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 #ifdef HAVE_UIDGID_HEADER
  39 # include <linux/uidgid.h>
  40 #endif
  41
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47 #include <lustre_acl.h>
  48
  49 #include "tgt_internal.h"
  50
  51 char *tgt_name(struct lu_target *tgt)
  52 {
  53         LASSERT(tgt->lut_obd != NULL);
  54         return tgt->lut_obd->obd_name;
  55 }
  56 EXPORT_SYMBOL(tgt_name);
  57
  58 /*
  59  * Generic code handling requests that have struct mdt_body passed in:
  60  *
  61  *  - extract mdt_body from request and save it in @tsi, if present;
  62  *
  63  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  64  *  @tsi;
  65  *
  66  *  - if HAS_BODY flag is set for this request type check whether object
  67  *  actually exists on storage (lu_object_exists()).
  68  *
  69  */
  70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  71 {
  72         const struct mdt_body   *body;
  73         struct lu_object        *obj;
  74         struct req_capsule      *pill = tsi->tsi_pill;
  75         int                      rc;
  76
  77         ENTRY;
  78
  79         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EFAULT);
  82
  83         tsi->tsi_mdt_body = body;
  84
  85         if (!(body->mbo_valid & OBD_MD_FLID))
  86                 RETURN(0);
  87
  88         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  89          * in any case in pre-2.5 clients. Fix that here if needed */
  90         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  91                 RETURN(0);
  92
  93         if (!fid_is_sane(&body->mbo_fid1)) {
  94                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  95                        PFID(&body->mbo_fid1));
  96                 RETURN(-EINVAL);
  97         }
  98
  99         obj = lu_object_find(tsi->tsi_env,
 100                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 101                              &body->mbo_fid1, NULL);
 102         if (!IS_ERR(obj)) {
 103                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 104                         lu_object_put(tsi->tsi_env, obj);
 105                         rc = -ENOENT;
 106                 } else {
 107                         tsi->tsi_corpus = obj;
 108                         rc = 0;
 109                 }
 110         } else {
 111                 rc = PTR_ERR(obj);
 112         }
 113
 114         tsi->tsi_fid = body->mbo_fid1;
 115
 116         RETURN(rc);
 117 }
 118
 119 /**
 120  * Validate oa from client.
 121  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 122  * req are valid.
 123  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 124  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 125  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 126  *       whether this is oi_fid or real ostid. So it will check
 127  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 128  *    c. Old FID-disable osc will send IDIF.
 129  *    d. new FID-enable osc/osp will send normal FID.
 130  *
 131  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 132  * be used for LAST_ID file, and only being accessed inside OST now.
 133  */
 134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 135 {
 136         struct ost_id   *oi     = &oa->o_oi;
 137         u64              seq    = ostid_seq(oi);
 138         u64              id     = ostid_id(oi);
 139         int              rc;
 140         ENTRY;
 141
 142         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 143                      fid_seq_is_echo(seq))) {
 144                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 145                  * during create, and we will reset this to 1, since this
 146                  * oi_id is basically useless in the following create process,
 147                  * but oi_id == 0 will make it difficult to tell whether it is
 148                  * real FID or ost_id. */
 149                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 150                 oi->oi_fid.f_oid = id ?: 1;
 151                 oi->oi_fid.f_ver = 0;
 152         } else {
 153                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 154
 155                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 156                         GOTO(out, rc = -EPROTO);
 157
 158                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 159                  * all of the requests are required to setup FLGROUP */
 160                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 161                         ostid_set_seq_mdt0(oi);
 162                         oa->o_valid |= OBD_MD_FLGROUP;
 163                         seq = ostid_seq(oi);
 164                 }
 165
 166                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 167                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 168                         GOTO(out, rc = -EPROTO);
 169
 170                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 171                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 172                 if (unlikely(rc != 0))
 173                         GOTO(out, rc);
 174
 175                 oi->oi_fid = tti->tti_fid1;
 176         }
 177
 178         RETURN(0);
 179
 180 out:
 181         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 182                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 183                seq, id, rc);
 184         return rc;
 185 }
 186 EXPORT_SYMBOL(tgt_validate_obdo);
 187
 188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 189 {
 190         unsigned                 max_brw;
 191         struct niobuf_remote    *rnb;
 192         struct obd_ioobj        *ioo;
 193         int                      obj_count;
 194
 195         ENTRY;
 196
 197         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 198         if (ioo == NULL)
 199                 RETURN(-EPROTO);
 200
 201         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 202         if (rnb == NULL)
 203                 RETURN(-EPROTO);
 204
