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                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 922                       == LOLND))
 923                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 924
 925                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 926                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 927                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 928                                "expect %x\n", tgt_name(tgt),
 929                                req->rq_flvr.sf_rpc,
 930                                libcfs_nid2str(req->rq_peer.nid),
 931                                exp->exp_flvr.sf_rpc);
 932                         rc = -EACCES;
 933                 }
 934                 spin_unlock(&exp->exp_lock);
 935         } else {
 936                 if (exp->exp_sp_peer != req->rq_sp_from) {
 937                         CERROR("%s: RPC source %s doesn't match %s\n",
 938                                tgt_name(tgt),
 939                                sptlrpc_part2name(req->rq_sp_from),
 940                                sptlrpc_part2name(exp->exp_sp_peer));
 941                         rc = -EACCES;
 942                 } else {
 943                         rc = sptlrpc_target_export_check(exp, req);
 944                 }
 945         }
 946
 947         return rc;
 948 }
 949
 950 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 951 {
 952         struct sptlrpc_rule_set  tmp_rset;
 953         int                      rc;
 954
 955         if (unlikely(tgt == NULL)) {
 956                 CERROR("No target passed");
 957                 return -EINVAL;
 958         }
 959
 960         sptlrpc_rule_set_init(&tmp_rset);
 961         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 962         if (rc) {
 963                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 964                        tgt_name(tgt), rc);
 965                 return rc;
 966         }
 967
 968         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 969
 970         write_lock(&tgt->lut_sptlrpc_lock);
 971         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 972         tgt->lut_sptlrpc_rset = tmp_rset;
 973         write_unlock(&tgt->lut_sptlrpc_lock);
 974
 975         return 0;
 976 }
 977 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 978
 979 int tgt_connect(struct tgt_session_info *tsi)
 980 {
 981         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 982         struct obd_connect_data *reply;
 983         int                      rc;
 984
 985         ENTRY;
 986
 987         /* XXX: better to call this check right after getting new export but
 988          * before last_rcvd slot allocation to avoid server load upon insecure
 989          * connects. This is to be fixed after unifiyng all targets.
 990          */
 991         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 992         if (rc)
 993                 GOTO(out, rc);
 994
 995         /* To avoid exposing partially initialized connection flags, changes up
 996          * to this point have been staged in reply->ocd_connect_flags. Now that
 997          * connection handling has completed successfully, atomically update
 998          * the connect flags in the shared export data structure. LU-1623 */
 999         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1000         spin_lock(&tsi->tsi_exp->exp_lock);
1001         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1002         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1003                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1004                         reply->ocd_connect_flags2;
1005         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1006         spin_unlock(&tsi->tsi_exp->exp_lock);
1007
1008         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1009                    LUSTRE_MDT_NAME) == 0) {
1010                 rc = req_check_sepol(tsi->tsi_pill);
1011                 if (rc)
1012                         GOTO(out, rc);
1013         }
1014
1015         RETURN(0);
1016 out:
1017         obd_disconnect(class_export_get(tsi->tsi_exp));
1018         return rc;
1019 }
1020 EXPORT_SYMBOL(tgt_connect);
1021
1022 int tgt_disconnect(struct tgt_session_info *tsi)
1023 {
1024         int rc;
1025
1026         ENTRY;
1027
1028         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1029
1030         rc = target_handle_disconnect(tgt_ses_req(tsi));
1031         if (rc)
1032                 RETURN(err_serious(rc));
1033
1034         RETURN(rc);
1035 }
1036 EXPORT_SYMBOL(tgt_disconnect);
1037
1038 /*
1039  * Unified target OBD handlers
1040  */
1041 int tgt_obd_ping(struct tgt_session_info *tsi)
1042 {
1043         int rc;
1044
1045         ENTRY;
1046
1047         /* The target-specific part of OBD_PING request handling.
1048          * It controls Filter Modification Data (FMD) expiration each time
1049          * PING is received.
1050          *
1051          * Valid only for replayable targets, e.g. MDT and OFD
1052          */
1053         if (tsi->tsi_exp->exp_obd->obd_replayable)
1054                 tgt_fmd_expire(tsi->tsi_exp);
1055
1056         rc = req_capsule_server_pack(tsi->tsi_pill);
1057         if (rc)
1058                 RETURN(err_serious(rc));
1059
1060         RETURN(rc);
1061 }
1062 EXPORT_SYMBOL(tgt_obd_ping);
1063
1064 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1065 {
1066         return err_serious(-EOPNOTSUPP);
1067 }
1068
1069 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1070 {
1071         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1072         struct obd_export       *exp = req->rq_export;
1073         struct ptlrpc_bulk_desc *desc;
1074         int                      i;
1075         int                      rc;
1076         int                      pages = 0;
1077
1078         ENTRY;
1079
1080         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1081                 unsigned int offset;
1082
1083                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1084                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1085                                       PAGE_SIZE);
1086         }
1087
1088         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1089                                   PTLRPC_BULK_PUT_SOURCE,
1090                                     MDS_BULK_PORTAL,
1091                                     &ptlrpc_bulk_kiov_nopin_ops);
1092         if (desc == NULL)
1093                 RETURN(-ENOMEM);
1094
1095         for (i = 0; i < rdbuf->rb_nbufs; i++)
1096                 desc->bd_frag_ops->add_iov_frag(desc,
1097                                         rdbuf->rb_bufs[i].lb_buf,
1098                                         rdbuf->rb_bufs[i].lb_len);
1099
1100         rc = target_bulk_io(exp, desc);
1101         ptlrpc_free_bulk(desc);
1102         RETURN(rc);
1103 }
1104 EXPORT_SYMBOL(tgt_send_buffer);
1105
1106 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1107 {
1108         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1109         struct obd_export       *exp = req->rq_export;
1110         struct ptlrpc_bulk_desc *desc;
1111         int                      tmpcount;
1112         int                      tmpsize;
1113         int                      i;
1114         int                      rc;
1115
1116         ENTRY;
1117
1118         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1119                                     PTLRPC_BULK_PUT_SOURCE,
1120                                     MDS_BULK_PORTAL,
1121                                     &ptlrpc_bulk_kiov_pin_ops);
1122         if (desc == NULL)
1123                 RETURN(-ENOMEM);
1124
1125         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1126                 /* old client requires reply size in it's PAGE_SIZE,
1127                  * which is rdpg->rp_count */
1128                 nob = rdpg->rp_count;
1129
1130         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1131              i++, tmpcount -= tmpsize) {
1132                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1133                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1134                                                  tmpsize);
1135         }
1136
1137         LASSERT(desc->bd_nob == nob);
1138         rc = target_bulk_io(exp, desc);
1139         ptlrpc_free_bulk(desc);
1140         RETURN(rc);
1141 }
1142 EXPORT_SYMBOL(tgt_sendpage);
1143
1144 /*
1145  * OBD_IDX_READ handler
1146  */
1147 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1148 {
1149         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1150         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1151         struct idx_info         *req_ii, *rep_ii;
1152         int                      rc, i;
1153
1154         ENTRY;
1155
1156         memset(rdpg, 0, sizeof(*rdpg));
1157         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1158
1159         /* extract idx_info buffer from request & reply */
1160         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1161         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1162                 RETURN(err_serious(-EPROTO));
1163
1164         rc = req_capsule_server_pack(tsi->tsi_pill);
