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 #ifdef HAVE_UIDGID_HEADER
  40 # include <linux/uidgid.h>
  41 #endif
  42
  43 #include <obd.h>
  44 #include <obd_class.h>
  45 #include <obd_cksum.h>
  46 #include <lustre_lfsck.h>
  47 #include <lustre_nodemap.h>
  48 #include <lustre_acl.h>
  49
  50 #include "tgt_internal.h"
  51
  52 char *tgt_name(struct lu_target *tgt)
  53 {
  54         LASSERT(tgt->lut_obd != NULL);
  55         return tgt->lut_obd->obd_name;
  56 }
  57 EXPORT_SYMBOL(tgt_name);
  58
  59 /*
  60  * Generic code handling requests that have struct mdt_body passed in:
  61  *
  62  *  - extract mdt_body from request and save it in @tsi, if present;
  63  *
  64  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  65  *  @tsi;
  66  *
  67  *  - if HAS_BODY flag is set for this request type check whether object
  68  *  actually exists on storage (lu_object_exists()).
  69  *
  70  */
  71 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  72 {
  73         const struct mdt_body   *body;
  74         struct lu_object        *obj;
  75         struct req_capsule      *pill = tsi->tsi_pill;
  76         int                      rc;
  77
  78         ENTRY;
  79
  80         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  81         if (body == NULL)
  82                 RETURN(-EFAULT);
  83
  84         tsi->tsi_mdt_body = body;
  85
  86         if (!(body->mbo_valid & OBD_MD_FLID))
  87                 RETURN(0);
  88
  89         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  90          * in any case in pre-2.5 clients. Fix that here if needed */
  91         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  92                 RETURN(0);
  93
  94         if (!fid_is_sane(&body->mbo_fid1)) {
  95                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  96                        PFID(&body->mbo_fid1));
  97                 RETURN(-EINVAL);
  98         }
  99
 100         obj = lu_object_find(tsi->tsi_env,
 101                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 102                              &body->mbo_fid1, NULL);
 103         if (!IS_ERR(obj)) {
 104                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 105                         lu_object_put(tsi->tsi_env, obj);
 106                         rc = -ENOENT;
 107                 } else {
 108                         tsi->tsi_corpus = obj;
 109                         rc = 0;
 110                 }
 111         } else {
 112                 rc = PTR_ERR(obj);
 113         }
 114
 115         tsi->tsi_fid = body->mbo_fid1;
 116
 117         RETURN(rc);
 118 }
 119
 120 /**
 121  * Validate oa from client.
 122  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 123  * req are valid.
 124  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 125  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 126  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 127  *       whether this is oi_fid or real ostid. So it will check
 128  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 129  *    c. Old FID-disable osc will send IDIF.
 130  *    d. new FID-enable osc/osp will send normal FID.
 131  *
 132  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 133  * be used for LAST_ID file, and only being accessed inside OST now.
 134  */
 135 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 136 {
 137         struct ost_id   *oi     = &oa->o_oi;
 138         u64              seq    = ostid_seq(oi);
 139         u64              id     = ostid_id(oi);
 140         int              rc;
 141         ENTRY;
 142
 143         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 144                      fid_seq_is_echo(seq))) {
 145                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 146                  * during create, and we will reset this to 1, since this
 147                  * oi_id is basically useless in the following create process,
 148                  * but oi_id == 0 will make it difficult to tell whether it is
 149                  * real FID or ost_id. */
 150                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 151                 oi->oi_fid.f_oid = id ?: 1;
 152                 oi->oi_fid.f_ver = 0;
 153         } else {
 154                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 155
 156                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 157                         GOTO(out, rc = -EPROTO);
 158
 159                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 160                  * all of the requests are required to setup FLGROUP */
 161                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 162                         ostid_set_seq_mdt0(oi);
 163                         oa->o_valid |= OBD_MD_FLGROUP;
 164                         seq = ostid_seq(oi);
 165                 }
 166
 167                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 168                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 169                         GOTO(out, rc = -EPROTO);
 170
 171                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 172                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 173                 if (unlikely(rc != 0))
 174                         GOTO(out, rc);
 175
 176                 oi->oi_fid = tti->tti_fid1;
 177         }
 178
 179         RETURN(0);
 180
 181 out:
 182         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 183                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 184                seq, id, rc);
 185         return rc;
 186 }
 187 EXPORT_SYMBOL(tgt_validate_obdo);
 188
 189 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 190 {
 191         unsigned                 max_brw;
 192         struct niobuf_remote    *rnb;
 193         struct obd_ioobj        *ioo;
 194         int                      obj_count;
 195
 196         ENTRY;
 197
 198         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 199         if (ioo == NULL)
 200                 RETURN(-EPROTO);
 201
 202         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 203         if (rnb == NULL)
 204                 RETURN(-EPROTO);
 205
 206         max_brw = ioobj_max_brw_get(ioo);
 207         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 208                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 209                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 210                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 211                        POSTID(oi), -EPROTO);
 212                 RETURN(-EPROTO);
 213         }
 214         ioo->ioo_oid = *oi;
 215
 216         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 217                                         RCL_CLIENT) / sizeof(*ioo);
 218         if (obj_count == 0) {
 219                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 220                 RETURN(-EPROTO);
 221         } else if (obj_count > 1) {
 222                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 223                        obj_count);
 224                 RETURN(-EPROTO);
 225         }
 226
 227         if (ioo->ioo_bufcnt == 0) {
 228                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 229                 RETURN(-EPROTO);
 230         }
 231
 232         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 233                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 234                           "bulk has too many pages (%d)",
 235                           ioo->ioo_bufcnt);
 236                 RETURN(-EPROTO);
 237         }
 238
 239         RETURN(0);
 240 }
 241
 242 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 243 {
 244         struct ost_body         *body;
 245         struct req_capsule      *pill = tsi->tsi_pill;
 246         struct lu_nodemap       *nodemap;
 247         int                      rc;
 248
 249         ENTRY;
 250
 251         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 252         if (body == NULL)
 253                 RETURN(-EFAULT);
 254
 255         rc = tgt_validate_obdo(tsi, &body->oa);
 256         if (rc)
 257                 RETURN(rc);
 258
 259         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 260         if (IS_ERR(nodemap))
 261                 RETURN(PTR_ERR(nodemap));
 262
 263         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 264                                         NODEMAP_CLIENT_TO_FS,
 265                                         body->oa.o_uid);
 266         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 267                                         NODEMAP_CLIENT_TO_FS,
 268                                         body->oa.o_gid);
 269         nodemap_putref(nodemap);
 270
 271         tsi->tsi_ost_body = body;
 272         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 273
 274         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 275                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 276                 if (rc < 0)
 277                         RETURN(rc);
 278         }
 279
 280         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 281                 if (flags & HAS_BODY) {
 282                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 283                                tgt_name(tsi->tsi_tgt));
 284                         RETURN(-EPROTO);
 285                 } else {
 286                         RETURN(0);
 287                 }
 288         }
 289
 290         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 291
 292         /*
 293          * OST doesn't get object in advance for further use to prevent
 294          * situations with nested object_find which is potential deadlock.
 295          */
 296         tsi->tsi_corpus = NULL;
 297         RETURN(rc);
 298 }
 299
 300 /*
 301  * Do necessary preprocessing according to handler ->th_flags.
 302  */
 303 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 304                                   struct tgt_handler *h,
 305                                   struct ptlrpc_request *req)
 306 {
 307         struct req_capsule      *pill = tsi->tsi_pill;
 308         __u32                    flags = h->th_flags;
 309         int                      rc = 0;
 310
 311         ENTRY;
 312
 313         if (tsi->tsi_preprocessed)
 314                 RETURN(0);
 315
 316         LASSERT(h->th_act != NULL);
 317         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 318         LASSERT(current->journal_info == NULL);
 319
 320         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 321                      h->th_fmt != NULL));
 322         if (h->th_fmt != NULL) {
 323                 req_capsule_set(pill, h->th_fmt);
 324                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
 325                         rc = tgt_mdt_body_unpack(tsi, flags);
 326                         if (rc < 0)
 327                                 RETURN(rc);
 328                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 329                                                  RCL_CLIENT)) {
 330                         rc = tgt_ost_body_unpack(tsi, flags);
 331                         if (rc < 0)
 332                                 RETURN(rc);
 333                 }
 334         }
 335
 336         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 337                 RETURN(-EROFS);
 338
 339         if (flags & HAS_KEY) {
 340                 struct ldlm_request *dlm_req;
 341
 342                 LASSERT(h->th_fmt != NULL);
 343
 344                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 345                 if (dlm_req != NULL) {
 346                         union ldlm_wire_policy_data *policy =
 347                                         &dlm_req->lock_desc.l_policy_data;
 348
 349                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 350                                      LDLM_IBITS &&
 351                                      (policy->l_inodebits.bits |
 352                                       policy->l_inodebits.try_bits) == 0)) {
 353                                 /*
 354                                  * Lock without inodebits makes no sense and
 355                                  * will oops later in ldlm. If client miss to
 356                                  * set such bits, do not trigger ASSERTION.
 357                                  *
 358                                  * For liblustre flock case, it maybe zero.
 359                                  */
 360                                 rc = -EPROTO;
 361                         } else {
 362                                 tsi->tsi_dlm_req = dlm_req;
 363                         }
 364                 } else {
 365                         rc = -EFAULT;
 366                 }
 367         }
 368         tsi->tsi_preprocessed = 1;
 369         RETURN(rc);
 370 }
 371
 372 /*
 373  * Invoke handler for this request opc. Also do necessary preprocessing
 374  * (according to handler ->th_flags), and post-processing (setting of
 375  * ->last_{xid,committed}).
 376  */
 377 static int tgt_handle_request0(struct tgt_session_info *tsi,
 378                                struct tgt_handler *h,
 379                                struct ptlrpc_request *req)
 380 {
 381         int      serious = 0;
 382         int      rc;
 383         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 384
 385         ENTRY;
 386
 387
 388         /* When dealing with sec context requests, no export is associated yet,
 389          * because these requests are sent before *_CONNECT requests.
 390          * A NULL req->rq_export means the normal *_common_slice handlers will
 391          * not be called, because there is no reference to the target.
 392          * So deal with them by hand and jump directly to target_send_reply().
 393          */
 394         switch (opc) {
 395         case SEC_CTX_INIT:
 396         case SEC_CTX_INIT_CONT:
 397         case SEC_CTX_FINI:
 398                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 399                 GOTO(out, rc = 0);
 400         }
 401
 402         /*
 403          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 404          * to put same checks into handlers like mdt_close(), mdt_reint(),
 405          * etc., without talking to mdt authors first. Checking same thing
 406          * there again is useless and returning 0 error without packing reply
 407          * is buggy! Handlers either pack reply or return error.
