
 787                 rc = process_req_last_xid(req);
 788                 if (rc) {
 789                         req->rq_status = rc;
 790                         rc = ptlrpc_error(req);
 791                         GOTO(out, rc);
 792                 }
 793         }
 794
 795         request_fail_id = tgt->lut_request_fail_id;
 796         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 797
 798         h = tgt_handler_find_check(req);
 799         if (IS_ERR(h)) {
 800                 req->rq_status = PTR_ERR(h);
 801                 rc = ptlrpc_error(req);
 802                 GOTO(out, rc);
 803         }
 804
 805         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 806                  h->th_opc, opc);
 807
 808         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 809              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 810                 GOTO(out, rc = 0);
 811
 812         rc = lustre_msg_check_version(msg, h->th_version);
 813         if (unlikely(rc)) {
 814                 DEBUG_REQ(D_ERROR, req,
 815                           "%s: drop malformed request version=%08x expect=%08x",
 816                           tgt_name(tgt), lustre_msg_get_version(msg),
 817                           h->th_version);
 818                 req->rq_status = -EINVAL;
 819                 rc = ptlrpc_error(req);
 820                 GOTO(out, rc);
 821         }
 822
 823 handle_recov:
 824         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 825         if (likely(rc == 1)) {
 826                 rc = tgt_handle_request0(tsi, h, req);
 827                 if (rc)
 828                         GOTO(out, rc);
 829         }
 830         EXIT;
 831 out:
 832         req_capsule_fini(tsi->tsi_pill);
 833         if (tsi->tsi_corpus != NULL) {
 834                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 835                 tsi->tsi_corpus = NULL;
 836         }
 837         return rc;
 838 }
 839 EXPORT_SYMBOL(tgt_request_handle);
 840
 841 /** Assign high priority operations to the request if needed. */
 842 int tgt_hpreq_handler(struct ptlrpc_request *req)
 843 {
 844         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 845         struct tgt_handler      *h;
 846         int                      rc;
 847
 848         ENTRY;
 849
 850         if (req->rq_export == NULL)
 851                 RETURN(0);
 852
 853         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 854         tsi->tsi_pill = &req->rq_pill;
 855         tsi->tsi_env = req->rq_svc_thread->t_env;
 856         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 857         tsi->tsi_exp = req->rq_export;
 858
 859         h = tgt_handler_find_check(req);
 860         if (IS_ERR(h)) {
 861                 rc = PTR_ERR(h);
 862                 RETURN(rc);
 863         }
 864
 865         rc = tgt_request_preprocess(tsi, h, req);
 866         if (unlikely(rc != 0))
 867                 RETURN(rc);
 868
 869         if (h->th_hp != NULL)
 870                 h->th_hp(tsi);
 871         RETURN(0);
 872 }
 873 EXPORT_SYMBOL(tgt_hpreq_handler);
 874
 875 void tgt_counter_incr(struct obd_export *exp, int opcode)
 876 {
 877         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 878         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 879                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 880 }
 881 EXPORT_SYMBOL(tgt_counter_incr);
 882
 883 /*
 884  * Unified target generic handlers.
 885  */
 886
 887 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 888 {
 889         struct lu_target *tgt = class_exp2tgt(exp);
 890         struct sptlrpc_flavor flvr;
 891         int rc = 0;
 892
 893         LASSERT(tgt);
 894         LASSERT(tgt->lut_obd);
 895         LASSERT(tgt->lut_slice);
 896
 897         /* always allow ECHO client */
 898         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 899                             LUSTRE_ECHO_NAME) == 0)) {
 900                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 901                 return 0;
 902         }
 903
 904         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 905                 read_lock(&tgt->lut_sptlrpc_lock);
 906                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 907                                              req->rq_sp_from,
 908                                              req->rq_peer.nid,
 909                                              &flvr);
 910                 read_unlock(&tgt->lut_sptlrpc_lock);
 911
 912                 spin_lock(&exp->exp_lock);
 913                 exp->exp_sp_peer = req->rq_sp_from;
 914                 exp->exp_flvr = flvr;
 915
 916                 /* when on mgs, if no restriction is set, or if client
 917                  * is loopback, allow any flavor */
 918                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 919                            LUSTRE_MGS_NAME) == 0) &&
 920                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 921                       exp->exp_connection->c_peer.nid == LNET_NID_LO_0))
 922                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 923
 924                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 925                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 926                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 927                                "expect %x\n", tgt_name(tgt),
 928                                req->rq_flvr.sf_rpc,
 929                                libcfs_nid2str(req->rq_peer.nid),
 930                                exp->exp_flvr.sf_rpc);
 931                         rc = -EACCES;
 932                 }
 933                 spin_unlock(&exp->exp_lock);
 934         } else {
 935                 if (exp->exp_sp_peer != req->rq_sp_from) {
 936                         CERROR("%s: RPC source %s doesn't match %s\n",
 937                                tgt_name(tgt),
 938                                sptlrpc_part2name(req->rq_sp_from),
 939                                sptlrpc_part2name(exp->exp_sp_peer));
 940                         rc = -EACCES;
 941                 } else {
 942                         rc = sptlrpc_target_export_check(exp, req);
 943                 }
 944         }
 945
 946         return rc;
 947 }
 948
 949 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 950 {
 951         struct sptlrpc_rule_set  tmp_rset;
 952         int                      rc;
 953
 954         if (unlikely(tgt == NULL)) {
 955                 CERROR("No target passed");
 956                 return -EINVAL;
 957         }
 958
 959         sptlrpc_rule_set_init(&tmp_rset);
 960         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 961         if (rc) {
 962                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 963                        tgt_name(tgt), rc);
 964                 return rc;
 965         }
 966
 967         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 968
 969         write_lock(&tgt->lut_sptlrpc_lock);
 970         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 971         tgt->lut_sptlrpc_rset = tmp_rset;
 972         write_unlock(&tgt->lut_sptlrpc_lock);
 973
 974         return 0;
 975 }
 976 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 977
 978 int tgt_connect(struct tgt_session_info *tsi)
 979 {
 980         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 981         struct obd_connect_data *reply;
 982         int                      rc;
 983
 984         ENTRY;
 985
 986         /* XXX: better to call this check right after getting new export but
 987          * before last_rcvd slot allocation to avoid server load upon insecure
 988          * connects. This is to be fixed after unifiyng all targets.
 989          */
 990         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 991         if (rc)
 992                 GOTO(out, rc);
 993
 994         /* To avoid exposing partially initialized connection flags, changes up
 995          * to this point have been staged in reply->ocd_connect_flags. Now that
 996          * connection handling has completed successfully, atomically update
 997          * the connect flags in the shared export data structure. LU-1623 */
 998         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
 999         spin_lock(&tsi->tsi_exp->exp_lock);
1000         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1001         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1002                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1003                         reply->ocd_connect_flags2;
1004         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1005         spin_unlock(&tsi->tsi_exp->exp_lock);
1006
1007         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1008                    LUSTRE_MDT_NAME) == 0) {
1009                 rc = req_check_sepol(tsi->tsi_pill);
1010                 if (rc)
1011                         GOTO(out, rc);
1012         }
1013
1014         RETURN(0);
1015 out:
1016         obd_disconnect(class_export_get(tsi->tsi_exp));
1017         return rc;
1018 }
1019 EXPORT_SYMBOL(tgt_connect);
1020
1021 int tgt_disconnect(struct tgt_session_info *tsi)
1022 {
1023         int rc;
1024
1025         ENTRY;
1026
1027         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1028
1029         rc = target_handle_disconnect(tgt_ses_req(tsi));
1030         if (rc)
1031                 RETURN(err_serious(rc));
1032
1033         RETURN(rc);
1034 }
1035 EXPORT_SYMBOL(tgt_disconnect);
1036
1037 /*
1038  * Unified target OBD handlers
1039  */
1040 int tgt_obd_ping(struct tgt_session_info *tsi)
1041 {
1042         int rc;
1043
1044         ENTRY;
1045
1046         /* The target-specific part of OBD_PING request handling.