1165         if (rc)
1166                 RETURN(err_serious(rc));
1167
1168         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1169         if (rep_ii == NULL)
1170                 RETURN(err_serious(-EFAULT));
1171         rep_ii->ii_magic = IDX_INFO_MAGIC;
1172
1173         /* extract hash to start with */
1174         rdpg->rp_hash = req_ii->ii_hash_start;
1175
1176         /* extract requested attributes */
1177         rdpg->rp_attrs = req_ii->ii_attrs;
1178
1179         /* check that fid packed in request is valid and supported */
1180         if (!fid_is_sane(&req_ii->ii_fid))
1181                 RETURN(-EINVAL);
1182         rep_ii->ii_fid = req_ii->ii_fid;
1183
1184         /* copy flags */
1185         rep_ii->ii_flags = req_ii->ii_flags;
1186
1187         /* compute number of pages to allocate, ii_count is the number of 4KB
1188          * containers */
1189         if (req_ii->ii_count <= 0)
1190                 GOTO(out, rc = -EFAULT);
1191         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1192                                exp_max_brw_size(tsi->tsi_exp));
1193         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1194
1195         /* allocate pages to store the containers */
1196         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1197         if (rdpg->rp_pages == NULL)
1198                 GOTO(out, rc = -ENOMEM);
1199         for (i = 0; i < rdpg->rp_npages; i++) {
1200                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1201                 if (rdpg->rp_pages[i] == NULL)
1202                         GOTO(out, rc = -ENOMEM);
1203         }
1204
1205         /* populate pages with key/record pairs */
1206         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1207         if (rc < 0)
1208                 GOTO(out, rc);
1209
1210         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1211                  "asked %d > %d\n", rc, rdpg->rp_count);
1212
1213         /* send pages to client */
1214         rc = tgt_sendpage(tsi, rdpg, rc);
1215         if (rc)
1216                 GOTO(out, rc);
1217         EXIT;
1218 out:
1219         if (rdpg->rp_pages) {
1220                 for (i = 0; i < rdpg->rp_npages; i++)
1221                         if (rdpg->rp_pages[i])
1222                                 __free_page(rdpg->rp_pages[i]);
1223                 OBD_FREE(rdpg->rp_pages,
1224                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1225         }
1226         return rc;
1227 }
1228
1229 struct tgt_handler tgt_obd_handlers[] = {
1230 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1231 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1232 };
1233 EXPORT_SYMBOL(tgt_obd_handlers);
1234
1235 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1236              struct dt_object *obj, __u64 start, __u64 end)
1237 {
1238         int rc = 0;
1239
1240         ENTRY;
1241
1242         /* if no objid is specified, it means "sync whole filesystem" */
1243         if (obj == NULL) {
1244                 rc = dt_sync(env, tgt->lut_bottom);
1245         } else if (dt_version_get(env, obj) >
1246                    tgt->lut_obd->obd_last_committed) {
1247                 rc = dt_object_sync(env, obj, start, end);
1248         }
1249         atomic_inc(&tgt->lut_sync_count);
1250
1251         RETURN(rc);
1252 }
1253 EXPORT_SYMBOL(tgt_sync);
1254 /*
1255  * Unified target DLM handlers.
1256  */
1257
1258 /**
1259  * Unified target BAST
1260  *
1261  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1262  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1263  * but if it's cross-MDT lock, because remote object version is not set, a
1264  * filesystem sync is needed.
1265  *
1266  * \param lock server side lock
1267  * \param desc lock desc
1268  * \param data ldlm_cb_set_arg
1269  * \param flag  indicates whether this cancelling or blocking callback
1270  * \retval      0 on success
1271  * \retval      negative number on error
1272  */
1273 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1274                             void *data, int flag)
1275 {
1276         struct lu_env            env;
1277         struct lu_target        *tgt;
1278         struct dt_object        *obj = NULL;
1279         struct lu_fid            fid;
1280         int                      rc = 0;
1281
1282         ENTRY;
1283
1284         tgt = class_exp2tgt(lock->l_export);
1285
1286         if (unlikely(tgt == NULL)) {
1287                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1288                        class_exp2obd(lock->l_export)->obd_name);
1289                 RETURN(-EINVAL);
1290         }
1291
1292         if (flag == LDLM_CB_CANCELING &&
1293             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1294             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1295              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1296               ldlm_is_cbpending(lock))) &&
1297             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1298              lock->l_resource->lr_type == LDLM_EXTENT)) {
1299                 __u64 start = 0;
1300                 __u64 end = OBD_OBJECT_EOF;
1301
1302                 rc = lu_env_init(&env, LCT_DT_THREAD);
1303                 if (unlikely(rc != 0))
1304                         GOTO(err, rc);
1305
1306                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1307                                    tgt->lut_lsd.lsd_osd_index);
1308
1309                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1310                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1311                         if (IS_ERR(obj))
1312                                 GOTO(err_env, rc = PTR_ERR(obj));
1313
1314                         if (!dt_object_exists(obj))
1315                                 GOTO(err_put, rc = -ENOENT);
1316
1317                         start = lock->l_policy_data.l_extent.start;
1318                         end = lock->l_policy_data.l_extent.end;
1319                 }
1320
1321                 rc = tgt_sync(&env, tgt, obj, start, end);
1322                 if (rc < 0) {
1323                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1324                                "cancel: rc = %d\n",
1325                                tgt_name(tgt), PFID(&fid),
1326                                lock->l_policy_data.l_extent.start,
1327                                lock->l_policy_data.l_extent.end, rc);
1328                 }
1329 err_put:
1330                 if (obj != NULL)
1331                         dt_object_put(&env, obj);
1332 err_env:
1333                 lu_env_fini(&env);
1334         }
1335 err:
1336         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1337         RETURN(rc);
1338 }
1339
1340 static struct ldlm_callback_suite tgt_dlm_cbs = {
1341         .lcs_completion = ldlm_server_completion_ast,
1342         .lcs_blocking   = tgt_blocking_ast,
1343         .lcs_glimpse    = ldlm_server_glimpse_ast
1344 };
1345
1346 int tgt_enqueue(struct tgt_session_info *tsi)
1347 {
1348         struct ptlrpc_request *req = tgt_ses_req(tsi);
1349         int rc;
1350
1351         ENTRY;
1352         /*
1353          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1354          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1355          */
1356         LASSERT(tsi->tsi_dlm_req != NULL);
1357         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1358                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1359         if (rc)
1360                 RETURN(err_serious(rc));
1361
1362         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1363         case LUT_FAIL_MDT:
1364                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1365                 break;
1366         case LUT_FAIL_OST:
1367                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1368                 break;
1369         case LUT_FAIL_MGT:
1370                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1371                 break;
1372         default:
1373                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1374                 break;
1375         }
1376         RETURN(req->rq_status);
1377 }
1378 EXPORT_SYMBOL(tgt_enqueue);
1379
1380 int tgt_convert(struct tgt_session_info *tsi)
1381 {
1382         struct ptlrpc_request *req = tgt_ses_req(tsi);
1383         int rc;
1384
1385         ENTRY;
1386         LASSERT(tsi->tsi_dlm_req);
1387         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1388         if (rc)
1389                 RETURN(err_serious(rc));
1390
1391         RETURN(req->rq_status);
1392 }
1393
1394 int tgt_bl_callback(struct tgt_session_info *tsi)
1395 {
1396         return err_serious(-EOPNOTSUPP);
1397 }
1398
1399 int tgt_cp_callback(struct tgt_session_info *tsi)
1400 {
1401         return err_serious(-EOPNOTSUPP);
1402 }
1403
1404 /* generic LDLM target handler */
1405 struct tgt_handler tgt_dlm_handlers[] = {
1406 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1407 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1408 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1409 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1410 };
1411 EXPORT_SYMBOL(tgt_dlm_handlers);
1412
1413 /*
1414  * Unified target LLOG handlers.