 408          *
 409          * We return 0 here and do not send any reply in order to emulate
 410          * network failure. Do not send any reply in case any of NET related
 411          * fail_id has occured.
 412          */
 413         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 414                 RETURN(0);
 415         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 416                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 417                 RETURN(0);
 418
 419         rc = tgt_request_preprocess(tsi, h, req);
 420         /* pack reply if reply format is fixed */
 421         if (rc == 0 && h->th_flags & HAS_REPLY) {
 422                 /* Pack reply */
 423                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 424                                           RCL_SERVER))
 425                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 426                                              RCL_SERVER,
 427                                              tsi->tsi_mdt_body->mbo_eadatasize);
 428                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 429                                           RCL_SERVER))
 430                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 431                                              RCL_SERVER, 0);
 432                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 433                         req_capsule_set_size(tsi->tsi_pill,
 434                                              &RMF_ACL, RCL_SERVER,
 435                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 436
 437                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 438                                           RCL_SERVER)) {
 439                         struct niobuf_remote *remote_nb =
 440                                 req_capsule_client_get(tsi->tsi_pill,
 441                                                        &RMF_NIOBUF_REMOTE);
 442                         struct ost_body *body = tsi->tsi_ost_body;
 443
 444                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 445                                          RCL_SERVER,
 446                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 447                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 448                                          remote_nb[0].rnb_len : 0);
 449                 }
 450
 451                 rc = req_capsule_server_pack(tsi->tsi_pill);
 452         }
 453
 454         if (likely(rc == 0)) {
 455                 /*
 456                  * Process request, there can be two types of rc:
 457                  * 1) errors with msg unpack/pack, other failures outside the
 458                  * operation itself. This is counted as serious errors;
 459                  * 2) errors during fs operation, should be placed in rq_status
 460                  * only
 461                  */
 462                 rc = h->th_act(tsi);
 463                 if (!is_serious(rc) &&
 464                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 465                         DEBUG_REQ(D_ERROR, req,
 466                                   "%s: %s handler did not pack reply but returned no error",
 467                                   tgt_name(tsi->tsi_tgt), h->th_name);
 468                         LBUG();
 469                 }
 470                 serious = is_serious(rc);
 471                 rc = clear_serious(rc);
 472         } else {
 473                 serious = 1;
 474         }
 475
 476         req->rq_status = rc;
 477
 478         /*
 479          * ELDLM_* codes which > 0 should be in rq_status only as well as
 480          * all non-serious errors.
 481          */
 482         if (rc > 0 || !serious)
 483                 rc = 0;
 484
 485         LASSERT(current->journal_info == NULL);
 486
 487         if (likely(rc == 0 && req->rq_export))
 488                 target_committed_to_req(req);
 489
 490 out:
 491         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 492         RETURN(0);
 493 }
 494
 495 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 496                                        struct obd_device *obd, int *process)
 497 {
 498         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 499         case MDS_DISCONNECT:
 500         case OST_DISCONNECT:
 501         case OBD_IDX_READ:
 502                 *process = 1;
 503                 RETURN(0);
 504         case MDS_CLOSE:
 505         case MDS_SYNC: /* used in unmounting */
 506         case OBD_PING:
 507         case MDS_REINT:
 508         case OUT_UPDATE:
 509         case SEQ_QUERY:
 510         case FLD_QUERY:
 511         case FLD_READ:
 512         case LDLM_ENQUEUE:
 513         case OST_CREATE:
 514         case OST_DESTROY:
 515         case OST_PUNCH:
 516         case OST_SETATTR:
 517         case OST_SYNC:
 518         case OST_WRITE:
 519         case MDS_HSM_PROGRESS:
 520         case MDS_HSM_STATE_SET:
 521         case MDS_HSM_REQUEST:
 522                 *process = target_queue_recovery_request(req, obd);
 523                 RETURN(0);
 524
 525         default:
 526                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 527                 *process = -EAGAIN;
 528                 RETURN(0);
 529         }
 530 }
 531
 532 /*
 533  * Handle recovery. Return:
 534  *        +1: continue request processing;
 535  *       -ve: abort immediately with the given error code;
 536  *         0: send reply with error code in req->rq_status;
 537  */
 538 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 539 {
 540         ENTRY;
 541
 542         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 543         case MDS_CONNECT:
 544         case OST_CONNECT:
 545         case MGS_CONNECT:
 546         case SEC_CTX_INIT:
 547         case SEC_CTX_INIT_CONT:
 548         case SEC_CTX_FINI:
 549                 RETURN(+1);
 550         }
 551
 552         if (!req->rq_export->exp_obd->obd_replayable)
 553                 RETURN(+1);
 554
 555         /* sanity check: if the xid matches, the request must be marked as a
 556          * resent or replayed */
 557         if (req_can_reconstruct(req, NULL) == 1) {
 558                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 559                       (MSG_RESENT | MSG_REPLAY))) {
 560                         DEBUG_REQ(D_WARNING, req,
 561                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 562                                   req->rq_xid,
 563                                   lustre_msg_get_flags(req->rq_reqmsg));
 564                         req->rq_status = -ENOTCONN;
 565                         RETURN(-ENOTCONN);
 566                 }
 567         }
 568         /* else: note the opposite is not always true; a RESENT req after a
 569          * failover will usually not match the last_xid, since it was likely
 570          * never committed. A REPLAYed request will almost never match the
 571          * last xid, however it could for a committed, but still retained,
 572          * open. */
 573
 574         /* Check for aborted recovery... */
 575         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 576                 int rc;
 577                 int should_process;
 578
 579                 DEBUG_REQ(D_INFO, req, "Got new replay");
 580                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 581                                                  &should_process);
 582                 if (rc != 0 || !should_process)
 583                         RETURN(rc);
 584                 else if (should_process < 0) {
 585                         req->rq_status = should_process;
 586                         rc = ptlrpc_error(req);
 587                         RETURN(rc);
 588                 }
 589         }
 590         RETURN(+1);
 591 }
 592
 593 /* Initial check for request, it is validation mostly */
 594 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 595 {
 596         struct tgt_handler      *h;
 597         struct tgt_opc_slice    *s;
 598         struct lu_target        *tgt;
 599         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 600
 601         ENTRY;
 602
 603         tgt = class_exp2tgt(req->rq_export);
 604         if (unlikely(tgt == NULL)) {
 605                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 606                           class_exp2obd(req->rq_export)->obd_name);
 607                 RETURN(ERR_PTR(-EINVAL));
 608         }
 609
 610         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 611                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 612                         break;
 613
 614         /* opcode was not found in slice */
 615         if (unlikely(s->tos_hs == NULL)) {
 616                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 617                        opc);
 618                 RETURN(ERR_PTR(-ENOTSUPP));
 619         }
 620
 621         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 622         h = s->tos_hs + (opc - s->tos_opc_start);
 623         if (unlikely(h->th_opc == 0)) {
 624                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 625                 RETURN(ERR_PTR(-ENOTSUPP));
 626         }
 627
 628         RETURN(h);
 629 }
 630
 631 static int process_req_last_xid(struct ptlrpc_request *req)
 632 {
 633         __u64   last_xid;
 634         int rc = 0;
 635         struct obd_export *exp = req->rq_export;
 636         struct tg_export_data *ted = &exp->exp_target_data;
 637         bool need_lock = tgt_is_multimodrpcs_client(exp);
 638         ENTRY;
 639
 640         if (need_lock)
 641                 mutex_lock(&ted->ted_lcd_lock);
 642         /* check request's xid is consistent with export's last_xid */
 643         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 644         if (last_xid > exp->exp_last_xid)
 645                 exp->exp_last_xid = last_xid;
 646
 647         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 648                 /* Some request is allowed to be sent during replay,
 649                  * such as OUT update requests, FLD requests, so it
 650                  * is possible that replay requests has smaller XID
 651                  * than the exp_last_xid.
 652                  *
 653                  * Some non-replay requests may have smaller XID as
 654                  * well:
 655                  *
 656                  * - Client send a no_resend RPC, like statfs;
 657                  * - The RPC timedout (or some other error) on client,
 658                  *   then it's removed from the unreplied list;
 659                  * - Client send some other request to bump the
 660                  *   exp_last_xid on server;
 661                  * - The former RPC got chance to be processed;
 662                  */
 663                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 664                         rc = -EPROTO;
 665
 666                 DEBUG_REQ(D_WARNING, req,
 667                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 668                           req->rq_xid, exp->exp_last_xid, rc);
 669                 if (rc)
 670                         GOTO(out, rc);
 671         }
 672
 673         /* The "last_xid" is the minimum xid among unreplied requests,
 674          * if the request is from the previous connection, its xid can
 675          * still be larger than "exp_last_xid", then the above check of
 676          * xid is not enough to determine whether the request is delayed.