1047          * It controls Filter Modification Data (FMD) expiration each time
1048          * PING is received.
1049          *
1050          * Valid only for replayable targets, e.g. MDT and OFD
1051          */
1052         if (tsi->tsi_exp->exp_obd->obd_replayable)
1053                 tgt_fmd_expire(tsi->tsi_exp);
1054
1055         rc = req_capsule_server_pack(tsi->tsi_pill);
1056         if (rc)
1057                 RETURN(err_serious(rc));
1058
1059         RETURN(rc);
1060 }
1061 EXPORT_SYMBOL(tgt_obd_ping);
1062
1063 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1064 {
1065         return err_serious(-EOPNOTSUPP);
1066 }
1067
1068 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1069 {
1070         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1071         struct obd_export       *exp = req->rq_export;
1072         struct ptlrpc_bulk_desc *desc;
1073         int                      i;
1074         int                      rc;
1075         int                      pages = 0;
1076
1077         ENTRY;
1078
1079         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1080                 unsigned int offset;
1081
1082                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1083                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1084                                       PAGE_SIZE);
1085         }
1086
1087         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1088                                   PTLRPC_BULK_PUT_SOURCE,
1089                                     MDS_BULK_PORTAL,
1090                                     &ptlrpc_bulk_kiov_nopin_ops);
1091         if (desc == NULL)
1092                 RETURN(-ENOMEM);
1093
1094         for (i = 0; i < rdbuf->rb_nbufs; i++)
1095                 desc->bd_frag_ops->add_iov_frag(desc,
1096                                         rdbuf->rb_bufs[i].lb_buf,
1097                                         rdbuf->rb_bufs[i].lb_len);
1098
1099         rc = target_bulk_io(exp, desc);
1100         ptlrpc_free_bulk(desc);
1101         RETURN(rc);
1102 }
1103 EXPORT_SYMBOL(tgt_send_buffer);
1104
1105 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1106 {
1107         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1108         struct obd_export       *exp = req->rq_export;
1109         struct ptlrpc_bulk_desc *desc;
1110         int                      tmpcount;
1111         int                      tmpsize;
1112         int                      i;
1113         int                      rc;
1114
1115         ENTRY;
1116
1117         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1118                                     PTLRPC_BULK_PUT_SOURCE,
1119                                     MDS_BULK_PORTAL,
1120                                     &ptlrpc_bulk_kiov_pin_ops);
1121         if (desc == NULL)
1122                 RETURN(-ENOMEM);
1123
1124         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1125                 /* old client requires reply size in it's PAGE_SIZE,
1126                  * which is rdpg->rp_count */
1127                 nob = rdpg->rp_count;
1128
1129         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1130              i++, tmpcount -= tmpsize) {
1131                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1132                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1133                                                  tmpsize);
1134         }
1135
1136         LASSERT(desc->bd_nob == nob);
1137         rc = target_bulk_io(exp, desc);
1138         ptlrpc_free_bulk(desc);
1139         RETURN(rc);
1140 }
1141 EXPORT_SYMBOL(tgt_sendpage);
1142
1143 /*
1144  * OBD_IDX_READ handler
1145  */
1146 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1147 {
1148         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1149         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1150         struct idx_info         *req_ii, *rep_ii;
1151         int                      rc, i;
1152
1153         ENTRY;
1154
1155         memset(rdpg, 0, sizeof(*rdpg));
1156         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1157
1158         /* extract idx_info buffer from request & reply */
1159         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1160         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1161                 RETURN(err_serious(-EPROTO));
1162
1163         rc = req_capsule_server_pack(tsi->tsi_pill);
1164         if (rc)
1165                 RETURN(err_serious(rc));
1166
1167         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1168         if (rep_ii == NULL)
1169                 RETURN(err_serious(-EFAULT));
1170         rep_ii->ii_magic = IDX_INFO_MAGIC;
1171
1172         /* extract hash to start with */
1173         rdpg->rp_hash = req_ii->ii_hash_start;
1174
1175         /* extract requested attributes */
1176         rdpg->rp_attrs = req_ii->ii_attrs;
1177
1178         /* check that fid packed in request is valid and supported */
1179         if (!fid_is_sane(&req_ii->ii_fid))
1180                 RETURN(-EINVAL);
1181         rep_ii->ii_fid = req_ii->ii_fid;
1182
1183         /* copy flags */
1184         rep_ii->ii_flags = req_ii->ii_flags;
1185
1186         /* compute number of pages to allocate, ii_count is the number of 4KB
1187          * containers */
1188         if (req_ii->ii_count <= 0)
1189                 GOTO(out, rc = -EFAULT);
1190         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1191                                exp_max_brw_size(tsi->tsi_exp));
1192         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1193
1194         /* allocate pages to store the containers */
1195         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1196         if (rdpg->rp_pages == NULL)
1197                 GOTO(out, rc = -ENOMEM);
1198         for (i = 0; i < rdpg->rp_npages; i++) {
1199                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1200                 if (rdpg->rp_pages[i] == NULL)
1201                         GOTO(out, rc = -ENOMEM);
1202         }
1203
1204         /* populate pages with key/record pairs */
1205         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1206         if (rc < 0)
1207                 GOTO(out, rc);
1208
1209         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1210                  "asked %d > %d\n", rc, rdpg->rp_count);
1211
1212         /* send pages to client */
1213         rc = tgt_sendpage(tsi, rdpg, rc);
1214         if (rc)
1215                 GOTO(out, rc);
1216         EXIT;
1217 out:
1218         if (rdpg->rp_pages) {
1219                 for (i = 0; i < rdpg->rp_npages; i++)
1220                         if (rdpg->rp_pages[i])
1221                                 __free_page(rdpg->rp_pages[i]);
1222                 OBD_FREE(rdpg->rp_pages,
1223                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1224         }
1225         return rc;
1226 }
1227
1228 struct tgt_handler tgt_obd_handlers[] = {
1229 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1230 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1231 };
1232 EXPORT_SYMBOL(tgt_obd_handlers);
1233
1234 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1235              struct dt_object *obj, __u64 start, __u64 end)
1236 {
1237         int rc = 0;
1238
1239         ENTRY;
1240
1241         /* if no objid is specified, it means "sync whole filesystem" */
1242         if (obj == NULL) {
1243                 rc = dt_sync(env, tgt->lut_bottom);
1244         } else if (dt_version_get(env, obj) >
1245                    tgt->lut_obd->obd_last_committed) {
1246                 rc = dt_object_sync(env, obj, start, end);
1247         }
1248         atomic_inc(&tgt->lut_sync_count);
1249
1250         RETURN(rc);
1251 }
1252 EXPORT_SYMBOL(tgt_sync);
1253 /*
1254  * Unified target DLM handlers.
1255  */
1256
1257 /**
1258  * Unified target BAST
1259  *
1260  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1261  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1262  * but if it's cross-MDT lock, because remote object version is not set, a
1263  * filesystem sync is needed.