1415  */
1416 int tgt_llog_open(struct tgt_session_info *tsi)
1417 {
1418         int rc;
1419
1420         ENTRY;
1421
1422         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1423
1424         RETURN(rc);
1425 }
1426 EXPORT_SYMBOL(tgt_llog_open);
1427
1428 int tgt_llog_read_header(struct tgt_session_info *tsi)
1429 {
1430         int rc;
1431
1432         ENTRY;
1433
1434         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1435
1436         RETURN(rc);
1437 }
1438 EXPORT_SYMBOL(tgt_llog_read_header);
1439
1440 int tgt_llog_next_block(struct tgt_session_info *tsi)
1441 {
1442         int rc;
1443
1444         ENTRY;
1445
1446         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1447
1448         RETURN(rc);
1449 }
1450 EXPORT_SYMBOL(tgt_llog_next_block);
1451
1452 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1453 {
1454         int rc;
1455
1456         ENTRY;
1457
1458         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1459
1460         RETURN(rc);
1461 }
1462 EXPORT_SYMBOL(tgt_llog_prev_block);
1463
1464 /* generic llog target handler */
1465 struct tgt_handler tgt_llog_handlers[] = {
1466 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1467 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1468 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1469 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1470 };
1471 EXPORT_SYMBOL(tgt_llog_handlers);
1472
1473 /*
1474  * sec context handlers
1475  */
1476 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1477 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1478 {
1479         return 0;
1480 }
1481
1482 struct tgt_handler tgt_sec_ctx_handlers[] = {
1483 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1484 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1485 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1486 };
1487 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1488
1489 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1490                                  struct dt_device *key,
1491                                  struct lfsck_req_local *lrl,
1492                                  struct thandle *th) = NULL;
1493
1494 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1495                                                       struct dt_device *,
1496                                                       struct lfsck_req_local *,
1497                                                       struct thandle *))
1498 {
1499         tgt_lfsck_in_notify_local = notify;
1500 }
1501 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1502
1503 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1504                            struct dt_device *key,
1505                            struct lfsck_request *lr) = NULL;
1506
1507 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1508                                                 struct dt_device *,
1509                                                 struct lfsck_request *))
1510 {
1511         tgt_lfsck_in_notify = notify;
1512 }
1513 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1514
1515 static int (*tgt_lfsck_query)(const struct lu_env *env,
1516                               struct dt_device *key,
1517                               struct lfsck_request *req,
1518                               struct lfsck_reply *rep,
1519                               struct lfsck_query *que) = NULL;
1520
1521 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1522                                            struct dt_device *,
1523                                            struct lfsck_request *,
1524                                            struct lfsck_reply *,
1525                                            struct lfsck_query *))
1526 {
1527         tgt_lfsck_query = query;
1528 }
1529 EXPORT_SYMBOL(tgt_register_lfsck_query);
1530
1531 /* LFSCK request handlers */
1532 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1533 {
1534         const struct lu_env     *env = tsi->tsi_env;
1535         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1536         struct lfsck_request    *lr;
1537         int                      rc;
1538         ENTRY;
1539
1540         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1541         if (lr == NULL)
1542                 RETURN(-EPROTO);
1543
1544         rc = tgt_lfsck_in_notify(env, key, lr);
1545
1546         RETURN(rc);
1547 }
1548
1549 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1550 {
1551         struct lfsck_request    *request;
1552         struct lfsck_reply      *reply;
1553         int                      rc;
1554         ENTRY;
1555
1556         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1557         if (request == NULL)
1558                 RETURN(-EPROTO);
1559
1560         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1561         if (reply == NULL)
1562                 RETURN(-ENOMEM);
1563
1564         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1565                              request, reply, NULL);
1566
1567         RETURN(rc < 0 ? rc : 0);
1568 }
1569
1570 struct tgt_handler tgt_lfsck_handlers[] = {
1571 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1572 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1573 };
1574 EXPORT_SYMBOL(tgt_lfsck_handlers);
1575
1576 /*
1577  * initialize per-thread page pool (bug 5137).
1578  */
1579 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1580 {
1581         struct tgt_thread_big_cache *tbc;
1582
1583         ENTRY;
1584
1585         LASSERT(thread != NULL);
1586         LASSERT(thread->t_data == NULL);
1587
1588         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1589         if (tbc == NULL)
1590                 RETURN(-ENOMEM);
1591         thread->t_data = tbc;
1592         RETURN(0);
1593 }
1594 EXPORT_SYMBOL(tgt_io_thread_init);
1595
1596 /*
1597  * free per-thread pool created by tgt_thread_init().
1598  */
1599 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1600 {
1601         struct tgt_thread_big_cache *tbc;
1602
1603         ENTRY;
1604
1605         LASSERT(thread != NULL);
1606
1607         /*
1608          * be prepared to handle partially-initialized pools (because this is
1609          * called from ost_io_thread_init() for cleanup.
1610          */
1611         tbc = thread->t_data;
1612         if (tbc != NULL) {
1613                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1614                 thread->t_data = NULL;
1615         }
1616         EXIT;
1617 }
1618 EXPORT_SYMBOL(tgt_io_thread_done);
1619
1620 /**
1621  * Helper function for getting Data-on-MDT file server DLM lock
1622  * if asked by client.
1623  */
1624 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1625                       struct lustre_handle *lh, int mode, __u64 *flags)
1626 {
1627         union ldlm_policy_data policy = {
1628                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1629         };
1630         int rc;
1631
1632         ENTRY;
1633
1634         LASSERT(lh != NULL);
1635         LASSERT(ns != NULL);
1636         LASSERT(!lustre_handle_is_used(lh));
1637
1638         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1639                                     flags, ldlm_blocking_ast,
1640                                     ldlm_completion_ast, ldlm_glimpse_ast,
1641                                     NULL, 0, LVB_T_NONE, NULL, lh);
1642
1643         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1644 }
1645 EXPORT_SYMBOL(tgt_mdt_data_lock);
1646
1647 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1648 {
1649         LASSERT(lustre_handle_is_used(lh));
1650         ldlm_lock_decref(lh, mode);
1651 }
1652 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1653
1654 /**
1655  * Helper function for getting server side [start, start+count] DLM lock
1656  * if asked by client.