 677          *
 678          * For example, if some replay request was delayed and caused
 679          * timeout at client and the replay is restarted, the delayed
 680          * replay request will have the larger xid than "exp_last_xid"
 681          */
 682         if (req->rq_export->exp_conn_cnt >
 683             lustre_msg_get_conn_cnt(req->rq_reqmsg))
 684                 GOTO(out, rc = -ESTALE);
 685
 686         /* try to release in-memory reply data */
 687         if (tgt_is_multimodrpcs_client(exp)) {
 688                 tgt_handle_received_xid(exp, last_xid);
 689                 rc = tgt_handle_tag(req);
 690         }
 691
 692 out:
 693         if (need_lock)
 694                 mutex_unlock(&ted->ted_lcd_lock);
 695
 696         RETURN(rc);
 697 }
 698
 699 int tgt_request_handle(struct ptlrpc_request *req)
 700 {
 701         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 702
 703         struct lustre_msg       *msg = req->rq_reqmsg;
 704         struct tgt_handler      *h;
 705         struct lu_target        *tgt;
 706         int                      request_fail_id = 0;
 707         __u32                    opc = lustre_msg_get_opc(msg);
 708         struct obd_device       *obd;
 709         int                      rc;
 710         bool                     is_connect = false;
 711         ENTRY;
 712
 713         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 714                 if (cfs_fail_val == 0 &&
 715                     lustre_msg_get_opc(msg) != OBD_PING &&
 716                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 717                         cfs_fail_val = 1;
 718                         cfs_race_state = 0;
 719                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 720                 }
 721         }
 722
 723         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 724         tsi->tsi_pill = &req->rq_pill;
 725         tsi->tsi_env = req->rq_svc_thread->t_env;
 726
 727         /* if request has export then get handlers slice from corresponding
 728          * target, otherwise that should be connect operation */
 729         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 730             opc == MGS_CONNECT) {
 731                 is_connect = true;
 732                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 733                 rc = target_handle_connect(req);
 734                 if (rc != 0) {
 735                         rc = ptlrpc_error(req);
 736                         GOTO(out, rc);
 737                 }
 738                 /* recovery-small test 18c asks to drop connect reply */
 739                 if (unlikely(opc == OST_CONNECT &&
 740                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 741                         GOTO(out, rc = 0);
 742         }
 743
 744         if (unlikely(!class_connected_export(req->rq_export))) {
 745                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 746                     opc == SEC_CTX_FINI) {
 747                         /* sec context initialization has to be handled
 748                          * by hand in tgt_handle_request0() */
 749                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 750                         h = NULL;
 751                         GOTO(handle_recov, rc = 0);
 752                 }
 753                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 754                        opc, libcfs_id2str(req->rq_peer));
 755                 req->rq_status = -ENOTCONN;
 756                 rc = ptlrpc_error(req);
 757                 GOTO(out, rc);
 758         }
 759
 760         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 761         tsi->tsi_exp = req->rq_export;
 762         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 763                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 764         else
 765                 tsi->tsi_jobid = NULL;
 766
 767         if (tgt == NULL) {
 768                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 769                           class_exp2obd(req->rq_export)->obd_name);
 770                 req->rq_status = -EINVAL;
 771                 rc = ptlrpc_error(req);
 772                 GOTO(out, rc);
 773         }
 774
 775         /* Skip last_xid processing for the recovery thread, otherwise, the
 776          * last_xid on same request could be processed twice: first time when
 777          * processing the incoming request, second time when the request is
 778          * being processed by recovery thread. */
 779         obd = class_exp2obd(req->rq_export);
 780         if (is_connect) {
 781                 /* reset the exp_last_xid on each connection. */
 782                 req->rq_export->exp_last_xid = 0;
 783         } else if (obd->obd_recovery_data.trd_processing_task !=
 784                    current_pid()) {
 785                 rc = process_req_last_xid(req);
 786                 if (rc) {
 787                         req->rq_status = rc;
 788                         rc = ptlrpc_error(req);
 789                         GOTO(out, rc);
 790                 }
 791         }
 792
 793         request_fail_id = tgt->lut_request_fail_id;
 794         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 795
 796         h = tgt_handler_find_check(req);
 797         if (IS_ERR(h)) {
 798                 req->rq_status = PTR_ERR(h);
 799                 rc = ptlrpc_error(req);
 800                 GOTO(out, rc);
 801         }
 802
 803         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 804                  h->th_opc, opc);
 805
 806         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 807              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 808                 GOTO(out, rc = 0);
 809
 810         rc = lustre_msg_check_version(msg, h->th_version);
 811         if (unlikely(rc)) {
 812                 DEBUG_REQ(D_ERROR, req,
 813                           "%s: drop malformed request version=%08x expect=%08x",
 814                           tgt_name(tgt), lustre_msg_get_version(msg),
 815                           h->th_version);
 816                 req->rq_status = -EINVAL;
 817                 rc = ptlrpc_error(req);
 818                 GOTO(out, rc);
 819         }
 820
 821 handle_recov:
 822         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 823         if (likely(rc == 1)) {
 824                 rc = tgt_handle_request0(tsi, h, req);
 825                 if (rc)
 826                         GOTO(out, rc);
 827         }
 828         EXIT;
 829 out:
 830         req_capsule_fini(tsi->tsi_pill);
 831         if (tsi->tsi_corpus != NULL) {
 832                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 833                 tsi->tsi_corpus = NULL;
 834         }
 835         return rc;
 836 }
 837 EXPORT_SYMBOL(tgt_request_handle);
 838
 839 /** Assign high priority operations to the request if needed. */
 840 int tgt_hpreq_handler(struct ptlrpc_request *req)
 841 {
 842         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 843         struct tgt_handler      *h;
 844         int                      rc;
 845
 846         ENTRY;
 847
 848         if (req->rq_export == NULL)
 849                 RETURN(0);
 850
 851         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 852         tsi->tsi_pill = &req->rq_pill;
 853         tsi->tsi_env = req->rq_svc_thread->t_env;
 854         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 855         tsi->tsi_exp = req->rq_export;
 856
 857         h = tgt_handler_find_check(req);
 858         if (IS_ERR(h)) {
 859                 rc = PTR_ERR(h);
 860                 RETURN(rc);
 861         }
 862
 863         rc = tgt_request_preprocess(tsi, h, req);
 864         if (unlikely(rc != 0))
 865                 RETURN(rc);
 866
 867         if (h->th_hp != NULL)
 868                 h->th_hp(tsi);
 869         RETURN(0);
 870 }
 871 EXPORT_SYMBOL(tgt_hpreq_handler);
 872
 873 void tgt_counter_incr(struct obd_export *exp, int opcode)
 874 {
 875         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 876         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 877                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 878 }
 879 EXPORT_SYMBOL(tgt_counter_incr);
 880
 881 /*
 882  * Unified target generic handlers.
 883  */
 884
 885 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 886 {
 887         struct lu_target        *tgt = class_exp2tgt(exp);
 888         struct sptlrpc_flavor    flvr;
 889         int                      rc = 0;
 890
 891         LASSERT(tgt);
 892         LASSERT(tgt->lut_obd);
 893         LASSERT(tgt->lut_slice);
 894
 895         /* always allow ECHO client */
 896         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 897                             LUSTRE_ECHO_NAME) == 0)) {
 898                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 899                 return 0;
 900         }
 901
 902         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 903                 read_lock(&tgt->lut_sptlrpc_lock);
 904                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 905                                              req->rq_sp_from,
 906                                              req->rq_peer.nid,
 907                                              &flvr);
 908                 read_unlock(&tgt->lut_sptlrpc_lock);
 909
 910                 spin_lock(&exp->exp_lock);
 911                 exp->exp_sp_peer = req->rq_sp_from;
 912                 exp->exp_flvr = flvr;
 913
 914                 /* when on mgs, if no restriction is set, or if client
 915                  * is loopback, allow any flavor */
 916                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 917                            LUSTRE_MGS_NAME) == 0) &&
 918                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 919                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 920                       == LOLND))
 921                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 922
 923                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 924                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 925                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 926                                "expect %x\n", tgt_name(tgt),
 927                                req->rq_flvr.sf_rpc,
 928                                libcfs_nid2str(req->rq_peer.nid),
 929                                exp->exp_flvr.sf_rpc);
 930                         rc = -EACCES;
 931                 }
 932                 spin_unlock(&exp->exp_lock);
 933         } else {
 934                 if (exp->exp_sp_peer != req->rq_sp_from) {
 935                         CERROR("%s: RPC source %s doesn't match %s\n",
 936                                tgt_name(tgt),
 937                                sptlrpc_part2name(req->rq_sp_from),
 938                                sptlrpc_part2name(exp->exp_sp_peer));
 939                         rc = -EACCES;
 940                 } else {
 941                         rc = sptlrpc_target_export_check(exp, req);
 942                 }
 943         }
 944
 945         return rc;
 946 }
 947
 948 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 949 {
 950         struct sptlrpc_rule_set  tmp_rset;
 951         int                      rc;
 952
 953         if (unlikely(tgt == NULL)) {
 954                 CERROR("No target passed");
 955                 return -EINVAL;
 956         }
 957
 958         sptlrpc_rule_set_init(&tmp_rset);
 959         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 960         if (rc) {
 961                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 962                        tgt_name(tgt), rc);
 963                 return rc;
 964         }
 965
 966         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 967
 968         write_lock(&tgt->lut_sptlrpc_lock);
 969         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 970         tgt->lut_sptlrpc_rset = tmp_rset;
 971         write_unlock(&tgt->lut_sptlrpc_lock);
 972
 973         return 0;
 974 }
 975 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 976
 977 int tgt_connect(struct tgt_session_info *tsi)
 978 {
 979         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 980         struct obd_connect_data *reply;
 981         int                      rc;
 982
 983         ENTRY;
 984
 985         /* XXX: better to call this check right after getting new export but
 986          * before last_rcvd slot allocation to avoid server load upon insecure
 987          * connects. This is to be fixed after unifiyng all targets.
 988          */
 989         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 990         if (rc)
 991                 GOTO(out, rc);
 992
 993         /* To avoid exposing partially initialized connection flags, changes up
 994          * to this point have been staged in reply->ocd_connect_flags. Now that
 995          * connection handling has completed successfully, atomically update
 996          * the connect flags in the shared export data structure. LU-1623 */
 997         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
 998         spin_lock(&tsi->tsi_exp->exp_lock);
 999         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1000         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1001                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1002                         reply->ocd_connect_flags2;
1003         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1004         spin_unlock(&tsi->tsi_exp->exp_lock);
1005
1006         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1007                    LUSTRE_MDT_NAME) == 0) {
1008                 rc = req_check_sepol(tsi->tsi_pill);
1009                 if (rc)
1010                         GOTO(out, rc);
1011         }
1012
1013         RETURN(0);
1014 out:
1015         obd_disconnect(class_export_get(tsi->tsi_exp));
1016         return rc;
1017 }
1018 EXPORT_SYMBOL(tgt_connect);
1019
1020 int tgt_disconnect(struct tgt_session_info *tsi)
1021 {
1022         int rc;
1023
1024         ENTRY;
1025
1026         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1027
1028         rc = target_handle_disconnect(tgt_ses_req(tsi));
1029         if (rc)
1030                 RETURN(err_serious(rc));
1031
1032         RETURN(rc);
1033 }
1034 EXPORT_SYMBOL(tgt_disconnect);
1035
1036 /*
1037  * Unified target OBD handlers
1038  */
1039 int tgt_obd_ping(struct tgt_session_info *tsi)
1040 {
1041         int rc;
1042
1043         ENTRY;
1044
1045         /* The target-specific part of OBD_PING request handling.
1046          * It controls Filter Modification Data (FMD) expiration each time
1047          * PING is received.