1264  *
1265  * \param lock server side lock
1266  * \param desc lock desc
1267  * \param data ldlm_cb_set_arg
1268  * \param flag  indicates whether this cancelling or blocking callback
1269  * \retval      0 on success
1270  * \retval      negative number on error
1271  */
1272 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1273                             void *data, int flag)
1274 {
1275         struct lu_env            env;
1276         struct lu_target        *tgt;
1277         struct dt_object        *obj = NULL;
1278         struct lu_fid            fid;
1279         int                      rc = 0;
1280
1281         ENTRY;
1282
1283         tgt = class_exp2tgt(lock->l_export);
1284
1285         if (unlikely(tgt == NULL)) {
1286                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1287                        class_exp2obd(lock->l_export)->obd_name);
1288                 RETURN(-EINVAL);
1289         }
1290
1291         if (flag == LDLM_CB_CANCELING &&
1292             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1293             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1294              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1295               ldlm_is_cbpending(lock))) &&
1296             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1297              lock->l_resource->lr_type == LDLM_EXTENT)) {
1298                 __u64 start = 0;
1299                 __u64 end = OBD_OBJECT_EOF;
1300
1301                 rc = lu_env_init(&env, LCT_DT_THREAD);
1302                 if (unlikely(rc != 0))
1303                         GOTO(err, rc);
1304
1305                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1306                                    tgt->lut_lsd.lsd_osd_index);
1307
1308                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1309                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1310                         if (IS_ERR(obj))
1311                                 GOTO(err_env, rc = PTR_ERR(obj));
1312
1313                         if (!dt_object_exists(obj))
1314                                 GOTO(err_put, rc = -ENOENT);
1315
1316                         start = lock->l_policy_data.l_extent.start;
1317                         end = lock->l_policy_data.l_extent.end;
1318                 }
1319
1320                 rc = tgt_sync(&env, tgt, obj, start, end);
1321                 if (rc < 0) {
1322                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1323                                "cancel: rc = %d\n",
1324                                tgt_name(tgt), PFID(&fid),
1325                                lock->l_policy_data.l_extent.start,
1326                                lock->l_policy_data.l_extent.end, rc);
1327                 }
1328 err_put:
1329                 if (obj != NULL)
1330                         dt_object_put(&env, obj);
1331 err_env:
1332                 lu_env_fini(&env);
1333         }
1334 err:
1335         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1336         RETURN(rc);
1337 }
1338
1339 static struct ldlm_callback_suite tgt_dlm_cbs = {
1340         .lcs_completion = ldlm_server_completion_ast,
1341         .lcs_blocking   = tgt_blocking_ast,
1342         .lcs_glimpse    = ldlm_server_glimpse_ast
1343 };
1344
1345 int tgt_enqueue(struct tgt_session_info *tsi)
1346 {
1347         struct ptlrpc_request *req = tgt_ses_req(tsi);
1348         int rc;
1349
1350         ENTRY;
1351         /*
1352          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1353          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1354          */
1355         LASSERT(tsi->tsi_dlm_req != NULL);
1356         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1357                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1358         if (rc)
1359                 RETURN(err_serious(rc));
1360
1361         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1362         case LUT_FAIL_MDT:
1363                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1364                 break;
1365         case LUT_FAIL_OST:
1366                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1367                 break;
1368         case LUT_FAIL_MGT:
1369                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1370                 break;
1371         default:
1372                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1373                 break;
1374         }
1375         RETURN(req->rq_status);
1376 }
1377 EXPORT_SYMBOL(tgt_enqueue);
1378
1379 int tgt_convert(struct tgt_session_info *tsi)
1380 {
1381         struct ptlrpc_request *req = tgt_ses_req(tsi);
1382         int rc;
1383
1384         ENTRY;
1385         LASSERT(tsi->tsi_dlm_req);
1386         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1387         if (rc)
1388                 RETURN(err_serious(rc));
1389
1390         RETURN(req->rq_status);
1391 }
1392
1393 int tgt_bl_callback(struct tgt_session_info *tsi)
1394 {
1395         return err_serious(-EOPNOTSUPP);
1396 }
1397
1398 int tgt_cp_callback(struct tgt_session_info *tsi)
1399 {
1400         return err_serious(-EOPNOTSUPP);
1401 }
1402
1403 /* generic LDLM target handler */
1404 struct tgt_handler tgt_dlm_handlers[] = {
1405 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1406 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1407 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1408 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1409 };
1410 EXPORT_SYMBOL(tgt_dlm_handlers);
1411
1412 /*
1413  * Unified target LLOG handlers.
1414  */
1415 int tgt_llog_open(struct tgt_session_info *tsi)
1416 {
1417         int rc;
1418
1419         ENTRY;
1420
1421         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1422
1423         RETURN(rc);
1424 }
1425 EXPORT_SYMBOL(tgt_llog_open);
1426
1427 int tgt_llog_read_header(struct tgt_session_info *tsi)
1428 {
1429         int rc;
1430
1431         ENTRY;
1432
1433         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1434
1435         RETURN(rc);
1436 }
1437 EXPORT_SYMBOL(tgt_llog_read_header);
1438
1439 int tgt_llog_next_block(struct tgt_session_info *tsi)
1440 {
1441         int rc;
1442
1443         ENTRY;
1444
1445         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1446
1447         RETURN(rc);
1448 }
1449 EXPORT_SYMBOL(tgt_llog_next_block);
1450
1451 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1452 {
1453         int rc;
1454
1455         ENTRY;
1456
1457         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1458
1459         RETURN(rc);
1460 }
1461 EXPORT_SYMBOL(tgt_llog_prev_block);
1462
1463 /* generic llog target handler */
1464 struct tgt_handler tgt_llog_handlers[] = {
1465 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1466 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1467 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1468 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1469 };
1470 EXPORT_SYMBOL(tgt_llog_handlers);
1471
1472 /*
1473  * sec context handlers
1474  */
1475 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1476 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1477 {
1478         return 0;
1479 }
1480
1481 struct tgt_handler tgt_sec_ctx_handlers[] = {
1482 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1483 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1484 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1485 };
1486 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1487
1488 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1489                                  struct dt_device *key,
1490                                  struct lfsck_req_local *lrl,
1491                                  struct thandle *th) = NULL;
1492
1493 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1494                                                       struct dt_device *,
1495                                                       struct lfsck_req_local *,
1496                                                       struct thandle *))
1497 {
1498         tgt_lfsck_in_notify_local = notify;
1499 }
1500 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1501
1502 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1503                            struct dt_device *key,
1504                            struct lfsck_request *lr) = NULL;
1505
1506 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1507                                                 struct dt_device *,
1508                                                 struct lfsck_request *))
1509 {
1510         tgt_lfsck_in_notify = notify;
1511 }
1512 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1513
1514 static int (*tgt_lfsck_query)(const struct lu_env *env,
1515                               struct dt_device *key,
1516                               struct lfsck_request *req,
1517                               struct lfsck_reply *rep,
1518                               struct lfsck_query *que) = NULL;
1519
1520 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1521                                            struct dt_device *,
1522                                            struct lfsck_request *,
1523                                            struct lfsck_reply *,
1524                                            struct lfsck_query *))
1525 {
1526         tgt_lfsck_query = query;
1527 }
1528 EXPORT_SYMBOL(tgt_register_lfsck_query);
1529
1530 /* LFSCK request handlers */
1531 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1532 {
1533         const struct lu_env     *env = tsi->tsi_env;
1534         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1535         struct lfsck_request    *lr;
1536         int                      rc;
1537         ENTRY;
1538
1539         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1540         if (lr == NULL)
1541                 RETURN(-EPROTO);
1542
1543         rc = tgt_lfsck_in_notify(env, key, lr);
1544
1545         RETURN(rc);
1546 }
1547
1548 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1549 {
1550         struct lfsck_request    *request;
1551         struct lfsck_reply      *reply;
1552         int                      rc;
1553         ENTRY;
1554
1555         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1556         if (request == NULL)
1557                 RETURN(-EPROTO);
1558
1559         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1560         if (reply == NULL)
1561                 RETURN(-ENOMEM);
1562
1563         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1564                              request, reply, NULL);
1565
1566         RETURN(rc < 0 ? rc : 0);
1567 }
1568
1569 struct tgt_handler tgt_lfsck_handlers[] = {
1570 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1571 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1572 };
1573 EXPORT_SYMBOL(tgt_lfsck_handlers);
1574
1575 /*
1576  * initialize per-thread page pool (bug 5137).
