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