1048          *
1049          * Valid only for replayable targets, e.g. MDT and OFD
1050          */
1051         if (tsi->tsi_exp->exp_obd->obd_replayable)
1052                 tgt_fmd_expire(tsi->tsi_exp);
1053
1054         rc = req_capsule_server_pack(tsi->tsi_pill);
1055         if (rc)
1056                 RETURN(err_serious(rc));
1057
1058         RETURN(rc);
1059 }
1060 EXPORT_SYMBOL(tgt_obd_ping);
1061
1062 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1063 {
1064         return err_serious(-EOPNOTSUPP);
1065 }
1066
1067 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1068 {
1069         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1070         struct obd_export       *exp = req->rq_export;
1071         struct ptlrpc_bulk_desc *desc;
1072         int                      i;
1073         int                      rc;
1074
1075         ENTRY;
1076
1077         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1078                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1079                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1080         if (desc == NULL)
1081                 RETURN(-ENOMEM);
1082
1083         for (i = 0; i < rdbuf->rb_nbufs; i++)
1084                 desc->bd_frag_ops->add_iov_frag(desc,
1085                                         rdbuf->rb_bufs[i].lb_buf,
1086                                         rdbuf->rb_bufs[i].lb_len);
1087
1088         rc = target_bulk_io(exp, desc);
1089         ptlrpc_free_bulk(desc);
1090         RETURN(rc);
1091 }
1092 EXPORT_SYMBOL(tgt_send_buffer);
1093
1094 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1095 {
1096         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1097         struct obd_export       *exp = req->rq_export;
1098         struct ptlrpc_bulk_desc *desc;
1099         int                      tmpcount;
1100         int                      tmpsize;
1101         int                      i;
1102         int                      rc;
1103
1104         ENTRY;
1105
1106         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1107                                     PTLRPC_BULK_PUT_SOURCE |
1108                                         PTLRPC_BULK_BUF_KIOV,
1109                                     MDS_BULK_PORTAL,
1110                                     &ptlrpc_bulk_kiov_pin_ops);
1111         if (desc == NULL)
1112                 RETURN(-ENOMEM);
1113
1114         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1115                 /* old client requires reply size in it's PAGE_SIZE,
1116                  * which is rdpg->rp_count */
1117                 nob = rdpg->rp_count;
1118
1119         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1120              i++, tmpcount -= tmpsize) {
1121                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1122                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1123                                                  tmpsize);
1124         }
1125
1126         LASSERT(desc->bd_nob == nob);
1127         rc = target_bulk_io(exp, desc);
1128         ptlrpc_free_bulk(desc);
1129         RETURN(rc);
1130 }
1131 EXPORT_SYMBOL(tgt_sendpage);
1132
1133 /*
1134  * OBD_IDX_READ handler
1135  */
1136 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1137 {
1138         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1139         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1140         struct idx_info         *req_ii, *rep_ii;
1141         int                      rc, i;
1142
1143         ENTRY;
1144
1145         memset(rdpg, 0, sizeof(*rdpg));
1146         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1147
1148         /* extract idx_info buffer from request & reply */
1149         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1150         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1151                 RETURN(err_serious(-EPROTO));
1152
1153         rc = req_capsule_server_pack(tsi->tsi_pill);
1154         if (rc)
1155                 RETURN(err_serious(rc));
1156
1157         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1158         if (rep_ii == NULL)
1159                 RETURN(err_serious(-EFAULT));
1160         rep_ii->ii_magic = IDX_INFO_MAGIC;
1161
1162         /* extract hash to start with */
1163         rdpg->rp_hash = req_ii->ii_hash_start;
1164
1165         /* extract requested attributes */
1166         rdpg->rp_attrs = req_ii->ii_attrs;
1167
1168         /* check that fid packed in request is valid and supported */
1169         if (!fid_is_sane(&req_ii->ii_fid))
1170                 RETURN(-EINVAL);
1171         rep_ii->ii_fid = req_ii->ii_fid;
1172
1173         /* copy flags */
1174         rep_ii->ii_flags = req_ii->ii_flags;
1175
1176         /* compute number of pages to allocate, ii_count is the number of 4KB
1177          * containers */
1178         if (req_ii->ii_count <= 0)
1179                 GOTO(out, rc = -EFAULT);
1180         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1181                                exp_max_brw_size(tsi->tsi_exp));
1182         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1183
1184         /* allocate pages to store the containers */
1185         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1186         if (rdpg->rp_pages == NULL)
1187                 GOTO(out, rc = -ENOMEM);
1188         for (i = 0; i < rdpg->rp_npages; i++) {
1189                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1190                 if (rdpg->rp_pages[i] == NULL)
1191                         GOTO(out, rc = -ENOMEM);
1192         }
1193
1194         /* populate pages with key/record pairs */
1195         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1196         if (rc < 0)
1197                 GOTO(out, rc);
1198
1199         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1200                  "asked %d > %d\n", rc, rdpg->rp_count);
1201
1202         /* send pages to client */
1203         rc = tgt_sendpage(tsi, rdpg, rc);
1204         if (rc)
1205                 GOTO(out, rc);
1206         EXIT;
1207 out:
1208         if (rdpg->rp_pages) {
1209                 for (i = 0; i < rdpg->rp_npages; i++)
1210                         if (rdpg->rp_pages[i])
1211                                 __free_page(rdpg->rp_pages[i]);
1212                 OBD_FREE(rdpg->rp_pages,
1213                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1214         }
1215         return rc;
1216 }
1217
1218 struct tgt_handler tgt_obd_handlers[] = {
1219 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1220 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1221 };
1222 EXPORT_SYMBOL(tgt_obd_handlers);
1223
1224 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1225              struct dt_object *obj, __u64 start, __u64 end)
1226 {
1227         int rc = 0;
1228
1229         ENTRY;
1230
1231         /* if no objid is specified, it means "sync whole filesystem" */
1232         if (obj == NULL) {
1233                 rc = dt_sync(env, tgt->lut_bottom);
1234         } else if (dt_version_get(env, obj) >
1235                    tgt->lut_obd->obd_last_committed) {
1236                 rc = dt_object_sync(env, obj, start, end);
1237         }
1238         atomic_inc(&tgt->lut_sync_count);
1239
1240         RETURN(rc);
1241 }
1242 EXPORT_SYMBOL(tgt_sync);
1243 /*
1244  * Unified target DLM handlers.
1245  */
1246
1247 /**
1248  * Unified target BAST
1249  *
1250  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1251  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1252  * but if it's cross-MDT lock, because remote object version is not set, a
1253  * filesystem sync is needed.
1254  *
1255  * \param lock server side lock
1256  * \param desc lock desc
1257  * \param data ldlm_cb_set_arg
1258  * \param flag  indicates whether this cancelling or blocking callback
1259  * \retval      0 on success
1260  * \retval      negative number on error
1261  */
1262 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1263                             void *data, int flag)
1264 {
1265         struct lu_env            env;
1266         struct lu_target        *tgt;
1267         struct dt_object        *obj = NULL;
1268         struct lu_fid            fid;
1269         int                      rc = 0;
1270
1271         ENTRY;
1272
1273         tgt = class_exp2tgt(lock->l_export);
1274
1275         if (unlikely(tgt == NULL)) {
1276                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1277                        class_exp2obd(lock->l_export)->obd_name);
1278                 RETURN(-EINVAL);
1279         }
1280
1281         if (flag == LDLM_CB_CANCELING &&
1282             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1283             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1284              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1285               ldlm_is_cbpending(lock))) &&
1286             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1287              lock->l_resource->lr_type == LDLM_EXTENT)) {
1288                 __u64 start = 0;
1289                 __u64 end = OBD_OBJECT_EOF;
1290
1291                 rc = lu_env_init(&env, LCT_DT_THREAD);
1292                 if (unlikely(rc != 0))
1293                         GOTO(err, rc);
1294
1295                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1296                                    tgt->lut_lsd.lsd_osd_index);
1297
1298                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1299                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1300                         if (IS_ERR(obj))
1301                                 GOTO(err_env, rc = PTR_ERR(obj));
1302
1303                         if (!dt_object_exists(obj))
1304                                 GOTO(err_put, rc = -ENOENT);
1305
1306                         start = lock->l_policy_data.l_extent.start;
1307                         end = lock->l_policy_data.l_extent.end;
1308                 }
1309
1310                 rc = tgt_sync(&env, tgt, obj, start, end);
1311                 if (rc < 0) {
1312                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1313                                "cancel: rc = %d\n",
1314                                tgt_name(tgt), PFID(&fid),
1315                                lock->l_policy_data.l_extent.start,
1316                                lock->l_policy_data.l_extent.end, rc);
1317                 }
1318 err_put:
1319                 if (obj != NULL)
1320                         dt_object_put(&env, obj);
1321 err_env:
1322                 lu_env_fini(&env);
1323         }
1324 err:
1325         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1326         RETURN(rc);
1327 }
1328
1329 static struct ldlm_callback_suite tgt_dlm_cbs = {
1330         .lcs_completion = ldlm_server_completion_ast,
1331         .lcs_blocking   = tgt_blocking_ast,
1332         .lcs_glimpse    = ldlm_server_glimpse_ast
1333 };
1334
1335 int tgt_enqueue(struct tgt_session_info *tsi)
1336 {
1337         struct ptlrpc_request *req = tgt_ses_req(tsi);
1338         int rc;
1339
1340         ENTRY;
1341         /*
1342          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1343          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1344          */
1345         LASSERT(tsi->tsi_dlm_req != NULL);
1346         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1347                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1348         if (rc)
1349                 RETURN(err_serious(rc));
1350
1351         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1352         case LUT_FAIL_MDT:
1353                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1354                 break;
1355         case LUT_FAIL_OST:
1356                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1357                 break;
1358         case LUT_FAIL_MGT:
1359                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1360                 break;
1361         default:
1362                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1363                 break;
1364         }
1365         RETURN(req->rq_status);
1366 }
1367 EXPORT_SYMBOL(tgt_enqueue);
1368
1369 int tgt_convert(struct tgt_session_info *tsi)
1370 {
1371         struct ptlrpc_request *req = tgt_ses_req(tsi);
1372         int rc;
1373
1374         ENTRY;
1375         LASSERT(tsi->tsi_dlm_req);
1376         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1377         if (rc)
1378                 RETURN(err_serious(rc));
1379
1380         RETURN(req->rq_status);
1381 }
1382
1383 int tgt_bl_callback(struct tgt_session_info *tsi)
1384 {
1385         return err_serious(-EOPNOTSUPP);
1386 }
1387
1388 int tgt_cp_callback(struct tgt_session_info *tsi)
1389 {
1390         return err_serious(-EOPNOTSUPP);
1391 }
1392
1393 /* generic LDLM target handler */
1394 struct tgt_handler tgt_dlm_handlers[] = {
1395 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1396 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1397 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1398 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1399 };
1400 EXPORT_SYMBOL(tgt_dlm_handlers);
1401
1402 /*
1403  * Unified target LLOG handlers.