1577  */
1578 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1579 {
1580         struct tgt_thread_big_cache *tbc;
1581
1582         ENTRY;
1583
1584         LASSERT(thread != NULL);
1585         LASSERT(thread->t_data == NULL);
1586
1587         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1588         if (tbc == NULL)
1589                 RETURN(-ENOMEM);
1590         thread->t_data = tbc;
1591         RETURN(0);
1592 }
1593 EXPORT_SYMBOL(tgt_io_thread_init);
1594
1595 /*
1596  * free per-thread pool created by tgt_thread_init().
1597  */
1598 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1599 {
1600         struct tgt_thread_big_cache *tbc;
1601
1602         ENTRY;
1603
1604         LASSERT(thread != NULL);
1605
1606         /*
1607          * be prepared to handle partially-initialized pools (because this is
1608          * called from ost_io_thread_init() for cleanup.
1609          */
1610         tbc = thread->t_data;
1611         if (tbc != NULL) {
1612                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1613                 thread->t_data = NULL;
1614         }
1615         EXIT;
1616 }
1617 EXPORT_SYMBOL(tgt_io_thread_done);
1618
1619 /**
1620  * Helper function for getting Data-on-MDT file server DLM lock
1621  * if asked by client.
1622  */
1623 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1624                       struct lustre_handle *lh, int mode, __u64 *flags)
1625 {
1626         union ldlm_policy_data policy = {
1627                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1628         };
1629         int rc;
1630
1631         ENTRY;
1632
1633         LASSERT(lh != NULL);
1634         LASSERT(ns != NULL);
1635         LASSERT(!lustre_handle_is_used(lh));
1636
1637         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1638                                     flags, ldlm_blocking_ast,
1639                                     ldlm_completion_ast, ldlm_glimpse_ast,
1640                                     NULL, 0, LVB_T_NONE, NULL, lh);
1641
1642         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1643 }
1644 EXPORT_SYMBOL(tgt_mdt_data_lock);
1645
1646 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1647 {
1648         LASSERT(lustre_handle_is_used(lh));
1649         ldlm_lock_decref(lh, mode);
1650 }
1651 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1652
1653 /**
1654  * Helper function for getting server side [start, start+count] DLM lock
1655  * if asked by client.
1656  */
1657 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1658                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1659                     struct lustre_handle *lh, int mode, __u64 *flags)
1660 {
1661         union ldlm_policy_data policy;
1662         int rc;
1663
1664         ENTRY;
1665
1666         LASSERT(lh != NULL);
1667         LASSERT(ns != NULL);
1668         LASSERT(!lustre_handle_is_used(lh));
1669
1670         policy.l_extent.gid = 0;
1671         policy.l_extent.start = start & PAGE_MASK;
1672
1673         /*
1674          * If ->o_blocks is EOF it means "lock till the end of the file".
1675          * Otherwise, it's size of an extent or hole being punched (in bytes).
1676          */
1677         if (end == OBD_OBJECT_EOF || end < start)
1678                 policy.l_extent.end = OBD_OBJECT_EOF;
1679         else
1680                 policy.l_extent.end = end | ~PAGE_MASK;
1681
1682         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1683                                     mode, flags, ldlm_blocking_ast,
1684                                     ldlm_completion_ast, ldlm_glimpse_ast,
1685                                     NULL, 0, LVB_T_NONE, NULL, lh);
1686         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1687 }
1688 EXPORT_SYMBOL(tgt_extent_lock);
1689
1690 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1691 {
1692         LASSERT(lustre_handle_is_used(lh));
1693         ldlm_lock_decref(lh, mode);
1694 }
1695 EXPORT_SYMBOL(tgt_extent_unlock);
1696
1697 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1698                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1699                         struct niobuf_remote *nb, struct lustre_handle *lh,
1700                         enum ldlm_mode mode)
1701 {
1702         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1703         __u64                    flags = 0;
1704         int                      nrbufs = obj->ioo_bufcnt;
1705         int                      i;
1706         int                      rc;
1707
1708         ENTRY;
1709
1710         LASSERT(mode == LCK_PR || mode == LCK_PW);
1711         LASSERT(!lustre_handle_is_used(lh));
1712
1713         if (ns->ns_obd->obd_recovering)
1714                 RETURN(0);
1715
1716         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1717                 RETURN(0);
1718
1719         for (i = 1; i < nrbufs; i++)
1720                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1721                         RETURN(-EFAULT);
1722
1723         /* MDT IO for data-on-mdt */
1724         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1725                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1726         else
1727                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1728                                      nb[nrbufs - 1].rnb_offset +
1729                                      nb[nrbufs - 1].rnb_len - 1,
1730                                      lh, mode, &flags);
1731         RETURN(rc);
1732 }
1733
1734 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1735                            struct lustre_handle *lh, enum ldlm_mode mode)
1736 {
1737         ENTRY;
1738
1739         LASSERT(mode == LCK_PR || mode == LCK_PW);
1740         LASSERT((obj->ioo_bufcnt > 0 &&
1741                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1742                 lustre_handle_is_used(lh));
1743
1744         if (lustre_handle_is_used(lh))
1745                 tgt_extent_unlock(lh, mode);
1746         EXIT;
1747 }
1748 static int tgt_checksum_niobuf(struct lu_target *tgt,
1749                                  struct niobuf_local *local_nb, int npages,
1750                                  int opc, enum cksum_types cksum_type,
1751                                  __u32 *cksum)
1752 {
1753         struct ahash_request           *req;
1754         unsigned int                    bufsize;
1755         int                             i, err;
1756         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1757
1758         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1759         if (IS_ERR(req)) {
1760                 CERROR("%s: unable to initialize checksum hash %s\n",
1761                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1762                 return PTR_ERR(req);
1763         }
1764
1765         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1766         for (i = 0; i < npages; i++) {
1767                 /* corrupt the data before we compute the checksum, to
1768                  * simulate a client->OST data error */
1769                 if (i == 0 && opc == OST_WRITE &&
1770                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1771                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1772                         int len = local_nb[i].lnb_len;
1773                         struct page *np = tgt_page_to_corrupt;
1774
1775                         if (np) {
1776                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1777                                 char *ptr2 = page_address(np);
1778
1779                                 memcpy(ptr2 + off, ptr + off, len);
1780                                 memcpy(ptr2 + off, "bad3", min(4, len));
1781                                 kunmap_atomic(ptr);
1782
1783                                 /* LU-8376 to preserve original index for
1784                                  * display in dump_all_bulk_pages() */
1785                                 np->index = i;
1786
1787                                 cfs_crypto_hash_update_page(req, np, off,
1788                                                             len);
1789                                 continue;
1790                         } else {
1791                                 CERROR("%s: can't alloc page for corruption\n",
1792                                        tgt_name(tgt));
1793                         }
1794                 }
1795                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1796                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1797                                   local_nb[i].lnb_len);
1798
1799                  /* corrupt the data after we compute the checksum, to
1800                  * simulate an OST->client data error */
1801                 if (i == 0 && opc == OST_READ &&
1802                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1803                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1804                         int len = local_nb[i].lnb_len;
1805                         struct page *np = tgt_page_to_corrupt;
1806
1807                         if (np) {
1808                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1809                                 char *ptr2 = page_address(np);
1810
1811                                 memcpy(ptr2 + off, ptr + off, len);
1812                                 memcpy(ptr2 + off, "bad4", min(4, len));
1813                                 kunmap_atomic(ptr);
1814
1815                                 /* LU-8376 to preserve original index for
1816                                  * display in dump_all_bulk_pages() */
1817                                 np->index = i;
1818
1819                                 cfs_crypto_hash_update_page(req, np, off,
1820                                                             len);
1821                                 continue;
1822                         } else {
1823                                 CERROR("%s: can't alloc page for corruption\n",
1824                                        tgt_name(tgt));
1825                         }
1826                 }
1827         }
1828
1829         bufsize = sizeof(*cksum);
1830         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1831
1832         return 0;
1833 }
1834
1835 char dbgcksum_file_name[PATH_MAX];
1836
1837 static void dump_all_bulk_pages(struct obdo *oa, int count,
1838                                 struct niobuf_local *local_nb,