1657  */
1658 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1659                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1660                     struct lustre_handle *lh, int mode, __u64 *flags)
1661 {
1662         union ldlm_policy_data policy;
1663         int rc;
1664
1665         ENTRY;
1666
1667         LASSERT(lh != NULL);
1668         LASSERT(ns != NULL);
1669         LASSERT(!lustre_handle_is_used(lh));
1670
1671         policy.l_extent.gid = 0;
1672         policy.l_extent.start = start & PAGE_MASK;
1673
1674         /*
1675          * If ->o_blocks is EOF it means "lock till the end of the file".
1676          * Otherwise, it's size of an extent or hole being punched (in bytes).
1677          */
1678         if (end == OBD_OBJECT_EOF || end < start)
1679                 policy.l_extent.end = OBD_OBJECT_EOF;
1680         else
1681                 policy.l_extent.end = end | ~PAGE_MASK;
1682
1683         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1684                                     mode, flags, ldlm_blocking_ast,
1685                                     ldlm_completion_ast, ldlm_glimpse_ast,
1686                                     NULL, 0, LVB_T_NONE, NULL, lh);
1687         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1688 }
1689 EXPORT_SYMBOL(tgt_extent_lock);
1690
1691 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1692 {
1693         LASSERT(lustre_handle_is_used(lh));
1694         ldlm_lock_decref(lh, mode);
1695 }
1696 EXPORT_SYMBOL(tgt_extent_unlock);
1697
1698 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1699                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1700                         struct niobuf_remote *nb, struct lustre_handle *lh,
1701                         enum ldlm_mode mode)
1702 {
1703         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1704         __u64                    flags = 0;
1705         int                      nrbufs = obj->ioo_bufcnt;
1706         int                      i;
1707         int                      rc;
1708
1709         ENTRY;
1710
1711         LASSERT(mode == LCK_PR || mode == LCK_PW);
1712         LASSERT(!lustre_handle_is_used(lh));
1713
1714         if (ns->ns_obd->obd_recovering)
1715                 RETURN(0);
1716
1717         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1718                 RETURN(0);
1719
1720         for (i = 1; i < nrbufs; i++)
1721                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1722                         RETURN(-EFAULT);
1723
1724         /* MDT IO for data-on-mdt */
1725         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1726                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1727         else
1728                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1729                                      nb[nrbufs - 1].rnb_offset +
1730                                      nb[nrbufs - 1].rnb_len - 1,
1731                                      lh, mode, &flags);
1732         RETURN(rc);
1733 }
1734
1735 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1736                            struct lustre_handle *lh, enum ldlm_mode mode)
1737 {
1738         ENTRY;
1739
1740         LASSERT(mode == LCK_PR || mode == LCK_PW);
1741         LASSERT((obj->ioo_bufcnt > 0 &&
1742                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1743                 lustre_handle_is_used(lh));
1744
1745         if (lustre_handle_is_used(lh))
1746                 tgt_extent_unlock(lh, mode);
1747         EXIT;
1748 }
1749 static int tgt_checksum_niobuf(struct lu_target *tgt,
1750                                  struct niobuf_local *local_nb, int npages,
1751                                  int opc, enum cksum_types cksum_type,
1752                                  __u32 *cksum)
1753 {
1754         struct ahash_request           *req;
1755         unsigned int                    bufsize;
1756         int                             i, err;
1757         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1758
1759         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1760         if (IS_ERR(req)) {
1761                 CERROR("%s: unable to initialize checksum hash %s\n",
1762                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1763                 return PTR_ERR(req);
1764         }
1765
1766         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1767         for (i = 0; i < npages; i++) {
1768                 /* corrupt the data before we compute the checksum, to
1769                  * simulate a client->OST data error */
1770                 if (i == 0 && opc == OST_WRITE &&
1771                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1772                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1773                         int len = local_nb[i].lnb_len;
1774                         struct page *np = tgt_page_to_corrupt;
1775
1776                         if (np) {
1777                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1778                                 char *ptr2 = page_address(np);
1779
1780                                 memcpy(ptr2 + off, ptr + off, len);
1781                                 memcpy(ptr2 + off, "bad3", min(4, len));
1782                                 kunmap_atomic(ptr);
1783
1784                                 /* LU-8376 to preserve original index for
1785                                  * display in dump_all_bulk_pages() */
1786                                 np->index = i;
1787
1788                                 cfs_crypto_hash_update_page(req, np, off,
1789                                                             len);
1790                                 continue;
1791                         } else {
1792                                 CERROR("%s: can't alloc page for corruption\n",
1793                                        tgt_name(tgt));
1794                         }
1795                 }
1796                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1797                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1798                                   local_nb[i].lnb_len);
1799
1800                  /* corrupt the data after we compute the checksum, to
1801                  * simulate an OST->client data error */
1802                 if (i == 0 && opc == OST_READ &&
1803                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1804                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1805                         int len = local_nb[i].lnb_len;
1806                         struct page *np = tgt_page_to_corrupt;
1807
1808                         if (np) {
1809                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1810                                 char *ptr2 = page_address(np);
1811
1812                                 memcpy(ptr2 + off, ptr + off, len);
1813                                 memcpy(ptr2 + off, "bad4", min(4, len));
1814                                 kunmap_atomic(ptr);
1815
1816                                 /* LU-8376 to preserve original index for
1817                                  * display in dump_all_bulk_pages() */
1818                                 np->index = i;
1819
1820                                 cfs_crypto_hash_update_page(req, np, off,
1821                                                             len);
1822                                 continue;
1823                         } else {
1824                                 CERROR("%s: can't alloc page for corruption\n",
1825                                        tgt_name(tgt));
1826                         }
1827                 }
1828         }
1829
1830         bufsize = sizeof(*cksum);
1831         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1832
1833         return 0;
1834 }
1835
1836 char dbgcksum_file_name[PATH_MAX];
1837
1838 static void dump_all_bulk_pages(struct obdo *oa, int count,
1839                                 struct niobuf_local *local_nb,
1840                                 __u32 server_cksum, __u32 client_cksum)
1841 {
1842         struct file *filp;
1843         int rc, i;
1844         unsigned int len;
1845         char *buf;
1846
1847         /* will only keep dump of pages on first error for the same range in
1848          * file/fid, not during the resends/retries. */
1849         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1850                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1851                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1852                   libcfs_debug_file_path_arr :
1853                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1854                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1855                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1856                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1857                  local_nb[0].lnb_file_offset,
1858                  local_nb[count-1].lnb_file_offset +
1859                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1860         filp = filp_open(dbgcksum_file_name,
1861                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1862         if (IS_ERR(filp)) {
1863                 rc = PTR_ERR(filp);
1864                 if (rc == -EEXIST)
1865                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1866                                "checksum error: rc = %d\n", dbgcksum_file_name,
1867                                rc);