1404  */
1405 int tgt_llog_open(struct tgt_session_info *tsi)
1406 {
1407         int rc;
1408
1409         ENTRY;
1410
1411         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1412
1413         RETURN(rc);
1414 }
1415 EXPORT_SYMBOL(tgt_llog_open);
1416
1417 int tgt_llog_read_header(struct tgt_session_info *tsi)
1418 {
1419         int rc;
1420
1421         ENTRY;
1422
1423         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1424
1425         RETURN(rc);
1426 }
1427 EXPORT_SYMBOL(tgt_llog_read_header);
1428
1429 int tgt_llog_next_block(struct tgt_session_info *tsi)
1430 {
1431         int rc;
1432
1433         ENTRY;
1434
1435         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1436
1437         RETURN(rc);
1438 }
1439 EXPORT_SYMBOL(tgt_llog_next_block);
1440
1441 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1442 {
1443         int rc;
1444
1445         ENTRY;
1446
1447         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1448
1449         RETURN(rc);
1450 }
1451 EXPORT_SYMBOL(tgt_llog_prev_block);
1452
1453 /* generic llog target handler */
1454 struct tgt_handler tgt_llog_handlers[] = {
1455 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1456 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1457 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1458 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1459 };
1460 EXPORT_SYMBOL(tgt_llog_handlers);
1461
1462 /*
1463  * sec context handlers
1464  */
1465 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1466 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1467 {
1468         return 0;
1469 }
1470
1471 struct tgt_handler tgt_sec_ctx_handlers[] = {
1472 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1473 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1474 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1475 };
1476 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1477
1478 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1479                                  struct dt_device *key,
1480                                  struct lfsck_req_local *lrl,
1481                                  struct thandle *th) = NULL;
1482
1483 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1484                                                       struct dt_device *,
1485                                                       struct lfsck_req_local *,
1486                                                       struct thandle *))
1487 {
1488         tgt_lfsck_in_notify_local = notify;
1489 }
1490 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1491
1492 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1493                            struct dt_device *key,
1494                            struct lfsck_request *lr) = NULL;
1495
1496 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1497                                                 struct dt_device *,
1498                                                 struct lfsck_request *))
1499 {
1500         tgt_lfsck_in_notify = notify;
1501 }
1502 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1503
1504 static int (*tgt_lfsck_query)(const struct lu_env *env,
1505                               struct dt_device *key,
1506                               struct lfsck_request *req,
1507                               struct lfsck_reply *rep,
1508                               struct lfsck_query *que) = NULL;
1509
1510 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1511                                            struct dt_device *,
1512                                            struct lfsck_request *,
1513                                            struct lfsck_reply *,
1514                                            struct lfsck_query *))
1515 {
1516         tgt_lfsck_query = query;
1517 }
1518 EXPORT_SYMBOL(tgt_register_lfsck_query);
1519
1520 /* LFSCK request handlers */
1521 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1522 {
1523         const struct lu_env     *env = tsi->tsi_env;
1524         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1525         struct lfsck_request    *lr;
1526         int                      rc;
1527         ENTRY;
1528
1529         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1530         if (lr == NULL)
1531                 RETURN(-EPROTO);
1532
1533         rc = tgt_lfsck_in_notify(env, key, lr);
1534
1535         RETURN(rc);
1536 }
1537
1538 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1539 {
1540         struct lfsck_request    *request;
1541         struct lfsck_reply      *reply;
1542         int                      rc;
1543         ENTRY;
1544
1545         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1546         if (request == NULL)
1547                 RETURN(-EPROTO);
1548
1549         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1550         if (reply == NULL)
1551                 RETURN(-ENOMEM);
1552
1553         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1554                              request, reply, NULL);
1555
1556         RETURN(rc < 0 ? rc : 0);
1557 }
1558
1559 struct tgt_handler tgt_lfsck_handlers[] = {
1560 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1561 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1562 };
1563 EXPORT_SYMBOL(tgt_lfsck_handlers);
1564
1565 /*
1566  * initialize per-thread page pool (bug 5137).
1567  */
1568 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1569 {
1570         struct tgt_thread_big_cache *tbc;
1571
1572         ENTRY;
1573
1574         LASSERT(thread != NULL);
1575         LASSERT(thread->t_data == NULL);
1576
1577         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1578         if (tbc == NULL)
1579                 RETURN(-ENOMEM);
1580         thread->t_data = tbc;
1581         RETURN(0);
1582 }
1583 EXPORT_SYMBOL(tgt_io_thread_init);
1584
1585 /*
1586  * free per-thread pool created by tgt_thread_init().
1587  */
1588 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1589 {
1590         struct tgt_thread_big_cache *tbc;
1591
1592         ENTRY;
1593
1594         LASSERT(thread != NULL);
1595
1596         /*
1597          * be prepared to handle partially-initialized pools (because this is
1598          * called from ost_io_thread_init() for cleanup.
1599          */
1600         tbc = thread->t_data;
1601         if (tbc != NULL) {
1602                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1603                 thread->t_data = NULL;
1604         }
1605         EXIT;
1606 }
1607 EXPORT_SYMBOL(tgt_io_thread_done);
1608
1609 /**
1610  * Helper function for getting Data-on-MDT file server DLM lock
1611  * if asked by client.
1612  */
1613 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1614                       struct lustre_handle *lh, int mode, __u64 *flags)
1615 {
1616         union ldlm_policy_data policy = {
1617                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1618         };
1619         int rc;
1620
1621         ENTRY;
1622
1623         LASSERT(lh != NULL);
1624         LASSERT(ns != NULL);
1625         LASSERT(!lustre_handle_is_used(lh));
1626
1627         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1628                                     flags, ldlm_blocking_ast,
1629                                     ldlm_completion_ast, ldlm_glimpse_ast,
1630                                     NULL, 0, LVB_T_NONE, NULL, lh);
1631
1632         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1633 }
1634 EXPORT_SYMBOL(tgt_mdt_data_lock);
1635
1636 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1637 {
1638         LASSERT(lustre_handle_is_used(lh));
1639         ldlm_lock_decref(lh, mode);
1640 }
1641 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1642
1643 /**
1644  * Helper function for getting server side [start, start+count] DLM lock
1645  * if asked by client.
1646  */
1647 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1648                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1649                     struct lustre_handle *lh, int mode, __u64 *flags)
1650 {
1651         union ldlm_policy_data policy;
1652         int rc;
1653
1654         ENTRY;
1655
1656         LASSERT(lh != NULL);
1657         LASSERT(ns != NULL);
1658         LASSERT(!lustre_handle_is_used(lh));
1659
1660         policy.l_extent.gid = 0;
1661         policy.l_extent.start = start & PAGE_MASK;
1662
1663         /*
1664          * If ->o_blocks is EOF it means "lock till the end of the file".
1665          * Otherwise, it's size of an extent or hole being punched (in bytes).
1666          */
1667         if (end == OBD_OBJECT_EOF || end < start)
1668                 policy.l_extent.end = OBD_OBJECT_EOF;
1669         else
1670                 policy.l_extent.end = end | ~PAGE_MASK;
1671
1672         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1673                                     mode, flags, ldlm_blocking_ast,
1674                                     ldlm_completion_ast, ldlm_glimpse_ast,
1675                                     NULL, 0, LVB_T_NONE, NULL, lh);
1676         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1677 }
1678 EXPORT_SYMBOL(tgt_extent_lock);
1679
1680 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1681 {
1682         LASSERT(lustre_handle_is_used(lh));
1683         ldlm_lock_decref(lh, mode);
1684 }
1685 EXPORT_SYMBOL(tgt_extent_unlock);
1686
1687 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1688                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1689                         struct niobuf_remote *nb, struct lustre_handle *lh,
1690                         enum ldlm_mode mode)
1691 {
1692         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1693         __u64                    flags = 0;
1694         int                      nrbufs = obj->ioo_bufcnt;
1695         int                      i;
1696         int                      rc;
1697
1698         ENTRY;
1699
1700         LASSERT(mode == LCK_PR || mode == LCK_PW);
1701         LASSERT(!lustre_handle_is_used(lh));
1702
1703         if (ns->ns_obd->obd_recovering)
1704                 RETURN(0);
1705
1706         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1707                 RETURN(0);
1708
1709         for (i = 1; i < nrbufs; i++)
1710                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1711                         RETURN(-EFAULT);
1712
1713         /* MDT IO for data-on-mdt */
1714         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1715                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1716         else
1717                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1718                                      nb[nrbufs - 1].rnb_offset +
1719                                      nb[nrbufs - 1].rnb_len - 1,
1720                                      lh, mode, &flags);
1721         RETURN(rc);
1722 }
1723
1724 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1725                            struct lustre_handle *lh, enum ldlm_mode mode)
1726 {
1727         ENTRY;
1728
1729         LASSERT(mode == LCK_PR || mode == LCK_PW);
1730         LASSERT((obj->ioo_bufcnt > 0 &&
1731                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1732                 lustre_handle_is_used(lh));
1733
1734         if (lustre_handle_is_used(lh))
1735                 tgt_extent_unlock(lh, mode);
1736         EXIT;
1737 }
1738 static int tgt_checksum_niobuf(struct lu_target *tgt,
1739                                  struct niobuf_local *local_nb, int npages,
1740                                  int opc, enum cksum_types cksum_type,
1741                                  __u32 *cksum)
1742 {
1743         struct ahash_request           *req;
1744         unsigned int                    bufsize;
1745         int                             i, err;
1746         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1747
1748         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1749         if (IS_ERR(req)) {
1750                 CERROR("%s: unable to initialize checksum hash %s\n",
1751                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1752                 return PTR_ERR(req);
1753         }
1754
1755         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1756         for (i = 0; i < npages; i++) {
1757                 /* corrupt the data before we compute the checksum, to
1758                  * simulate a client->OST data error */
1759                 if (i == 0 && opc == OST_WRITE &&
1760                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1761                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1762                         int len = local_nb[i].lnb_len;
1763                         struct page *np = tgt_page_to_corrupt;
1764
1765                         if (np) {
1766                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1767                                                         KM_USER0);
1768                                 char *ptr2 = page_address(np);
1769
1770                                 memcpy(ptr2 + off, ptr + off, len);
1771                                 memcpy(ptr2 + off, "bad3", min(4, len));
1772                                 ll_kunmap_atomic(ptr, KM_USER0);
1773
1774                                 /* LU-8376 to preserve original index for
1775                                  * display in dump_all_bulk_pages() */
1776                                 np->index = i;
1777
1778                                 cfs_crypto_hash_update_page(req, np, off,