1839                                 __u32 server_cksum, __u32 client_cksum)
1840 {
1841         struct file *filp;
1842         int rc, i;
1843         unsigned int len;
1844         char *buf;
1845
1846         /* will only keep dump of pages on first error for the same range in
1847          * file/fid, not during the resends/retries. */
1848         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1849                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1850                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1851                   libcfs_debug_file_path_arr :
1852                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1853                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1854                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1855                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1856                  local_nb[0].lnb_file_offset,
1857                  local_nb[count-1].lnb_file_offset +
1858                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1859         filp = filp_open(dbgcksum_file_name,
1860                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1861         if (IS_ERR(filp)) {
1862                 rc = PTR_ERR(filp);
1863                 if (rc == -EEXIST)
1864                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1865                                "checksum error: rc = %d\n", dbgcksum_file_name,
1866                                rc);
1867                 else
1868                         CERROR("%s: can't open to dump pages with checksum "
1869                                "error: rc = %d\n", dbgcksum_file_name, rc);
1870                 return;
1871         }
1872
1873         for (i = 0; i < count; i++) {
1874                 len = local_nb[i].lnb_len;
1875                 buf = kmap(local_nb[i].lnb_page);
1876                 while (len != 0) {
1877                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1878                         if (rc < 0) {
1879                                 CERROR("%s: wanted to write %u but got %d "
1880                                        "error\n", dbgcksum_file_name, len, rc);
1881                                 break;
1882                         }
1883                         len -= rc;
1884                         buf += rc;
1885                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1886                                dbgcksum_file_name, rc);
1887                 }
1888                 kunmap(local_nb[i].lnb_page);
1889         }
1890
1891         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1892         if (rc)
1893                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1894         filp_close(filp, NULL);
1895 }
1896
1897 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1898                                struct obd_export *exp, struct obdo *oa,
1899                                const struct lnet_process_id *peer,
1900                                __u32 client_cksum, __u32 server_cksum,
1901                                enum cksum_types server_cksum_type)
1902 {
1903         char *msg;
1904         enum cksum_types cksum_type;
1905         loff_t start, end;
1906
1907         /* unlikely to happen and only if resend does not occur due to cksum
1908          * control failure on Client */
1909         if (unlikely(server_cksum == client_cksum)) {
1910                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1911                        client_cksum);
1912                 return 0;
1913         }
1914
1915         if (exp->exp_obd->obd_checksum_dump)
1916                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1917                                     client_cksum);
1918
1919         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1920                                            oa->o_flags : 0);
1921
1922         if (cksum_type != server_cksum_type)
1923                 msg = "the server may have not used the checksum type specified"
1924                       " in the original request - likely a protocol problem";
1925         else
1926                 msg = "should have changed on the client or in transit";
1927
1928         start = local_nb[0].lnb_file_offset;
1929         end = local_nb[npages-1].lnb_file_offset +
1930                                         local_nb[npages-1].lnb_len - 1;
1931
1932         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1933                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1934                 " %x (type %x), server csum %x (type %x)\n",
1935                 exp->exp_obd->obd_name,
1936                 msg, libcfs_nid2str(peer->nid),
1937                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1938                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1939                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1940                 POSTID(&oa->o_oi),
1941                 start, end, client_cksum, cksum_type, server_cksum,
1942                 server_cksum_type);
1943
1944         return 1;
1945 }
1946
1947 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1948                              unsigned char *buf, int size)
1949 {
1950         int     i, off, len, copied = size;
1951         char    *ptr;
1952
1953         for (i = 0; i < npages; i++) {
1954                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1955                 len = local[i].lnb_len;
1956
1957                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1958                        i, off, len, size);
1959                 if (len > size)
1960                         return -EINVAL;
1961
1962                 ptr = kmap_atomic(local[i].lnb_page);
1963                 memcpy(buf + off, ptr, len);
1964                 kunmap_atomic(ptr);
1965                 buf += len;
1966                 size -= len;
1967         }
1968         return copied - size;
1969 }
1970
1971 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1972                                      struct niobuf_local *local_nb,
1973                                      int npages, int opc,
1974                                      obd_dif_csum_fn *fn,
1975                                      int sector_size,
1976                                      u32 *check_sum)
1977 {
1978         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1979         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1980         const char *obd_name = tgt->lut_obd->obd_name;
1981         struct ahash_request *req;
1982         unsigned int bufsize;
1983         unsigned char *buffer;
1984         struct page *__page;
1985         __u16 *guard_start;
1986         int guard_number;
1987         int used_number = 0;
1988         __u32 cksum;
1989         int rc = 0;
1990         int used;
1991         int i;
1992
1993         __page = alloc_page(GFP_KERNEL);
1994         if (__page == NULL)
1995                 return -ENOMEM;
1996
1997         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1998         if (IS_ERR(req)) {
1999                 CERROR("%s: unable to initialize checksum hash %s\n",
2000                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2001                 return PTR_ERR(req);
2002         }
2003
2004         buffer = kmap(__page);
2005         guard_start = (__u16 *)buffer;
2006         guard_number = PAGE_SIZE / sizeof(*guard_start);
2007         for (i = 0; i < npages; i++) {
2008                 /* corrupt the data before we compute the checksum, to
2009                  * simulate a client->OST data error */
2010                 if (i == 0 && opc == OST_WRITE &&
2011                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2012                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2013                         int len = local_nb[i].lnb_len;
2014                         struct page *np = tgt_page_to_corrupt;
2015
2016                         if (np) {
2017                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2018                                 char *ptr2 = page_address(np);
2019
2020                                 memcpy(ptr2 + off, ptr + off, len);
2021                                 memcpy(ptr2 + off, "bad3", min(4, len));
2022                                 kunmap_atomic(ptr);
2023
2024                                 /* LU-8376 to preserve original index for
2025                                  * display in dump_all_bulk_pages() */
2026                                 np->index = i;
2027
2028                                 cfs_crypto_hash_update_page(req, np, off,
2029                                                             len);
2030                                 continue;
2031                         } else {
2032                                 CERROR("%s: can't alloc page for corruption\n",
2033                                        tgt_name(tgt));
2034                         }
2035                 }
2036
2037                 /*
2038                  * The left guard number should be able to hold checksums of a
2039                  * whole page
2040                  */
2041                 if (t10_cksum_type && opc == OST_READ &&
2042                     local_nb[i].lnb_guard_disk) {
2043                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2044                         if (used > (guard_number - used_number)) {
2045                                 rc = -E2BIG;
2046                                 break;
2047                         }
2048                         memcpy(guard_start + used_number,
2049                                local_nb[i].lnb_guards,
2050                                used * sizeof(*local_nb[i].lnb_guards));
2051                 } else {
2052                         rc = obd_page_dif_generate_buffer(obd_name,
2053                                 local_nb[i].lnb_page,
2054                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2055                                 local_nb[i].lnb_len, guard_start + used_number,
2056                                 guard_number - used_number, &used, sector_size,
2057                                 fn);
2058                         if (rc)
2059                                 break;
2060                 }
2061
2062                 LASSERT(used <= MAX_GUARD_NUMBER);
2063                 /*
2064                  * If disk support T10PI checksum, copy guards to local_nb.