1868                 else
1869                         CERROR("%s: can't open to dump pages with checksum "
1870                                "error: rc = %d\n", dbgcksum_file_name, rc);
1871                 return;
1872         }
1873
1874         for (i = 0; i < count; i++) {
1875                 len = local_nb[i].lnb_len;
1876                 buf = kmap(local_nb[i].lnb_page);
1877                 while (len != 0) {
1878                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1879                         if (rc < 0) {
1880                                 CERROR("%s: wanted to write %u but got %d "
1881                                        "error\n", dbgcksum_file_name, len, rc);
1882                                 break;
1883                         }
1884                         len -= rc;
1885                         buf += rc;
1886                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1887                                dbgcksum_file_name, rc);
1888                 }
1889                 kunmap(local_nb[i].lnb_page);
1890         }
1891
1892         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1893         if (rc)
1894                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1895         filp_close(filp, NULL);
1896 }
1897
1898 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1899                                struct obd_export *exp, struct obdo *oa,
1900                                const struct lnet_process_id *peer,
1901                                __u32 client_cksum, __u32 server_cksum,
1902                                enum cksum_types server_cksum_type)
1903 {
1904         char *msg;
1905         enum cksum_types cksum_type;
1906         loff_t start, end;
1907
1908         /* unlikely to happen and only if resend does not occur due to cksum
1909          * control failure on Client */
1910         if (unlikely(server_cksum == client_cksum)) {
1911                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1912                        client_cksum);
1913                 return 0;
1914         }
1915
1916         if (exp->exp_obd->obd_checksum_dump)
1917                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1918                                     client_cksum);
1919
1920         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1921                                            oa->o_flags : 0);
1922
1923         if (cksum_type != server_cksum_type)
1924                 msg = "the server may have not used the checksum type specified"
1925                       " in the original request - likely a protocol problem";
1926         else
1927                 msg = "should have changed on the client or in transit";
1928
1929         start = local_nb[0].lnb_file_offset;
1930         end = local_nb[npages-1].lnb_file_offset +
1931                                         local_nb[npages-1].lnb_len - 1;
1932
1933         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1934                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1935                 " %x (type %x), server csum %x (type %x)\n",
1936                 exp->exp_obd->obd_name,
1937                 msg, libcfs_nid2str(peer->nid),
1938                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1939                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1940                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1941                 POSTID(&oa->o_oi),
1942                 start, end, client_cksum, cksum_type, server_cksum,
1943                 server_cksum_type);
1944
1945         return 1;
1946 }
1947
1948 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1949                              unsigned char *buf, int size)
1950 {
1951         int     i, off, len, copied = size;
1952         char    *ptr;
1953
1954         for (i = 0; i < npages; i++) {
1955                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1956                 len = local[i].lnb_len;
1957
1958                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1959                        i, off, len, size);
1960                 if (len > size)
1961                         return -EINVAL;
1962
1963                 ptr = kmap_atomic(local[i].lnb_page);
1964                 memcpy(buf + off, ptr, len);
1965                 kunmap_atomic(ptr);
1966                 buf += len;
1967                 size -= len;
1968         }
1969         return copied - size;
1970 }
1971
1972 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1973                                      struct niobuf_local *local_nb,
1974                                      int npages, int opc,
1975                                      obd_dif_csum_fn *fn,
1976                                      int sector_size,
1977                                      u32 *check_sum)
1978 {
1979         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1980         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1981         const char *obd_name = tgt->lut_obd->obd_name;
1982         struct ahash_request *req;
1983         unsigned int bufsize;
1984         unsigned char *buffer;
1985         struct page *__page;
1986         __u16 *guard_start;
1987         int guard_number;
1988         int used_number = 0;
1989         __u32 cksum;
1990         int rc = 0;
1991         int used;
1992         int i;
1993
1994         __page = alloc_page(GFP_KERNEL);
1995         if (__page == NULL)
1996                 return -ENOMEM;
1997
1998         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1999         if (IS_ERR(req)) {
2000                 CERROR("%s: unable to initialize checksum hash %s\n",
2001                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2002                 return PTR_ERR(req);
2003         }
2004
2005         buffer = kmap(__page);
2006         guard_start = (__u16 *)buffer;
2007         guard_number = PAGE_SIZE / sizeof(*guard_start);
2008         for (i = 0; i < npages; i++) {
2009                 /* corrupt the data before we compute the checksum, to
2010                  * simulate a client->OST data error */
2011                 if (i == 0 && opc == OST_WRITE &&
2012                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2013                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2014                         int len = local_nb[i].lnb_len;
2015                         struct page *np = tgt_page_to_corrupt;
2016
2017                         if (np) {
2018                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2019                                 char *ptr2 = page_address(np);
2020
2021                                 memcpy(ptr2 + off, ptr + off, len);
2022                                 memcpy(ptr2 + off, "bad3", min(4, len));
2023                                 kunmap_atomic(ptr);
2024
2025                                 /* LU-8376 to preserve original index for
2026                                  * display in dump_all_bulk_pages() */
2027                                 np->index = i;
2028
2029                                 cfs_crypto_hash_update_page(req, np, off,
2030                                                             len);
2031                                 continue;
2032                         } else {
2033                                 CERROR("%s: can't alloc page for corruption\n",
2034                                        tgt_name(tgt));
2035                         }
2036                 }
2037
2038                 /*
2039                  * The left guard number should be able to hold checksums of a
2040                  * whole page
2041                  */
2042                 if (t10_cksum_type && opc == OST_READ &&
2043                     local_nb[i].lnb_guard_disk) {
2044                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2045                         if (used > (guard_number - used_number)) {
2046                                 rc = -E2BIG;
2047                                 break;
2048                         }
2049                         memcpy(guard_start + used_number,
2050                                local_nb[i].lnb_guards,
2051                                used * sizeof(*local_nb[i].lnb_guards));
2052                 } else {
2053                         rc = obd_page_dif_generate_buffer(obd_name,
2054                                 local_nb[i].lnb_page,
2055                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2056                                 local_nb[i].lnb_len, guard_start + used_number,
2057                                 guard_number - used_number, &used, sector_size,
2058                                 fn);
2059                         if (rc)
2060                                 break;
2061                 }
2062
2063                 LASSERT(used <= MAX_GUARD_NUMBER);
2064                 /*
2065                  * If disk support T10PI checksum, copy guards to local_nb.
2066                  * If the write is partial page, do not use the guards for bio
2067                  * submission since the data might not be full-sector. The bio
2068                  * guards will be generated later based on the full sectors. If
2069                  * the sector size is 512B rather than 4 KB, or the page size
2070                  * is larger than 4KB, this might drop some useful guards for
2071                  * partial page write, but it will only add minimal extra time
2072                  * of checksum calculation.