1779                                                             len);
1780                                 continue;
1781                         } else {
1782                                 CERROR("%s: can't alloc page for corruption\n",
1783                                        tgt_name(tgt));
1784                         }
1785                 }
1786                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1787                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1788                                   local_nb[i].lnb_len);
1789
1790                  /* corrupt the data after we compute the checksum, to
1791                  * simulate an OST->client data error */
1792                 if (i == 0 && opc == OST_READ &&
1793                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1794                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1795                         int len = local_nb[i].lnb_len;
1796                         struct page *np = tgt_page_to_corrupt;
1797
1798                         if (np) {
1799                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1800                                                         KM_USER0);
1801                                 char *ptr2 = page_address(np);
1802
1803                                 memcpy(ptr2 + off, ptr + off, len);
1804                                 memcpy(ptr2 + off, "bad4", min(4, len));
1805                                 ll_kunmap_atomic(ptr, KM_USER0);
1806
1807                                 /* LU-8376 to preserve original index for
1808                                  * display in dump_all_bulk_pages() */
1809                                 np->index = i;
1810
1811                                 cfs_crypto_hash_update_page(req, np, off,
1812                                                             len);
1813                                 continue;
1814                         } else {
1815                                 CERROR("%s: can't alloc page for corruption\n",
1816                                        tgt_name(tgt));
1817                         }
1818                 }
1819         }
1820
1821         bufsize = sizeof(*cksum);
1822         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1823
1824         return 0;
1825 }
1826
1827 char dbgcksum_file_name[PATH_MAX];
1828
1829 static void dump_all_bulk_pages(struct obdo *oa, int count,
1830                                 struct niobuf_local *local_nb,
1831                                 __u32 server_cksum, __u32 client_cksum)
1832 {
1833         struct file *filp;
1834         int rc, i;
1835         unsigned int len;
1836         char *buf;
1837
1838         /* will only keep dump of pages on first error for the same range in
1839          * file/fid, not during the resends/retries. */
1840         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1841                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1842                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1843                   libcfs_debug_file_path_arr :
1844                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1845                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1846                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1847                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1848                  local_nb[0].lnb_file_offset,
1849                  local_nb[count-1].lnb_file_offset +
1850                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1851         filp = filp_open(dbgcksum_file_name,
1852                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1853         if (IS_ERR(filp)) {
1854                 rc = PTR_ERR(filp);
1855                 if (rc == -EEXIST)
1856                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1857                                "checksum error: rc = %d\n", dbgcksum_file_name,
1858                                rc);
1859                 else
1860                         CERROR("%s: can't open to dump pages with checksum "
1861                                "error: rc = %d\n", dbgcksum_file_name, rc);
1862                 return;
1863         }
1864
1865         for (i = 0; i < count; i++) {
1866                 len = local_nb[i].lnb_len;
1867                 buf = kmap(local_nb[i].lnb_page);
1868                 while (len != 0) {
1869                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1870                         if (rc < 0) {
1871                                 CERROR("%s: wanted to write %u but got %d "
1872                                        "error\n", dbgcksum_file_name, len, rc);
1873                                 break;
1874                         }
1875                         len -= rc;
1876                         buf += rc;
1877                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1878                                dbgcksum_file_name, rc);
1879                 }
1880                 kunmap(local_nb[i].lnb_page);
1881         }
1882
1883         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1884         if (rc)
1885                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1886         filp_close(filp, NULL);
1887 }
1888
1889 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1890                                struct obd_export *exp, struct obdo *oa,
1891                                const struct lnet_process_id *peer,
1892                                __u32 client_cksum, __u32 server_cksum,
1893                                enum cksum_types server_cksum_type)
1894 {
1895         char *msg;
1896         enum cksum_types cksum_type;
1897         loff_t start, end;
1898
1899         /* unlikely to happen and only if resend does not occur due to cksum
1900          * control failure on Client */
1901         if (unlikely(server_cksum == client_cksum)) {
1902                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1903                        client_cksum);
1904                 return 0;
1905         }
1906
1907         if (exp->exp_obd->obd_checksum_dump)
1908                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1909                                     client_cksum);
1910
1911         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1912                                            oa->o_flags : 0);
1913
1914         if (cksum_type != server_cksum_type)
1915                 msg = "the server may have not used the checksum type specified"
1916                       " in the original request - likely a protocol problem";
1917         else
1918                 msg = "should have changed on the client or in transit";
1919
1920         start = local_nb[0].lnb_file_offset;
1921         end = local_nb[npages-1].lnb_file_offset +
1922                                         local_nb[npages-1].lnb_len - 1;
1923
1924         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1925                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1926                 " %x (type %x), server csum %x (type %x)\n",
1927                 exp->exp_obd->obd_name,
1928                 msg, libcfs_nid2str(peer->nid),
1929                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1930                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1931                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1932                 POSTID(&oa->o_oi),
1933                 start, end, client_cksum, cksum_type, server_cksum,
1934                 server_cksum_type);
1935
1936         return 1;
1937 }
1938
1939 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1940                              unsigned char *buf, int size)
1941 {
1942         int     i, off, len, copied = size;
1943         char    *ptr;
1944
1945         for (i = 0; i < npages; i++) {
1946                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1947                 len = local[i].lnb_len;
1948
1949                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1950                        i, off, len, size);
1951                 if (len > size)
1952                         return -EINVAL;
1953
1954                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1955                 memcpy(buf + off, ptr, len);
1956                 ll_kunmap_atomic(ptr, KM_USER0);
1957                 buf += len;
1958                 size -= len;
1959         }
1960         return copied - size;
1961 }
1962
1963 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1964                                      struct niobuf_local *local_nb,
1965                                      int npages, int opc,
1966                                      obd_dif_csum_fn *fn,
1967                                      int sector_size,
1968                                      u32 *check_sum)
1969 {
1970         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1971         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1972         const char *obd_name = tgt->lut_obd->obd_name;
1973         struct ahash_request *req;
1974         unsigned int bufsize;
1975         unsigned char *buffer;
1976         struct page *__page;
1977         __u16 *guard_start;
1978         int guard_number;
1979         int used_number = 0;
1980         __u32 cksum;
1981         int rc = 0;
1982         int used;
1983         int i;
1984
1985         __page = alloc_page(GFP_KERNEL);
1986         if (__page == NULL)
1987                 return -ENOMEM;
1988
1989         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1990         if (IS_ERR(req)) {
1991                 CERROR("%s: unable to initialize checksum hash %s\n",
1992                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1993                 return PTR_ERR(req);
1994         }
1995
1996         buffer = kmap(__page);
1997         guard_start = (__u16 *)buffer;
1998         guard_number = PAGE_SIZE / sizeof(*guard_start);
1999         for (i = 0; i < npages; i++) {
2000                 /* corrupt the data before we compute the checksum, to
2001                  * simulate a client->OST data error */
2002                 if (i == 0 && opc == OST_WRITE &&
2003                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2004                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2005                         int len = local_nb[i].lnb_len;
2006                         struct page *np = tgt_page_to_corrupt;
2007
2008                         if (np) {
2009                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2010                                                         KM_USER0);
2011                                 char *ptr2 = page_address(np);
2012
2013                                 memcpy(ptr2 + off, ptr + off, len);
2014                                 memcpy(ptr2 + off, "bad3", min(4, len));
2015                                 ll_kunmap_atomic(ptr, KM_USER0);
2016
2017                                 /* LU-8376 to preserve original index for
2018                                  * display in dump_all_bulk_pages() */
2019                                 np->index = i;
2020
2021                                 cfs_crypto_hash_update_page(req, np, off,
2022                                                             len);
2023                                 continue;
2024                         } else {
2025                                 CERROR("%s: can't alloc page for corruption\n",
2026                                        tgt_name(tgt));
2027                         }
2028                 }
2029
2030                 /*
2031                  * The left guard number should be able to hold checksums of a
2032                  * whole page
2033                  */
2034                 if (t10_cksum_type && opc == OST_READ &&
2035                     local_nb[i].lnb_guard_disk) {
2036                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2037                         if (used > (guard_number - used_number)) {
2038                                 rc = -E2BIG;
2039                                 break;
2040                         }
2041                         memcpy(guard_start + used_number,
2042                                local_nb[i].lnb_guards,
2043                                used * sizeof(*local_nb[i].lnb_guards));
2044                 } else {
2045                         rc = obd_page_dif_generate_buffer(obd_name,
2046                                 local_nb[i].lnb_page,
2047                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2048                                 local_nb[i].lnb_len, guard_start + used_number,
2049                                 guard_number - used_number, &used, sector_size,
2050                                 fn);
2051                         if (rc)
2052                                 break;
2053                 }
2054
2055                 LASSERT(used <= MAX_GUARD_NUMBER);
2056                 /*
2057                  * If disk support T10PI checksum, copy guards to local_nb.
2058                  * If the write is partial page, do not use the guards for bio
2059                  * submission since the data might not be full-sector. The bio
2060                  * guards will be generated later based on the full sectors. If
2061                  * the sector size is 512B rather than 4 KB, or the page size
2062                  * is larger than 4KB, this might drop some useful guards for
2063                  * partial page write, but it will only add minimal extra time
2064                  * of checksum calculation.