2065                  * If the write is partial page, do not use the guards for bio
2066                  * submission since the data might not be full-sector. The bio
2067                  * guards will be generated later based on the full sectors. If
2068                  * the sector size is 512B rather than 4 KB, or the page size
2069                  * is larger than 4KB, this might drop some useful guards for
2070                  * partial page write, but it will only add minimal extra time
2071                  * of checksum calculation.
2072                  */
2073                 if (t10_cksum_type && opc == OST_WRITE &&
2074                     local_nb[i].lnb_len == PAGE_SIZE) {
2075                         local_nb[i].lnb_guard_rpc = 1;
2076                         memcpy(local_nb[i].lnb_guards,
2077                                guard_start + used_number,
2078                                used * sizeof(*local_nb[i].lnb_guards));
2079                 }
2080
2081                 used_number += used;
2082                 if (used_number == guard_number) {
2083                         cfs_crypto_hash_update_page(req, __page, 0,
2084                                 used_number * sizeof(*guard_start));
2085                         used_number = 0;
2086                 }
2087
2088                  /* corrupt the data after we compute the checksum, to
2089                  * simulate an OST->client data error */
2090                 if (unlikely(i == 0 && opc == OST_READ &&
2091                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2092                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2093                         int len = local_nb[i].lnb_len;
2094                         struct page *np = tgt_page_to_corrupt;
2095
2096                         if (np) {
2097                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2098                                 char *ptr2 = page_address(np);
2099
2100                                 memcpy(ptr2 + off, ptr + off, len);
2101                                 memcpy(ptr2 + off, "bad4", min(4, len));
2102                                 kunmap_atomic(ptr);
2103
2104                                 /* LU-8376 to preserve original index for
2105                                  * display in dump_all_bulk_pages() */
2106                                 np->index = i;
2107
2108                                 cfs_crypto_hash_update_page(req, np, off,
2109                                                             len);
2110                                 continue;
2111                         } else {
2112                                 CERROR("%s: can't alloc page for corruption\n",
2113                                        tgt_name(tgt));
2114                         }
2115                 }
2116         }
2117         kunmap(__page);
2118         if (rc)
2119                 GOTO(out, rc);
2120
2121         if (used_number != 0)
2122                 cfs_crypto_hash_update_page(req, __page, 0,
2123                         used_number * sizeof(*guard_start));
2124
2125         bufsize = sizeof(cksum);
2126         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2127
2128         if (rc == 0)
2129                 *check_sum = cksum;
2130 out:
2131         __free_page(__page);
2132         return rc;
2133 }
2134
2135 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2136                                   enum cksum_types cksum_type,
2137                                   struct niobuf_local *local_nb,
2138                                   int npages, int opc, u32 *check_sum)
2139 {
2140         obd_dif_csum_fn *fn = NULL;
2141         int sector_size = 0;
2142         int rc;
2143
2144         ENTRY;
2145         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2146
2147         if (fn)
2148                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2149                                                opc, fn, sector_size,
2150                                                check_sum);
2151         else
2152                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2153                                          cksum_type, check_sum);
2154         RETURN(rc);
2155 }
2156
2157 int tgt_brw_read(struct tgt_session_info *tsi)
2158 {
2159         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2160         struct ptlrpc_bulk_desc *desc = NULL;
2161         struct obd_export       *exp = tsi->tsi_exp;
2162         struct niobuf_remote    *remote_nb;
2163         struct niobuf_local     *local_nb;
2164         struct obd_ioobj        *ioo;
2165         struct ost_body         *body, *repbody;
2166         struct lustre_handle     lockh = { 0 };
2167         int                      npages, nob = 0, rc, i, no_reply = 0,
2168                                  npages_read;
2169         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2170         const char *obd_name = exp->exp_obd->obd_name;
2171
2172         ENTRY;
2173
2174         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2175             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2176                 CERROR("%s: deny read request from %s to portal %u\n",
2177                        tgt_name(tsi->tsi_tgt),
2178                        obd_export_nid2str(req->rq_export),
2179                        ptlrpc_req2svc(req)->srv_req_portal);
2180                 RETURN(-EPROTO);
2181         }
2182
2183         req->rq_bulk_read = 1;
2184
2185         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2186                 RETURN(-EIO);
2187
2188         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2189                          cfs_fail_val : (obd_timeout + 1) / 4);
2190
2191         /* Check if there is eviction in progress, and if so, wait for it to
2192          * finish */
2193         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2194                 /* We do not care how long it takes */
2195                 wait_event_idle(
2196                         exp->exp_obd->obd_evict_inprogress_waitq,
2197                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2198         }
2199
2200         /* There must be big cache in current thread to process this request
2201          * if it is NULL then something went wrong and it wasn't allocated,
2202          * report -ENOMEM in that case */
2203         if (tbc == NULL)
2204                 RETURN(-ENOMEM);
2205
2206         body = tsi->tsi_ost_body;
2207         LASSERT(body != NULL);
2208
2209         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2210         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2211
2212         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2213         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2214
2215         local_nb = tbc->local;
2216
2217         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2218                           &lockh, LCK_PR);
2219         if (rc != 0)
2220                 RETURN(rc);
2221
2222         /*
2223          * If getting the lock took more time than
2224          * client was willing to wait, drop it. b=11330
2225          */
2226         if (ktime_get_real_seconds() > req->rq_deadline ||
2227             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2228                 no_reply = 1;
2229                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2230                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2231                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2232                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2233                 GOTO(out_lock, rc = -ETIMEDOUT);
2234         }
2235
2236         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2237         repbody->oa = body->oa;
2238
2239         npages = PTLRPC_MAX_BRW_PAGES;
2240         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2241                         ioo, remote_nb, &npages, local_nb);
2242         if (rc != 0)
2243                 GOTO(out_lock, rc);
2244
2245         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2246             body->oa.o_flags & OBD_FL_SHORT_IO) {
2247                 desc = NULL;
2248         } else {
2249                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2250                                             PTLRPC_BULK_PUT_SOURCE,
2251                                             OST_BULK_PORTAL,
2252                                             &ptlrpc_bulk_kiov_nopin_ops);
2253                 if (desc == NULL)
2254                         GOTO(out_commitrw, rc = -ENOMEM);
2255         }
2256
2257         nob = 0;
2258         npages_read = npages;
2259         for (i = 0; i < npages; i++) {
2260                 int page_rc = local_nb[i].lnb_rc;
2261
2262                 if (page_rc < 0) {
2263                         rc = page_rc;
2264                         npages_read = i;
2265                         break;
2266                 }
2267
2268                 nob += page_rc;
2269                 if (page_rc != 0 && desc != NULL) { /* some data! */
2270                         LASSERT(local_nb[i].lnb_page != NULL);
2271                         desc->bd_frag_ops->add_kiov_frag
2272                           (desc, local_nb[i].lnb_page,
2273                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2274                            page_rc);
2275                 }
2276
2277                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2278                         local_nb[i].lnb_len = page_rc;
2279                         npages_read = i + (page_rc != 0 ? 1 : 0);
2280                         /* All subsequent pages should be 0 */
2281                         while (++i < npages)
2282                                 LASSERT(local_nb[i].lnb_rc == 0);
2283                         break;
2284                 }
2285         }
2286
2287         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2288                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2289                            body->oa.o_flags : 0;
2290                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2291
2292                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2293                                                           cksum_type);
2294                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2295
2296                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2297                                             local_nb, npages_read, OST_READ,
2298                                             &repbody->oa.o_cksum);
2299                 if (rc < 0)
2300                         GOTO(out_commitrw, rc);
2301                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2302                        repbody->oa.o_cksum);
2303