2073                  */
2074                 if (t10_cksum_type && opc == OST_WRITE &&
2075                     local_nb[i].lnb_len == PAGE_SIZE) {
2076                         local_nb[i].lnb_guard_rpc = 1;
2077                         memcpy(local_nb[i].lnb_guards,
2078                                guard_start + used_number,
2079                                used * sizeof(*local_nb[i].lnb_guards));
2080                 }
2081
2082                 used_number += used;
2083                 if (used_number == guard_number) {
2084                         cfs_crypto_hash_update_page(req, __page, 0,
2085                                 used_number * sizeof(*guard_start));
2086                         used_number = 0;
2087                 }
2088
2089                  /* corrupt the data after we compute the checksum, to
2090                  * simulate an OST->client data error */
2091                 if (unlikely(i == 0 && opc == OST_READ &&
2092                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2093                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2094                         int len = local_nb[i].lnb_len;
2095                         struct page *np = tgt_page_to_corrupt;
2096
2097                         if (np) {
2098                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2099                                 char *ptr2 = page_address(np);
2100
2101                                 memcpy(ptr2 + off, ptr + off, len);
2102                                 memcpy(ptr2 + off, "bad4", min(4, len));
2103                                 kunmap_atomic(ptr);
2104
2105                                 /* LU-8376 to preserve original index for
2106                                  * display in dump_all_bulk_pages() */
2107                                 np->index = i;
2108
2109                                 cfs_crypto_hash_update_page(req, np, off,
2110                                                             len);
2111                                 continue;
2112                         } else {
2113                                 CERROR("%s: can't alloc page for corruption\n",
2114                                        tgt_name(tgt));
2115                         }
2116                 }
2117         }
2118         kunmap(__page);
2119         if (rc)
2120                 GOTO(out, rc);
2121
2122         if (used_number != 0)
2123                 cfs_crypto_hash_update_page(req, __page, 0,
2124                         used_number * sizeof(*guard_start));
2125
2126         bufsize = sizeof(cksum);
2127         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2128
2129         if (rc == 0)
2130                 *check_sum = cksum;
2131 out:
2132         __free_page(__page);
2133         return rc;
2134 }
2135
2136 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2137                                   enum cksum_types cksum_type,
2138                                   struct niobuf_local *local_nb,
2139                                   int npages, int opc, u32 *check_sum)
2140 {
2141         obd_dif_csum_fn *fn = NULL;
2142         int sector_size = 0;
2143         int rc;
2144
2145         ENTRY;
2146         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2147
2148         if (fn)
2149                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2150                                                opc, fn, sector_size,
2151                                                check_sum);
2152         else
2153                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2154                                          cksum_type, check_sum);
2155         RETURN(rc);
2156 }
2157
2158 int tgt_brw_read(struct tgt_session_info *tsi)
2159 {
2160         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2161         struct ptlrpc_bulk_desc *desc = NULL;
2162         struct obd_export       *exp = tsi->tsi_exp;
2163         struct niobuf_remote    *remote_nb;
2164         struct niobuf_local     *local_nb;
2165         struct obd_ioobj        *ioo;
2166         struct ost_body         *body, *repbody;
2167         struct lustre_handle     lockh = { 0 };
2168         int                      npages, nob = 0, rc, i, no_reply = 0,
2169                                  npages_read;
2170         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2171         const char *obd_name = exp->exp_obd->obd_name;
2172
2173         ENTRY;
2174
2175         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2176             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2177                 CERROR("%s: deny read request from %s to portal %u\n",
2178                        tgt_name(tsi->tsi_tgt),
2179                        obd_export_nid2str(req->rq_export),
2180                        ptlrpc_req2svc(req)->srv_req_portal);
2181                 RETURN(-EPROTO);
2182         }
2183
2184         req->rq_bulk_read = 1;
2185
2186         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2187                 RETURN(-EIO);
2188
2189         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2190                          cfs_fail_val : (obd_timeout + 1) / 4);
2191
2192         /* Check if there is eviction in progress, and if so, wait for it to
2193          * finish */
2194         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2195                 /* We do not care how long it takes */
2196                 wait_event_idle(
2197                         exp->exp_obd->obd_evict_inprogress_waitq,
2198                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2199         }
2200
2201         /* There must be big cache in current thread to process this request
2202          * if it is NULL then something went wrong and it wasn't allocated,
2203          * report -ENOMEM in that case */
2204         if (tbc == NULL)
2205                 RETURN(-ENOMEM);
2206
2207         body = tsi->tsi_ost_body;
2208         LASSERT(body != NULL);
2209
2210         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2211         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2212
2213         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2214         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2215
2216         local_nb = tbc->local;
2217
2218         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2219                           &lockh, LCK_PR);
2220         if (rc != 0)
2221                 RETURN(rc);
2222
2223         /*
2224          * If getting the lock took more time than
2225          * client was willing to wait, drop it. b=11330
2226          */
2227         if (ktime_get_real_seconds() > req->rq_deadline ||
2228             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2229                 no_reply = 1;
2230                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2231                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2232                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2233                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2234                 GOTO(out_lock, rc = -ETIMEDOUT);
2235         }
2236
2237         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2238         repbody->oa = body->oa;
2239
2240         npages = PTLRPC_MAX_BRW_PAGES;
2241         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2242                         ioo, remote_nb, &npages, local_nb);
2243         if (rc != 0)
2244                 GOTO(out_lock, rc);
2245
2246         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2247             body->oa.o_flags & OBD_FL_SHORT_IO) {
2248                 desc = NULL;
2249         } else {
2250                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2251                                             PTLRPC_BULK_PUT_SOURCE,
2252                                             OST_BULK_PORTAL,
2253                                             &ptlrpc_bulk_kiov_nopin_ops);
2254                 if (desc == NULL)
2255                         GOTO(out_commitrw, rc = -ENOMEM);
2256         }
2257
2258         nob = 0;
2259         npages_read = npages;
2260         for (i = 0; i < npages; i++) {
2261                 int page_rc = local_nb[i].lnb_rc;
2262
2263                 if (page_rc < 0) {
2264                         rc = page_rc;
2265                         npages_read = i;
2266                         break;
2267                 }
2268
2269                 nob += page_rc;
2270                 if (page_rc != 0 && desc != NULL) { /* some data! */
2271                         LASSERT(local_nb[i].lnb_page != NULL);
2272                         desc->bd_frag_ops->add_kiov_frag
2273                           (desc, local_nb[i].lnb_page,
2274                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2275                            page_rc);
2276                 }
2277
2278                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2279                         local_nb[i].lnb_len = page_rc;
2280                         npages_read = i + (page_rc != 0 ? 1 : 0);
2281                         /* All subsequent pages should be 0 */
2282                         while (++i < npages)
2283                                 LASSERT(local_nb[i].lnb_rc == 0);
2284                         break;
2285                 }
2286         }
2287
2288         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2289                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2290                            body->oa.o_flags : 0;
2291                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2292
2293                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2294                                                           cksum_type);
2295                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2296
2297                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2298                                             local_nb, npages_read, OST_READ,
2299                                             &repbody->oa.o_cksum);
2300                 if (rc < 0)
2301                         GOTO(out_commitrw, rc);
2302                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2303                        repbody->oa.o_cksum);
2304
2305                 /* if a resend it could be for a cksum error, so check Server
2306                  * cksum with returned Client cksum (this should even cover
2307                  * zero-cksum case) */
2308                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2309                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2310                         check_read_checksum(local_nb, npages_read, exp,
