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