 205         max_brw = ioobj_max_brw_get(ioo);
 206         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 207                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 208                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 209                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 210                        POSTID(oi), -EPROTO);
 211                 RETURN(-EPROTO);
 212         }
 213         ioo->ioo_oid = *oi;
 214
 215         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 216                                         RCL_CLIENT) / sizeof(*ioo);
 217         if (obj_count == 0) {
 218                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 219                 RETURN(-EPROTO);
 220         } else if (obj_count > 1) {
 221                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 222                        obj_count);
 223                 RETURN(-EPROTO);
 224         }
 225
 226         if (ioo->ioo_bufcnt == 0) {
 227                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 228                 RETURN(-EPROTO);
 229         }
 230
 231         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 232                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 233                           "bulk has too many pages (%d)",
 234                           ioo->ioo_bufcnt);
 235                 RETURN(-EPROTO);
 236         }
 237
 238         RETURN(0);
 239 }
 240
 241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 242 {
 243         struct ost_body         *body;
 244         struct req_capsule      *pill = tsi->tsi_pill;
 245         struct lu_nodemap       *nodemap;
 246         int                      rc;
 247
 248         ENTRY;
 249
 250         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 251         if (body == NULL)
 252                 RETURN(-EFAULT);
 253
 254         rc = tgt_validate_obdo(tsi, &body->oa);
 255         if (rc)
 256                 RETURN(rc);
 257
 258         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 259         if (IS_ERR(nodemap))
 260                 RETURN(PTR_ERR(nodemap));
 261
 262         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_uid);
 265         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 266                                         NODEMAP_CLIENT_TO_FS,
 267                                         body->oa.o_gid);
 268         nodemap_putref(nodemap);
 269
 270         tsi->tsi_ost_body = body;
 271         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 272
 273         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 274                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 275                 if (rc < 0)
 276                         RETURN(rc);
 277         }
 278
 279         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 280                 if (flags & HAS_BODY) {
 281                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 282                                tgt_name(tsi->tsi_tgt));
 283                         RETURN(-EPROTO);
 284                 } else {
 285                         RETURN(0);
 286                 }
 287         }
 288
 289         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 290
 291         /*
 292          * OST doesn't get object in advance for further use to prevent
 293          * situations with nested object_find which is potential deadlock.
 294          */
 295         tsi->tsi_corpus = NULL;
 296         RETURN(rc);
 297 }
 298
 299 /*
 300  * Do necessary preprocessing according to handler ->th_flags.
 301  */
 302 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 303                                   struct tgt_handler *h,
 304                                   struct ptlrpc_request *req)
 305 {
 306         struct req_capsule      *pill = tsi->tsi_pill;
 307         __u32                    flags = h->th_flags;
 308         int                      rc = 0;
 309
 310         ENTRY;
 311
 312         if (tsi->tsi_preprocessed)
 313                 RETURN(0);
 314
 315         LASSERT(h->th_act != NULL);
 316         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 317         LASSERT(current->journal_info == NULL);
 318
 319         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 320                      h->th_fmt != NULL));
 321         if (h->th_fmt != NULL) {
 322                 req_capsule_set(pill, h->th_fmt);
 323                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
 324                         rc = tgt_mdt_body_unpack(tsi, flags);
 325                         if (rc < 0)
 326                                 RETURN(rc);
 327                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 328                                                  RCL_CLIENT)) {
 329                         rc = tgt_ost_body_unpack(tsi, flags);
 330                         if (rc < 0)
 331                                 RETURN(rc);
 332                 }
 333         }
 334
 335         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 336                 RETURN(-EROFS);
 337
 338         if (flags & HAS_KEY) {
 339                 struct ldlm_request *dlm_req;
 340
 341                 LASSERT(h->th_fmt != NULL);
 342
 343                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 344                 if (dlm_req != NULL) {
 345                         union ldlm_wire_policy_data *policy =
 346                                         &dlm_req->lock_desc.l_policy_data;
 347
 348                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 349                                      LDLM_IBITS &&
 350                                      (policy->l_inodebits.bits |
 351                                       policy->l_inodebits.try_bits) == 0)) {
 352                                 /*
 353                                  * Lock without inodebits makes no sense and
 354                                  * will oops later in ldlm. If client miss to
 355                                  * set such bits, do not trigger ASSERTION.