2311                                             &body->oa, &req->rq_peer,
2312                                             body->oa.o_cksum,
2313                                             repbody->oa.o_cksum, cksum_type);
2314         } else {
2315                 repbody->oa.o_valid = 0;
2316         }
2317         /* We're finishing using body->oa as an input variable */
2318
2319         /* Check if client was evicted while we were doing i/o before touching
2320          * network */
2321         if (rc == 0) {
2322                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2323                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2324                         unsigned char *short_io_buf;
2325                         int short_io_size;
2326
2327                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2328                                                               &RMF_SHORT_IO);
2329                         short_io_size = req_capsule_get_size(&req->rq_pill,
2330                                                              &RMF_SHORT_IO,
2331                                                              RCL_SERVER);
2332                         rc = tgt_pages2shortio(local_nb, npages_read,
2333                                                short_io_buf, short_io_size);
2334                         if (rc >= 0)
2335                                 req_capsule_shrink(&req->rq_pill,
2336                                                    &RMF_SHORT_IO, rc,
2337                                                    RCL_SERVER);
2338                         rc = rc > 0 ? 0 : rc;
2339                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2340                         rc = target_bulk_io(exp, desc);
2341                 }
2342                 no_reply = rc != 0;
2343         } else {
2344                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2345                     body->oa.o_flags & OBD_FL_SHORT_IO)
2346                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2347                                            RCL_SERVER);
2348         }
2349
2350 out_commitrw:
2351         /* Must commit after prep above in all cases */
2352         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2353                           remote_nb, npages, local_nb, rc);
2354 out_lock:
2355         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2356
2357         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2358                 ptlrpc_free_bulk(desc);
2359
2360         LASSERT(rc <= 0);
2361         if (rc == 0) {
2362                 rc = nob;
2363                 ptlrpc_lprocfs_brw(req, nob);
2364         } else if (no_reply) {
2365                 req->rq_no_reply = 1;
2366                 /* reply out callback would free */
2367                 ptlrpc_req_drop_rs(req);
2368                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2369                               "client will retry: rc %d\n",
2370                               obd_name,
2371                               obd_uuid2str(&exp->exp_client_uuid),
2372                               obd_export_nid2str(exp), rc);
2373         }
2374         /* send a bulk after reply to simulate a network delay or reordering
2375          * by a router - Note that !desc implies short io, so there is no bulk
2376          * to reorder. */
2377         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2378             desc) {
2379                 CDEBUG(D_INFO, "reorder BULK\n");
2380
2381                 ssleep(3);
2382                 target_bulk_io(exp, desc);
2383                 ptlrpc_free_bulk(desc);
2384         }
2385
2386         RETURN(rc);
2387 }
2388 EXPORT_SYMBOL(tgt_brw_read);
2389
2390 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2391                              unsigned char *buf, unsigned int size)
2392 {
2393         int     i, off, len;
2394         char    *ptr;
2395
2396         for (i = 0; i < npages; i++) {
2397                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2398                 len = local[i].lnb_len;
2399
2400                 if (len == 0)
2401                         continue;
2402
2403                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2404                        i, off, len, size);
2405                 ptr = kmap_atomic(local[i].lnb_page);
2406                 if (ptr == NULL)
2407                         return -EINVAL;
2408                 memcpy(ptr + off, buf, len < size ? len : size);
2409                 kunmap_atomic(ptr);
2410                 buf += len;
2411                 size -= len;
2412         }
2413         return 0;
2414 }
2415
2416 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2417                               struct ptlrpc_bulk_desc *desc,
2418                               struct niobuf_local *local_nb, int npages,
2419                               u32 client_cksum, u32 server_cksum,
2420                               bool mmap)
2421 {
2422         struct obd_export *exp = req->rq_export;
2423         struct ost_body *body;
2424         char *router = "";
2425         char *via = "";
2426
2427         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2428         LASSERT(body != NULL);
2429
2430         if (desc && req->rq_peer.nid != desc->bd_sender) {
2431                 via = " via ";
2432                 router = libcfs_nid2str(desc->bd_sender);
2433         }
2434
2435         if (exp->exp_obd->obd_checksum_dump)
2436                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2437                                     client_cksum);
2438
2439         if (mmap) {
2440                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2441                              client_cksum, server_cksum);
2442                 return;
2443         }
2444
2445         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2446                            DFID" object "DOSTID" extent [%llu-%llu"
2447                            "]: client csum %x, server csum %x\n",
2448                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2449                            via, router,
2450                            body->oa.o_valid & OBD_MD_FLFID ?
2451                            body->oa.o_parent_seq : (__u64)0,
2452                            body->oa.o_valid & OBD_MD_FLFID ?
2453                            body->oa.o_parent_oid : 0,
2454                            body->oa.o_valid & OBD_MD_FLFID ?
2455                            body->oa.o_parent_ver : 0,
2456                            POSTID(&body->oa.o_oi),
2457                            local_nb[0].lnb_file_offset,
2458                            local_nb[npages-1].lnb_file_offset +
2459                            local_nb[npages - 1].lnb_len - 1,
2460                            client_cksum, server_cksum);
2461 }
2462
2463 int tgt_brw_write(struct tgt_session_info *tsi)
2464 {
2465         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2466         struct ptlrpc_bulk_desc *desc = NULL;
2467         struct obd_export       *exp = req->rq_export;
2468         struct niobuf_remote    *remote_nb;
2469         struct niobuf_local     *local_nb;
2470         struct obd_ioobj        *ioo;
2471         struct ost_body         *body, *repbody;
2472         struct lustre_handle     lockh = {0};
2473         __u32                   *rcs;
2474         int                      objcount, niocount, npages;
2475         int                      rc, i, j;
2476         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2477         bool                     no_reply = false, mmap;
2478         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2479         bool wait_sync = false;
2480         const char *obd_name = exp->exp_obd->obd_name;
2481
2482         ENTRY;
2483
2484         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2485             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2486                 CERROR("%s: deny write request from %s to portal %u\n",
2487                        tgt_name(tsi->tsi_tgt),
2488                        obd_export_nid2str(req->rq_export),
2489                        ptlrpc_req2svc(req)->srv_req_portal);
2490                 RETURN(err_serious(-EPROTO));
2491         }
2492
2493         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2494                 RETURN(err_serious(-ENOSPC));
2495         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2496                 RETURN(err_serious(-EROFS));
2497
2498         req->rq_bulk_write = 1;
2499
2500         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2501                 RETURN(err_serious(-EIO));
2502         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2503                 RETURN(err_serious(-EFAULT));
2504
2505         /* pause before transaction has been started */
2506         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2507                          cfs_fail_val : (obd_timeout + 1) / 4);
2508
2509         /* Delay write commit to show stale size information */
2510         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2511
2512         /* There must be big cache in current thread to process this request
2513          * if it is NULL then something went wrong and it wasn't allocated,
2514          * report -ENOMEM in that case */
2515         if (tbc == NULL)
2516                 RETURN(-ENOMEM);
2517
2518         body = tsi->tsi_ost_body;
2519         LASSERT(body != NULL);
2520
2521         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2522         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2523
2524         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2525                                         RCL_CLIENT) / sizeof(*ioo);
2526
2527         for (niocount = i = 0; i < objcount; i++)
2528                 niocount += ioo[i].ioo_bufcnt;
2529
2530         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2531         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2532         if (niocount != req_capsule_get_size(&req->rq_pill,
2533                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2534                         sizeof(*remote_nb))
2535                 RETURN(err_serious(-EPROTO));
2536
2537         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2538             ptlrpc_connection_is_local(exp->exp_connection))
2539                 memory_pressure_set();
2540
2541         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2542                              niocount * sizeof(*rcs));
2543         rc = req_capsule_server_pack(&req->rq_pill);