2065                  */
2066                 if (t10_cksum_type && opc == OST_WRITE &&
2067                     local_nb[i].lnb_len == PAGE_SIZE) {
2068                         local_nb[i].lnb_guard_rpc = 1;
2069                         memcpy(local_nb[i].lnb_guards,
2070                                guard_start + used_number,
2071                                used * sizeof(*local_nb[i].lnb_guards));
2072                 }
2073
2074                 used_number += used;
2075                 if (used_number == guard_number) {
2076                         cfs_crypto_hash_update_page(req, __page, 0,
2077                                 used_number * sizeof(*guard_start));
2078                         used_number = 0;
2079                 }
2080
2081                  /* corrupt the data after we compute the checksum, to
2082                  * simulate an OST->client data error */
2083                 if (unlikely(i == 0 && opc == OST_READ &&
2084                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2085                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2086                         int len = local_nb[i].lnb_len;
2087                         struct page *np = tgt_page_to_corrupt;
2088
2089                         if (np) {
2090                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2091                                                         KM_USER0);
2092                                 char *ptr2 = page_address(np);
2093
2094                                 memcpy(ptr2 + off, ptr + off, len);
2095                                 memcpy(ptr2 + off, "bad4", min(4, len));
2096                                 ll_kunmap_atomic(ptr, KM_USER0);
2097
2098                                 /* LU-8376 to preserve original index for
2099                                  * display in dump_all_bulk_pages() */
2100                                 np->index = i;
2101
2102                                 cfs_crypto_hash_update_page(req, np, off,
2103                                                             len);
2104                                 continue;
2105                         } else {
2106                                 CERROR("%s: can't alloc page for corruption\n",
2107                                        tgt_name(tgt));
2108                         }
2109                 }
2110         }
2111         kunmap(__page);
2112         if (rc)
2113                 GOTO(out, rc);
2114
2115         if (used_number != 0)
2116                 cfs_crypto_hash_update_page(req, __page, 0,
2117                         used_number * sizeof(*guard_start));
2118
2119         bufsize = sizeof(cksum);
2120         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2121
2122         if (rc == 0)
2123                 *check_sum = cksum;
2124 out:
2125         __free_page(__page);
2126         return rc;
2127 }
2128
2129 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2130                                   enum cksum_types cksum_type,
2131                                   struct niobuf_local *local_nb,
2132                                   int npages, int opc, u32 *check_sum)
2133 {
2134         obd_dif_csum_fn *fn = NULL;
2135         int sector_size = 0;
2136         int rc;
2137
2138         ENTRY;
2139         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2140
2141         if (fn)
2142                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2143                                                opc, fn, sector_size,
2144                                                check_sum);
2145         else
2146                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2147                                          cksum_type, check_sum);
2148         RETURN(rc);
2149 }
2150
2151 int tgt_brw_read(struct tgt_session_info *tsi)
2152 {
2153         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2154         struct ptlrpc_bulk_desc *desc = NULL;
2155         struct obd_export       *exp = tsi->tsi_exp;
2156         struct niobuf_remote    *remote_nb;
2157         struct niobuf_local     *local_nb;
2158         struct obd_ioobj        *ioo;
2159         struct ost_body         *body, *repbody;
2160         struct lustre_handle     lockh = { 0 };
2161         int                      npages, nob = 0, rc, i, no_reply = 0,
2162                                  npages_read;
2163         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2164         const char *obd_name = exp->exp_obd->obd_name;
2165
2166         ENTRY;
2167
2168         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2169             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2170                 CERROR("%s: deny read request from %s to portal %u\n",
2171                        tgt_name(tsi->tsi_tgt),
2172                        obd_export_nid2str(req->rq_export),
2173                        ptlrpc_req2svc(req)->srv_req_portal);
2174                 RETURN(-EPROTO);
2175         }
2176
2177         req->rq_bulk_read = 1;
2178
2179         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2180                 RETURN(-EIO);
2181
2182         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2183                          cfs_fail_val : (obd_timeout + 1) / 4);
2184
2185         /* Check if there is eviction in progress, and if so, wait for it to
2186          * finish */
2187         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2188                 /* We do not care how long it takes */
2189                 wait_event_idle(
2190                         exp->exp_obd->obd_evict_inprogress_waitq,
2191                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2192         }
2193
2194         /* There must be big cache in current thread to process this request
2195          * if it is NULL then something went wrong and it wasn't allocated,
2196          * report -ENOMEM in that case */
2197         if (tbc == NULL)
2198                 RETURN(-ENOMEM);
2199
2200         body = tsi->tsi_ost_body;
2201         LASSERT(body != NULL);
2202
2203         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2204         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2205
2206         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2207         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2208
2209         local_nb = tbc->local;
2210
2211         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2212                           &lockh, LCK_PR);
2213         if (rc != 0)
2214                 RETURN(rc);
2215
2216         /*
2217          * If getting the lock took more time than
2218          * client was willing to wait, drop it. b=11330
2219          */
2220         if (ktime_get_real_seconds() > req->rq_deadline ||
2221             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2222                 no_reply = 1;
2223                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2224                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2225                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2226                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2227                 GOTO(out_lock, rc = -ETIMEDOUT);
2228         }
2229
2230         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2231         repbody->oa = body->oa;
2232
2233         npages = PTLRPC_MAX_BRW_PAGES;
2234         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2235                         ioo, remote_nb, &npages, local_nb);
2236         if (rc != 0)
2237                 GOTO(out_lock, rc);
2238
2239         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2240             body->oa.o_flags & OBD_FL_SHORT_IO) {
2241                 desc = NULL;
2242         } else {
2243                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2244                                             PTLRPC_BULK_PUT_SOURCE |
2245                                                 PTLRPC_BULK_BUF_KIOV,
2246                                             OST_BULK_PORTAL,
2247                                             &ptlrpc_bulk_kiov_nopin_ops);
2248                 if (desc == NULL)
2249                         GOTO(out_commitrw, rc = -ENOMEM);
2250         }
2251
2252         nob = 0;
2253         npages_read = npages;
2254         for (i = 0; i < npages; i++) {
2255                 int page_rc = local_nb[i].lnb_rc;
2256
2257                 if (page_rc < 0) {
2258                         rc = page_rc;
2259                         npages_read = i;
2260                         break;
2261                 }
2262
2263                 nob += page_rc;
2264                 if (page_rc != 0 && desc != NULL) { /* some data! */
2265                         LASSERT(local_nb[i].lnb_page != NULL);
2266                         desc->bd_frag_ops->add_kiov_frag
2267                           (desc, local_nb[i].lnb_page,
2268                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2269                            page_rc);
2270                 }
2271
2272                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2273                         local_nb[i].lnb_len = page_rc;
2274                         npages_read = i + (page_rc != 0 ? 1 : 0);
2275                         /* All subsequent pages should be 0 */
2276                         while (++i < npages)
2277                                 LASSERT(local_nb[i].lnb_rc == 0);
2278                         break;
2279                 }
2280         }
2281
2282         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2283                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2284                            body->oa.o_flags : 0;
2285                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2286
2287                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2288                                                           cksum_type);
2289                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2290
2291                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2292                                             local_nb, npages_read, OST_READ,
2293                                             &repbody->oa.o_cksum);
2294                 if (rc < 0)
2295                         GOTO(out_commitrw, rc);
2296                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2297                        repbody->oa.o_cksum);
2298
2299                 /* if a resend it could be for a cksum error, so check Server
2300                  * cksum with returned Client cksum (this should even cover
2301                  * zero-cksum case) */
2302                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2303                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2304                         check_read_checksum(local_nb, npages_read, exp,
2305                                             &body->oa, &req->rq_peer,
2306                                             body->oa.o_cksum,
2307                                             repbody->oa.o_cksum, cksum_type);
2308         } else {
2309                 repbody->oa.o_valid = 0;
2310         }
2311         /* We're finishing using body->oa as an input variable */
2312
2313         /* Check if client was evicted while we were doing i/o before touching
2314          * network */
2315         if (rc == 0) {
2316                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2317                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2318                         unsigned char *short_io_buf;
2319                         int short_io_size;
2320
2321                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2322                                                               &RMF_SHORT_IO);
2323                         short_io_size = req_capsule_get_size(&req->rq_pill,
2324                                                              &RMF_SHORT_IO,
2325                                                              RCL_SERVER);
2326                         rc = tgt_pages2shortio(local_nb, npages_read,
2327                                                short_io_buf, short_io_size);
2328                         if (rc >= 0)
2329                                 req_capsule_shrink(&req->rq_pill,
2330                                                    &RMF_SHORT_IO, rc,
2331                                                    RCL_SERVER);
2332                         rc = rc > 0 ? 0 : rc;
2333                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2334                         rc = target_bulk_io(exp, desc);
2335                 }
2336                 no_reply = rc != 0;
2337         } else {
2338                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2339                     body->oa.o_flags & OBD_FL_SHORT_IO)
2340                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2341                                            RCL_SERVER);
2342         }
2343
2344 out_commitrw:
2345         /* Must commit after prep above in all cases */
2346         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2347                           remote_nb, npages, local_nb, rc);
2348 out_lock:
2349         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2350
2351         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2352                 ptlrpc_free_bulk(desc);
2353
2354         LASSERT(rc <= 0);
2355         if (rc == 0) {
2356                 rc = nob;
2357                 ptlrpc_lprocfs_brw(req, nob);
2358         } else if (no_reply) {
2359                 req->rq_no_reply = 1;
2360                 /* reply out callback would free */
2361                 ptlrpc_req_drop_rs(req);
2362                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2363                               "client will retry: rc %d\n",
2364                               obd_name,
2365                               obd_uuid2str(&exp->exp_client_uuid),
2366                               obd_export_nid2str(exp), rc);
2367         }
2368         /* send a bulk after reply to simulate a network delay or reordering
2369          * by a router - Note that !desc implies short io, so there is no bulk
2370          * to reorder. */
2371         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2372             desc) {
2373                 CDEBUG(D_INFO, "reorder BULK\n");
2374
2375                 ssleep(3);
2376                 target_bulk_io(exp, desc);
2377                 ptlrpc_free_bulk(desc);
2378         }
2379
2380         RETURN(rc);
2381 }
2382 EXPORT_SYMBOL(tgt_brw_read);
2383
2384 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2385                              unsigned char *buf, unsigned int size)
2386 {
2387         int     i, off, len;
2388         char    *ptr;
2389
2390         for (i = 0; i < npages; i++) {
2391                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2392                 len = local[i].lnb_len;
2393
2394                 if (len == 0)
2395                         continue;
2396
2397                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2398                        i, off, len, size);
2399                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2400                 if (ptr == NULL)
2401                         return -EINVAL;
2402                 memcpy(ptr + off, buf, len < size ? len : size);
2403                 ll_kunmap_atomic(ptr, KM_USER0);
2404                 buf += len;
2405                 size -= len;
2406         }
2407         return 0;
2408 }
2409
2410 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2411                               struct ptlrpc_bulk_desc *desc,
2412                               struct niobuf_local *local_nb, int npages,
2413                               u32 client_cksum, u32 server_cksum,
2414                               bool mmap)
2415 {
2416         struct obd_export *exp = req->rq_export;
2417         struct ost_body *body;
2418         char *router = "";
2419         char *via = "";
2420
2421         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2422         LASSERT(body != NULL);
2423
2424         if (desc && req->rq_peer.nid != desc->bd_sender) {
2425                 via = " via ";
2426                 router = libcfs_nid2str(desc->bd_sender);
2427         }
2428
2429         if (exp->exp_obd->obd_checksum_dump)
2430                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2431                                     client_cksum);
2432
2433         if (mmap) {
2434                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2435                              client_cksum, server_cksum);
2436                 return;
2437         }
2438
2439         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2440                            DFID" object "DOSTID" extent [%llu-%llu"
2441                            "]: client csum %x, server csum %x\n",
2442                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2443                            via, router,
2444                            body->oa.o_valid & OBD_MD_FLFID ?