 356                                  *
 357                                  * For liblustre flock case, it maybe zero.
 358                                  */
 359                                 rc = -EPROTO;
 360                         } else {
 361                                 tsi->tsi_dlm_req = dlm_req;
 362                         }
 363                 } else {
 364                         rc = -EFAULT;
 365                 }
 366         }
 367         tsi->tsi_preprocessed = 1;
 368         RETURN(rc);
 369 }
 370
 371 /*
 372  * Invoke handler for this request opc. Also do necessary preprocessing
 373  * (according to handler ->th_flags), and post-processing (setting of
 374  * ->last_{xid,committed}).
 375  */
 376 static int tgt_handle_request0(struct tgt_session_info *tsi,
 377                                struct tgt_handler *h,
 378                                struct ptlrpc_request *req)
 379 {
 380         int      serious = 0;
 381         int      rc;
 382         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 383
 384         ENTRY;
 385
 386
 387         /* When dealing with sec context requests, no export is associated yet,
 388          * because these requests are sent before *_CONNECT requests.
 389          * A NULL req->rq_export means the normal *_common_slice handlers will
 390          * not be called, because there is no reference to the target.
 391          * So deal with them by hand and jump directly to target_send_reply().
 392          */
 393         switch (opc) {
 394         case SEC_CTX_INIT:
 395         case SEC_CTX_INIT_CONT:
 396         case SEC_CTX_FINI:
 397                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 398                 GOTO(out, rc = 0);
 399         }
 400
 401         /*
 402          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 403          * to put same checks into handlers like mdt_close(), mdt_reint(),
 404          * etc., without talking to mdt authors first. Checking same thing
 405          * there again is useless and returning 0 error without packing reply
 406          * is buggy! Handlers either pack reply or return error.
 407          *
 408          * We return 0 here and do not send any reply in order to emulate
 409          * network failure. Do not send any reply in case any of NET related
 410          * fail_id has occured.
 411          */
 412         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 413                 RETURN(0);
 414         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 415                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 416                 RETURN(0);
 417
 418         rc = tgt_request_preprocess(tsi, h, req);
 419         /* pack reply if reply format is fixed */
 420         if (rc == 0 && h->th_flags & HAS_REPLY) {
 421                 /* Pack reply */
 422                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 423                                           RCL_SERVER))
 424                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 425                                              RCL_SERVER,
 426                                              tsi->tsi_mdt_body->mbo_eadatasize);
 427                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 428                                           RCL_SERVER))
 429                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 430                                              RCL_SERVER, 0);
 431                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 432                         req_capsule_set_size(tsi->tsi_pill,
 433                                              &RMF_ACL, RCL_SERVER,
 434                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 435
 436                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 437                                           RCL_SERVER)) {
 438                         struct niobuf_remote *remote_nb =
 439                                 req_capsule_client_get(tsi->tsi_pill,
 440                                                        &RMF_NIOBUF_REMOTE);
 441                         struct ost_body *body = tsi->tsi_ost_body;
 442
 443                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 444                                          RCL_SERVER,
 445                                          (body->oa.o_flags & OBD_FL_SHORT_IO) ?
 446                                          remote_nb[0].rnb_len : 0);
 447                 }
 448
 449                 rc = req_capsule_server_pack(tsi->tsi_pill);
 450         }
 451
 452         if (likely(rc == 0)) {
 453                 /*
 454                  * Process request, there can be two types of rc:
 455                  * 1) errors with msg unpack/pack, other failures outside the
 456                  * operation itself. This is counted as serious errors;
 457                  * 2) errors during fs operation, should be placed in rq_status
 458                  * only
 459                  */
 460                 rc = h->th_act(tsi);
 461                 if (!is_serious(rc) &&
 462                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 463                         DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
 464                                   "pack reply and returned 0 error\n",
 465                                   tgt_name(tsi->tsi_tgt), h->th_name);
 466                         LBUG();
 467                 }
 468                 serious = is_serious(rc);
 469                 rc = clear_serious(rc);
 470         } else {
 471                 serious = 1;
 472         }
 473
 474         req->rq_status = rc;
 475
 476         /*
 477          * ELDLM_* codes which > 0 should be in rq_status only as well as
 478          * all non-serious errors.