2544         if (rc != 0)
2545                 GOTO(out, rc = err_serious(rc));
2546
2547         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2548         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2549
2550         local_nb = tbc->local;
2551
2552         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2553                           &lockh, LCK_PW);
2554         if (rc != 0)
2555                 GOTO(out, rc);
2556
2557         /*
2558          * If getting the lock took more time than
2559          * client was willing to wait, drop it. b=11330
2560          */
2561         if (ktime_get_real_seconds() > req->rq_deadline ||
2562             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2563                 no_reply = true;
2564                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2565                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2566                        POSTID(&ioo->ioo_oid),
2567                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2568                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2569                 GOTO(out_lock, rc = -ETIMEDOUT);
2570         }
2571
2572         /* Because we already sync grant info with client when reconnect,
2573          * grant info will be cleared for resent req, then fed_grant and
2574          * total_grant will not be modified in following preprw_write */
2575         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2576                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2577                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2578         }
2579
2580         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2581         if (repbody == NULL)
2582                 GOTO(out_lock, rc = -ENOMEM);
2583         repbody->oa = body->oa;
2584
2585         npages = PTLRPC_MAX_BRW_PAGES;
2586         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2587                         objcount, ioo, remote_nb, &npages, local_nb);
2588         if (rc < 0)
2589                 GOTO(out_lock, rc);
2590         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2591             body->oa.o_flags & OBD_FL_SHORT_IO) {
2592                 unsigned int short_io_size;
2593                 unsigned char *short_io_buf;
2594
2595                 short_io_size = req_capsule_get_size(&req->rq_pill,
2596                                                      &RMF_SHORT_IO,
2597                                                      RCL_CLIENT);
2598                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2599                                                       &RMF_SHORT_IO);
2600                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2601                                " size = %d\n", short_io_size);
2602
2603                 /* Copy short io buf to pages */
2604                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2605                                        short_io_size);
2606                 desc = NULL;
2607         } else {
2608                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2609                                             PTLRPC_BULK_GET_SINK,
2610                                             OST_BULK_PORTAL,
2611                                             &ptlrpc_bulk_kiov_nopin_ops);
2612                 if (desc == NULL)
2613                         GOTO(skip_transfer, rc = -ENOMEM);
2614
2615                 /* NB Having prepped, we must commit... */
2616                 for (i = 0; i < npages; i++)
2617                         desc->bd_frag_ops->add_kiov_frag(desc,
2618                                         local_nb[i].lnb_page,
2619                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2620                                         local_nb[i].lnb_len);
2621
2622                 rc = sptlrpc_svc_prep_bulk(req, desc);
2623                 if (rc != 0)
2624                         GOTO(skip_transfer, rc);
2625
2626                 rc = target_bulk_io(exp, desc);
2627         }
2628
2629         no_reply = rc != 0;
2630
2631 skip_transfer:
2632         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2633                 static int cksum_counter;
2634
2635                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2636                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2637
2638                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2639                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2640                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2641                                                            cksum_type);
2642
2643                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2644                                             local_nb, npages, OST_WRITE,
2645                                             &repbody->oa.o_cksum);
2646                 if (rc < 0)
2647                         GOTO(out_commitrw, rc);
2648
2649                 cksum_counter++;
2650
2651                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2652                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2653                                 body->oa.o_flags & OBD_FL_MMAP);
2654
2655                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2656                                           body->oa.o_cksum,
2657                                           repbody->oa.o_cksum, mmap);
2658                         cksum_counter = 0;
2659                 } else if ((cksum_counter & (-cksum_counter)) ==
2660                            cksum_counter) {
2661                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2662                                cksum_counter, libcfs_id2str(req->rq_peer),
2663                                repbody->oa.o_cksum);
2664                 }
2665         }
2666
2667 out_commitrw:
2668         /* Must commit after prep above in all cases */
2669         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2670                           objcount, ioo, remote_nb, npages, local_nb, rc);
2671         if (rc == -ENOTCONN)
2672                 /* quota acquire process has been given up because
2673                  * either the client has been evicted or the client
2674                  * has timed out the request already */
2675                 no_reply = true;
2676
2677         for (i = 0; i < niocount; i++) {
2678                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2679                         wait_sync = true;
2680                         break;
2681                 }
2682         }
2683         /*
2684          * Disable sending mtime back to the client. If the client locked the
2685          * whole object, then it has already updated the mtime on its side,
2686          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2687          */
2688         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2689
2690         if (rc == 0) {
2691                 int nob = 0;
2692
2693                 /* set per-requested niobuf return codes */
2694                 for (i = j = 0; i < niocount; i++) {
2695                         int len = remote_nb[i].rnb_len;
2696
2697                         nob += len;
2698                         rcs[i] = 0;
2699                         do {
2700                                 LASSERT(j < npages);
2701                                 if (local_nb[j].lnb_rc < 0)
2702                                         rcs[i] = local_nb[j].lnb_rc;
2703                                 len -= local_nb[j].lnb_len;
2704                                 j++;
2705                         } while (len > 0);
2706                         LASSERT(len == 0);
2707                 }
2708                 LASSERT(j == npages);
2709                 ptlrpc_lprocfs_brw(req, nob);
2710         }
2711 out_lock:
2712         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2713         if (desc)
2714                 ptlrpc_free_bulk(desc);
2715 out:
2716         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2717                 req->rq_no_reply = 1;
2718                 /* reply out callback would free */
2719                 ptlrpc_req_drop_rs(req);
2720                 if (!exp->exp_obd->obd_no_transno)
2721                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2722                                       " client will retry: rc = %d\n",
2723                                       obd_name,
2724                                       obd_uuid2str(&exp->exp_client_uuid),
2725                                       obd_export_nid2str(exp), rc);
2726         }
2727         memory_pressure_clr();
2728         RETURN(rc);
2729 }
2730 EXPORT_SYMBOL(tgt_brw_write);
2731
2732 /* Check if request can be reconstructed from saved reply data
2733  * A copy of the reply data is returned in @trd if the pointer is not NULL
2734  */
2735 int req_can_reconstruct(struct ptlrpc_request *req,
2736                          struct tg_reply_data *trd)
2737 {
2738         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2739         struct lsd_client_data *lcd = ted->ted_lcd;
2740         int found;
2741
2742         if (tgt_is_multimodrpcs_client(req->rq_export))
2743                 return tgt_lookup_reply(req, trd);
2744
2745         mutex_lock(&ted->ted_lcd_lock);
2746         found = req->rq_xid == lcd->lcd_last_xid ||
2747                 req->rq_xid == lcd->lcd_last_close_xid;
2748
2749         if (found && trd != NULL) {
2750                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2751                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2752                         trd->trd_reply.lrd_transno =
2753                                                 lcd->lcd_last_close_transno;
2754                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2755                 } else {
2756                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2757                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2758                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2759                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2760                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2761                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2762                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2763                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2764                 }
2765         }
2766         mutex_unlock(&ted->ted_lcd_lock);
2767
2768         return found;
2769 }
2770 EXPORT_SYMBOL(req_can_reconstruct);
2771