2304                 /* if a resend it could be for a cksum error, so check Server
2305                  * cksum with returned Client cksum (this should even cover
2306                  * zero-cksum case) */
2307                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2308                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2309                         check_read_checksum(local_nb, npages_read, exp,
2310                                             &body->oa, &req->rq_peer,
2311                                             body->oa.o_cksum,
2312                                             repbody->oa.o_cksum, cksum_type);
2313         } else {
2314                 repbody->oa.o_valid = 0;
2315         }
2316         /* We're finishing using body->oa as an input variable */
2317
2318         /* Check if client was evicted while we were doing i/o before touching
2319          * network */
2320         if (rc == 0) {
2321                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2322                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2323                         unsigned char *short_io_buf;
2324                         int short_io_size;
2325
2326                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2327                                                               &RMF_SHORT_IO);
2328                         short_io_size = req_capsule_get_size(&req->rq_pill,
2329                                                              &RMF_SHORT_IO,
2330                                                              RCL_SERVER);
2331                         rc = tgt_pages2shortio(local_nb, npages_read,
2332                                                short_io_buf, short_io_size);
2333                         if (rc >= 0)
2334                                 req_capsule_shrink(&req->rq_pill,
2335                                                    &RMF_SHORT_IO, rc,
2336                                                    RCL_SERVER);
2337                         rc = rc > 0 ? 0 : rc;
2338                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2339                         rc = target_bulk_io(exp, desc);
2340                 }
2341                 no_reply = rc != 0;
2342         } else {
2343                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2344                     body->oa.o_flags & OBD_FL_SHORT_IO)
2345                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2346                                            RCL_SERVER);
2347         }
2348
2349 out_commitrw:
2350         /* Must commit after prep above in all cases */
2351         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2352                           remote_nb, npages, local_nb, rc);
2353 out_lock:
2354         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2355
2356         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2357                 ptlrpc_free_bulk(desc);
2358
2359         LASSERT(rc <= 0);
2360         if (rc == 0) {
2361                 rc = nob;
2362                 ptlrpc_lprocfs_brw(req, nob);
2363         } else if (no_reply) {
2364                 req->rq_no_reply = 1;
2365                 /* reply out callback would free */
2366                 ptlrpc_req_drop_rs(req);
2367                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2368                               "client will retry: rc %d\n",
2369                               obd_name,
2370                               obd_uuid2str(&exp->exp_client_uuid),
2371                               obd_export_nid2str(exp), rc);
2372         }
2373         /* send a bulk after reply to simulate a network delay or reordering
2374          * by a router - Note that !desc implies short io, so there is no bulk
2375          * to reorder. */
2376         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2377             desc) {
2378                 CDEBUG(D_INFO, "reorder BULK\n");
2379
2380                 ssleep(3);
2381                 target_bulk_io(exp, desc);
2382                 ptlrpc_free_bulk(desc);
2383         }
2384
2385         RETURN(rc);
2386 }
2387 EXPORT_SYMBOL(tgt_brw_read);
2388
2389 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2390                              unsigned char *buf, unsigned int size)
2391 {
2392         int     i, off, len;
2393         char    *ptr;
2394
2395         for (i = 0; i < npages; i++) {
2396                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2397                 len = local[i].lnb_len;
2398
2399                 if (len == 0)
2400                         continue;
2401
2402                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2403                        i, off, len, size);
2404                 ptr = kmap_atomic(local[i].lnb_page);
2405                 if (ptr == NULL)
2406                         return -EINVAL;
2407                 memcpy(ptr + off, buf, len < size ? len : size);
2408                 kunmap_atomic(ptr);
2409                 buf += len;
2410                 size -= len;
2411         }
2412         return 0;
2413 }
2414
2415 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2416                               struct ptlrpc_bulk_desc *desc,
2417                               struct niobuf_local *local_nb, int npages,
2418                               u32 client_cksum, u32 server_cksum,
2419                               bool mmap)
2420 {
2421         struct obd_export *exp = req->rq_export;
2422         struct ost_body *body;
2423         char *router = "";
2424         char *via = "";
2425
2426         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2427         LASSERT(body != NULL);
2428
2429         if (desc && req->rq_peer.nid != desc->bd_sender) {
2430                 via = " via ";
2431                 router = libcfs_nid2str(desc->bd_sender);
2432         }
2433
2434         if (exp->exp_obd->obd_checksum_dump)
2435                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2436                                     client_cksum);
2437
2438         if (mmap) {
2439                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2440                              client_cksum, server_cksum);
2441                 return;
2442         }
2443
2444         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2445                            DFID" object "DOSTID" extent [%llu-%llu"
2446                            "]: client csum %x, server csum %x\n",
2447                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2448                            via, router,
2449                            body->oa.o_valid & OBD_MD_FLFID ?
2450                            body->oa.o_parent_seq : (__u64)0,
2451                            body->oa.o_valid & OBD_MD_FLFID ?
2452                            body->oa.o_parent_oid : 0,
2453                            body->oa.o_valid & OBD_MD_FLFID ?
2454                            body->oa.o_parent_ver : 0,
2455                            POSTID(&body->oa.o_oi),
2456                            local_nb[0].lnb_file_offset,
2457                            local_nb[npages-1].lnb_file_offset +
2458                            local_nb[npages - 1].lnb_len - 1,
2459                            client_cksum, server_cksum);
2460 }
2461
2462 int tgt_brw_write(struct tgt_session_info *tsi)
2463 {
2464         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2465         struct ptlrpc_bulk_desc *desc = NULL;
2466         struct obd_export       *exp = req->rq_export;
2467         struct niobuf_remote    *remote_nb;
2468         struct niobuf_local     *local_nb;
2469         struct obd_ioobj        *ioo;
2470         struct ost_body         *body, *repbody;
2471         struct lustre_handle     lockh = {0};
2472         __u32                   *rcs;
2473         int                      objcount, niocount, npages;
2474         int                      rc, i, j;
2475         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2476         bool                     no_reply = false, mmap;
2477         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2478         bool wait_sync = false;
2479         const char *obd_name = exp->exp_obd->obd_name;
2480
2481         ENTRY;
2482
2483         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2484             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2485                 CERROR("%s: deny write request from %s to portal %u\n",
2486                        tgt_name(tsi->tsi_tgt),
2487                        obd_export_nid2str(req->rq_export),
2488                        ptlrpc_req2svc(req)->srv_req_portal);
2489                 RETURN(err_serious(-EPROTO));
2490         }
2491
2492         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2493                 RETURN(err_serious(-ENOSPC));
2494         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2495                 RETURN(err_serious(-EROFS));
2496
2497         req->rq_bulk_write = 1;
2498
2499         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2500                 RETURN(err_serious(-EIO));
2501         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2502                 RETURN(err_serious(-EFAULT));
2503
2504         /* pause before transaction has been started */
2505         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2506                          cfs_fail_val : (obd_timeout + 1) / 4);
2507
2508         /* Delay write commit to show stale size information */
2509         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2510
2511         /* There must be big cache in current thread to process this request
2512          * if it is NULL then something went wrong and it wasn't allocated,
2513          * report -ENOMEM in that case */
2514         if (tbc == NULL)
2515                 RETURN(-ENOMEM);
2516
2517         body = tsi->tsi_ost_body;
2518         LASSERT(body != NULL);
2519
2520         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2521         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2522
2523         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2524                                         RCL_CLIENT) / sizeof(*ioo);
2525
2526         for (niocount = i = 0; i < objcount; i++)
2527                 niocount += ioo[i].ioo_bufcnt;
2528
2529         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2530         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2531         if (niocount != req_capsule_get_size(&req->rq_pill,
2532                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2533                         sizeof(*remote_nb))
2534                 RETURN(err_serious(-EPROTO));
2535
2536         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2537             ptlrpc_connection_is_local(exp->exp_connection))
2538                 memory_pressure_set();