2445                            body->oa.o_parent_seq : (__u64)0,
2446                            body->oa.o_valid & OBD_MD_FLFID ?
2447                            body->oa.o_parent_oid : 0,
2448                            body->oa.o_valid & OBD_MD_FLFID ?
2449                            body->oa.o_parent_ver : 0,
2450                            POSTID(&body->oa.o_oi),
2451                            local_nb[0].lnb_file_offset,
2452                            local_nb[npages-1].lnb_file_offset +
2453                            local_nb[npages - 1].lnb_len - 1,
2454                            client_cksum, server_cksum);
2455 }
2456
2457 int tgt_brw_write(struct tgt_session_info *tsi)
2458 {
2459         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2460         struct ptlrpc_bulk_desc *desc = NULL;
2461         struct obd_export       *exp = req->rq_export;
2462         struct niobuf_remote    *remote_nb;
2463         struct niobuf_local     *local_nb;
2464         struct obd_ioobj        *ioo;
2465         struct ost_body         *body, *repbody;
2466         struct lustre_handle     lockh = {0};
2467         __u32                   *rcs;
2468         int                      objcount, niocount, npages;
2469         int                      rc, i, j;
2470         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2471         bool                     no_reply = false, mmap;
2472         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2473         bool wait_sync = false;
2474         const char *obd_name = exp->exp_obd->obd_name;
2475
2476         ENTRY;
2477
2478         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2479             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2480                 CERROR("%s: deny write request from %s to portal %u\n",
2481                        tgt_name(tsi->tsi_tgt),
2482                        obd_export_nid2str(req->rq_export),
2483                        ptlrpc_req2svc(req)->srv_req_portal);
2484                 RETURN(err_serious(-EPROTO));
2485         }
2486
2487         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2488                 RETURN(err_serious(-ENOSPC));
2489         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2490                 RETURN(err_serious(-EROFS));
2491
2492         req->rq_bulk_write = 1;
2493
2494         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2495                 RETURN(err_serious(-EIO));
2496         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2497                 RETURN(err_serious(-EFAULT));
2498
2499         /* pause before transaction has been started */
2500         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2501                          cfs_fail_val : (obd_timeout + 1) / 4);
2502
2503         /* Delay write commit to show stale size information */
2504         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2505
2506         /* There must be big cache in current thread to process this request
2507          * if it is NULL then something went wrong and it wasn't allocated,
2508          * report -ENOMEM in that case */
2509         if (tbc == NULL)
2510                 RETURN(-ENOMEM);
2511
2512         body = tsi->tsi_ost_body;
2513         LASSERT(body != NULL);
2514
2515         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2516         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2517
2518         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2519                                         RCL_CLIENT) / sizeof(*ioo);
2520
2521         for (niocount = i = 0; i < objcount; i++)
2522                 niocount += ioo[i].ioo_bufcnt;
2523
2524         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2525         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2526         if (niocount != req_capsule_get_size(&req->rq_pill,
2527                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2528                         sizeof(*remote_nb))
2529                 RETURN(err_serious(-EPROTO));
2530
2531         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2532             ptlrpc_connection_is_local(exp->exp_connection))
2533                 memory_pressure_set();
2534
2535         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2536                              niocount * sizeof(*rcs));
2537         rc = req_capsule_server_pack(&req->rq_pill);
2538         if (rc != 0)
2539                 GOTO(out, rc = err_serious(rc));
2540
2541         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2542         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2543
2544         local_nb = tbc->local;
2545
2546         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2547                           &lockh, LCK_PW);
2548         if (rc != 0)
2549                 GOTO(out, rc);
2550
2551         /*
2552          * If getting the lock took more time than
2553          * client was willing to wait, drop it. b=11330
2554          */
2555         if (ktime_get_real_seconds() > req->rq_deadline ||
2556             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2557                 no_reply = true;
2558                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2559                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2560                        POSTID(&ioo->ioo_oid),
2561                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2562                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2563                 GOTO(out_lock, rc = -ETIMEDOUT);
2564         }
2565
2566         /* Because we already sync grant info with client when reconnect,
2567          * grant info will be cleared for resent req, then fed_grant and
2568          * total_grant will not be modified in following preprw_write */
2569         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2570                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2571                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2572         }
2573
2574         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2575         if (repbody == NULL)
2576                 GOTO(out_lock, rc = -ENOMEM);
2577         repbody->oa = body->oa;
2578
2579         npages = PTLRPC_MAX_BRW_PAGES;
2580         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2581                         objcount, ioo, remote_nb, &npages, local_nb);
2582         if (rc < 0)
2583                 GOTO(out_lock, rc);
2584         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2585             body->oa.o_flags & OBD_FL_SHORT_IO) {
2586                 unsigned int short_io_size;
2587                 unsigned char *short_io_buf;
2588
2589                 short_io_size = req_capsule_get_size(&req->rq_pill,
2590                                                      &RMF_SHORT_IO,
2591                                                      RCL_CLIENT);
2592                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2593                                                       &RMF_SHORT_IO);
2594                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2595                                " size = %d\n", short_io_size);
2596
2597                 /* Copy short io buf to pages */
2598                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2599                                        short_io_size);
2600                 desc = NULL;
2601         } else {
2602                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2603                                             PTLRPC_BULK_GET_SINK |
2604                                             PTLRPC_BULK_BUF_KIOV,
2605                                             OST_BULK_PORTAL,
2606                                             &ptlrpc_bulk_kiov_nopin_ops);
2607                 if (desc == NULL)
2608                         GOTO(skip_transfer, rc = -ENOMEM);
2609
2610                 /* NB Having prepped, we must commit... */
2611                 for (i = 0; i < npages; i++)
2612                         desc->bd_frag_ops->add_kiov_frag(desc,
2613                                         local_nb[i].lnb_page,
2614                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2615                                         local_nb[i].lnb_len);
2616
2617                 rc = sptlrpc_svc_prep_bulk(req, desc);
2618                 if (rc != 0)
2619                         GOTO(skip_transfer, rc);
2620
2621                 rc = target_bulk_io(exp, desc);
2622         }
2623
2624         no_reply = rc != 0;
2625
2626 skip_transfer:
2627         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2628                 static int cksum_counter;
2629
2630                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2631                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2632
2633                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2634                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2635                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2636                                                            cksum_type);
2637
2638                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2639                                             local_nb, npages, OST_WRITE,
2640                                             &repbody->oa.o_cksum);
2641                 if (rc < 0)
2642                         GOTO(out_commitrw, rc);
2643
2644                 cksum_counter++;
2645
2646                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2647                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2648                                 body->oa.o_flags & OBD_FL_MMAP);
2649
2650                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2651                                           body->oa.o_cksum,
2652                                           repbody->oa.o_cksum, mmap);
2653                         cksum_counter = 0;
2654                 } else if ((cksum_counter & (-cksum_counter)) ==
2655                            cksum_counter) {
2656                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2657                                cksum_counter, libcfs_id2str(req->rq_peer),
2658                                repbody->oa.o_cksum);
2659                 }
2660         }
2661
2662 out_commitrw:
2663         /* Must commit after prep above in all cases */
2664         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2665                           objcount, ioo, remote_nb, npages, local_nb, rc);
2666         if (rc == -ENOTCONN)
2667                 /* quota acquire process has been given up because
2668                  * either the client has been evicted or the client
2669                  * has timed out the request already */
2670                 no_reply = true;
2671
2672         for (i = 0; i < niocount; i++) {
2673                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2674                         wait_sync = true;
2675                         break;
2676                 }
2677         }
2678         /*
2679          * Disable sending mtime back to the client. If the client locked the
2680          * whole object, then it has already updated the mtime on its side,
2681          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2682          */
2683         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2684
2685         if (rc == 0) {
2686                 int nob = 0;
2687
2688                 /* set per-requested niobuf return codes */
2689                 for (i = j = 0; i < niocount; i++) {
2690                         int len = remote_nb[i].rnb_len;
2691
2692                         nob += len;
2693                         rcs[i] = 0;
2694                         do {
2695                                 LASSERT(j < npages);
2696                                 if (local_nb[j].lnb_rc < 0)
2697                                         rcs[i] = local_nb[j].lnb_rc;
2698                                 len -= local_nb[j].lnb_len;
2699                                 j++;
2700                         } while (len > 0);
2701                         LASSERT(len == 0);
2702                 }
2703                 LASSERT(j == npages);
2704                 ptlrpc_lprocfs_brw(req, nob);
2705         }
2706 out_lock:
2707         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2708         if (desc)
2709                 ptlrpc_free_bulk(desc);
2710 out:
2711         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2712                 req->rq_no_reply = 1;
2713                 /* reply out callback would free */
2714                 ptlrpc_req_drop_rs(req);
2715                 if (!exp->exp_obd->obd_no_transno)
2716                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2717                                       " client will retry: rc = %d\n",
2718                                       obd_name,
2719                                       obd_uuid2str(&exp->exp_client_uuid),
2720                                       obd_export_nid2str(exp), rc);
2721         }
2722         memory_pressure_clr();
2723         RETURN(rc);
2724 }
2725 EXPORT_SYMBOL(tgt_brw_write);
2726
2727 /* Check if request can be reconstructed from saved reply data
2728  * A copy of the reply data is returned in @trd if the pointer is not NULL
2729  */
2730 int req_can_reconstruct(struct ptlrpc_request *req,
2731                          struct tg_reply_data *trd)
2732 {
2733         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2734         struct lsd_client_data *lcd = ted->ted_lcd;
2735         int found;
2736
2737         if (tgt_is_multimodrpcs_client(req->rq_export))
2738                 return tgt_lookup_reply(req, trd);
2739
2740         mutex_lock(&ted->ted_lcd_lock);
2741         found = req->rq_xid == lcd->lcd_last_xid ||
2742                 req->rq_xid == lcd->lcd_last_close_xid;
2743
2744         if (found && trd != NULL) {
2745                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2746                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2747                         trd->trd_reply.lrd_transno =
2748                                                 lcd->lcd_last_close_transno;
2749                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2750                 } else {
2751                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2752                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2753                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2754                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2755                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2756                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2757                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2758                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2759                 }
2760         }
2761         mutex_unlock(&ted->ted_lcd_lock);
2762
2763         return found;
2764 }
2765 EXPORT_SYMBOL(req_can_reconstruct);
2766