2539
2540         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2541                              niocount * sizeof(*rcs));
2542         rc = req_capsule_server_pack(&req->rq_pill);
2543         if (rc != 0)
2544                 GOTO(out, rc = err_serious(rc));
2545
2546         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2547         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2548
2549         local_nb = tbc->local;
2550
2551         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2552                           &lockh, LCK_PW);
2553         if (rc != 0)
2554                 GOTO(out, rc);
2555
2556         /*
2557          * If getting the lock took more time than
2558          * client was willing to wait, drop it. b=11330
2559          */
2560         if (ktime_get_real_seconds() > req->rq_deadline ||
2561             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2562                 no_reply = true;
2563                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2564                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2565                        POSTID(&ioo->ioo_oid),
2566                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2567                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2568                 GOTO(out_lock, rc = -ETIMEDOUT);
2569         }
2570
2571         /* Because we already sync grant info with client when reconnect,
2572          * grant info will be cleared for resent req, then fed_grant and
2573          * total_grant will not be modified in following preprw_write */
2574         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2575                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2576                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2577         }
2578
2579         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2580         if (repbody == NULL)
2581                 GOTO(out_lock, rc = -ENOMEM);
2582         repbody->oa = body->oa;
2583
2584         npages = PTLRPC_MAX_BRW_PAGES;
2585         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2586                         objcount, ioo, remote_nb, &npages, local_nb);
2587         if (rc < 0)
2588                 GOTO(out_lock, rc);
2589         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2590             body->oa.o_flags & OBD_FL_SHORT_IO) {
2591                 unsigned int short_io_size;
2592                 unsigned char *short_io_buf;
2593
2594                 short_io_size = req_capsule_get_size(&req->rq_pill,
2595                                                      &RMF_SHORT_IO,
2596                                                      RCL_CLIENT);
2597                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2598                                                       &RMF_SHORT_IO);
2599                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2600                                " size = %d\n", short_io_size);
2601
2602                 /* Copy short io buf to pages */
2603                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2604                                        short_io_size);
2605                 desc = NULL;
2606         } else {
2607                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2608                                             PTLRPC_BULK_GET_SINK,
2609                                             OST_BULK_PORTAL,
2610                                             &ptlrpc_bulk_kiov_nopin_ops);
2611                 if (desc == NULL)
2612                         GOTO(skip_transfer, rc = -ENOMEM);
2613
2614                 /* NB Having prepped, we must commit... */
2615                 for (i = 0; i < npages; i++)
2616                         desc->bd_frag_ops->add_kiov_frag(desc,
2617                                         local_nb[i].lnb_page,
2618                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2619                                         local_nb[i].lnb_len);
2620
2621                 rc = sptlrpc_svc_prep_bulk(req, desc);
2622                 if (rc != 0)
2623                         GOTO(skip_transfer, rc);
2624
2625                 rc = target_bulk_io(exp, desc);
2626         }
2627
2628         no_reply = rc != 0;
2629
2630 skip_transfer:
2631         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2632                 static int cksum_counter;
2633
2634                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2635                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2636
2637                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2638                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2639                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2640                                                            cksum_type);
2641
2642                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2643                                             local_nb, npages, OST_WRITE,
2644                                             &repbody->oa.o_cksum);
2645                 if (rc < 0)
2646                         GOTO(out_commitrw, rc);
2647
2648                 cksum_counter++;
2649
2650                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2651                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2652                                 body->oa.o_flags & OBD_FL_MMAP);
2653
2654                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2655                                           body->oa.o_cksum,
2656                                           repbody->oa.o_cksum, mmap);
2657                         cksum_counter = 0;
2658                 } else if ((cksum_counter & (-cksum_counter)) ==
2659                            cksum_counter) {
2660                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2661                                cksum_counter, libcfs_id2str(req->rq_peer),
2662                                repbody->oa.o_cksum);
2663                 }
2664         }
2665
2666 out_commitrw:
2667         /* Must commit after prep above in all cases */
2668         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2669                           objcount, ioo, remote_nb, npages, local_nb, rc);
2670         if (rc == -ENOTCONN)
2671                 /* quota acquire process has been given up because
2672                  * either the client has been evicted or the client
2673                  * has timed out the request already */
2674                 no_reply = true;
2675
2676         for (i = 0; i < niocount; i++) {
2677                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2678                         wait_sync = true;
2679                         break;
2680                 }
2681         }
2682         /*
2683          * Disable sending mtime back to the client. If the client locked the
2684          * whole object, then it has already updated the mtime on its side,
2685          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2686          */
2687         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2688
2689         if (rc == 0) {
2690                 int nob = 0;
2691
2692                 /* set per-requested niobuf return codes */
2693                 for (i = j = 0; i < niocount; i++) {
2694                         int len = remote_nb[i].rnb_len;
2695
2696                         nob += len;
2697                         rcs[i] = 0;
2698                         do {
2699                                 LASSERT(j < npages);
2700                                 if (local_nb[j].lnb_rc < 0)
2701                                         rcs[i] = local_nb[j].lnb_rc;
2702                                 len -= local_nb[j].lnb_len;
2703                                 j++;
2704                         } while (len > 0);
2705                         LASSERT(len == 0);
2706                 }
2707                 LASSERT(j == npages);
2708                 ptlrpc_lprocfs_brw(req, nob);
2709         }
2710 out_lock:
2711         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2712         if (desc)
2713                 ptlrpc_free_bulk(desc);
2714 out:
2715         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2716                 req->rq_no_reply = 1;
2717                 /* reply out callback would free */
2718                 ptlrpc_req_drop_rs(req);
2719                 if (!exp->exp_obd->obd_no_transno)
2720                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2721                                       " client will retry: rc = %d\n",
2722                                       obd_name,
2723                                       obd_uuid2str(&exp->exp_client_uuid),
2724                                       obd_export_nid2str(exp), rc);
2725         }
2726         memory_pressure_clr();
2727         RETURN(rc);
2728 }
2729 EXPORT_SYMBOL(tgt_brw_write);
2730
2731 /* Check if request can be reconstructed from saved reply data
2732  * A copy of the reply data is returned in @trd if the pointer is not NULL
2733  */
2734 int req_can_reconstruct(struct ptlrpc_request *req,
2735                          struct tg_reply_data *trd)
2736 {
2737         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2738         struct lsd_client_data *lcd = ted->ted_lcd;
2739         int found;
2740
2741         if (tgt_is_multimodrpcs_client(req->rq_export))
2742                 return tgt_lookup_reply(req, trd);
2743
2744         mutex_lock(&ted->ted_lcd_lock);
2745         found = req->rq_xid == lcd->lcd_last_xid ||
2746                 req->rq_xid == lcd->lcd_last_close_xid;
2747
2748         if (found && trd != NULL) {
2749                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2750                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2751                         trd->trd_reply.lrd_transno =
2752                                                 lcd->lcd_last_close_transno;
2753                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2754                 } else {
2755                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2756                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2757                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2758                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2759                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2760                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2761                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2762                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2763                 }
2764         }
2765         mutex_unlock(&ted->ted_lcd_lock);
2766
2767         return found;
2768 }
2769 EXPORT_SYMBOL(req_can_reconstruct);
2770