 479          */
 480         if (rc > 0 || !serious)
 481                 rc = 0;
 482
 483         LASSERT(current->journal_info == NULL);
 484
 485         if (likely(rc == 0 && req->rq_export))
 486                 target_committed_to_req(req);
 487
 488 out:
 489         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 490         RETURN(0);
 491 }
 492
 493 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 494                                        struct obd_device *obd, int *process)
 495 {
 496         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 497         case MDS_DISCONNECT:
 498         case OST_DISCONNECT:
 499         case OBD_IDX_READ:
 500                 *process = 1;
 501                 RETURN(0);
 502         case MDS_CLOSE:
 503         case MDS_SYNC: /* used in unmounting */
 504         case OBD_PING:
 505         case MDS_REINT:
 506         case OUT_UPDATE:
 507         case SEQ_QUERY:
 508         case FLD_QUERY:
 509         case FLD_READ:
 510         case LDLM_ENQUEUE:
 511         case OST_CREATE:
 512         case OST_DESTROY:
 513         case OST_PUNCH:
 514         case OST_SETATTR:
 515         case OST_SYNC:
 516         case OST_WRITE:
 517         case MDS_HSM_PROGRESS:
 518         case MDS_HSM_STATE_SET:
 519         case MDS_HSM_REQUEST:
 520                 *process = target_queue_recovery_request(req, obd);
 521                 RETURN(0);
 522
 523         default:
 524                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 525                 *process = -EAGAIN;
 526                 RETURN(0);
 527         }
 528 }
 529
 530 /*
 531  * Handle recovery. Return:
 532  *        +1: continue request processing;
 533  *       -ve: abort immediately with the given error code;
 534  *         0: send reply with error code in req->rq_status;
 535  */
 536 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 537 {
 538         ENTRY;
 539
 540         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 541         case MDS_CONNECT:
 542         case OST_CONNECT:
 543         case MGS_CONNECT:
 544         case SEC_CTX_INIT:
 545         case SEC_CTX_INIT_CONT:
 546         case SEC_CTX_FINI:
 547                 RETURN(+1);
 548         }
 549
 550         if (!req->rq_export->exp_obd->obd_replayable)
 551                 RETURN(+1);
 552
 553         /* sanity check: if the xid matches, the request must be marked as a
 554          * resent or replayed */
 555         if (req_can_reconstruct(req, NULL)) {
 556                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 557                       (MSG_RESENT | MSG_REPLAY))) {
 558                         DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches "
 559                                   "saved xid, expected REPLAY or RESENT flag "
 560                                   "(%x)", req->rq_xid,
 561                                   lustre_msg_get_flags(req->rq_reqmsg));
 562                         req->rq_status = -ENOTCONN;
 563                         RETURN(-ENOTCONN);
 564                 }
 565         }
 566         /* else: note the opposite is not always true; a RESENT req after a
 567          * failover will usually not match the last_xid, since it was likely
 568          * never committed. A REPLAYed request will almost never match the
 569          * last xid, however it could for a committed, but still retained,
 570          * open. */
 571
 572         /* Check for aborted recovery... */
 573         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 574                 int rc;
 575                 int should_process;
 576
 577                 DEBUG_REQ(D_INFO, req, "Got new replay");
 578                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 579                                                  &should_process);
 580                 if (rc != 0 || !should_process)
 581                         RETURN(rc);
 582                 else if (should_process < 0) {
 583                         req->rq_status = should_process;
 584                         rc = ptlrpc_error(req);
 585                         RETURN(rc);
 586                 }
 587         }
 588         RETURN(+1);
 589 }
 590
 591 /* Initial check for request, it is validation mostly */
 592 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 593 {
 594         struct tgt_handler      *h;
 595         struct tgt_opc_slice    *s;
 596         struct lu_target        *tgt;
 597         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 598
 599         ENTRY;
 600
 601         tgt = class_exp2tgt(req->rq_export);
 602         if (unlikely(tgt == NULL)) {
 603                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
 604                           class_exp2obd(req->rq_export)->obd_name);
 605                 RETURN(ERR_PTR(-EINVAL));
 606         }
 607
 608         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 609                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 610                         break;
 611
 612         /* opcode was not found in slice */
 613         if (unlikely(s->tos_hs == NULL)) {
 614                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 615                        opc);
 616                 RETURN(ERR_PTR(-ENOTSUPP));
 617         }
 618
 619         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 620         h = s->tos_hs + (opc - s->tos_opc_start);
 621         if (unlikely(h->th_opc == 0)) {
 622                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 623                 RETURN(ERR_PTR(-ENOTSUPP));
 624         }
 625
 626         RETURN(h);
 627 }
 628
 629 static int process_req_last_xid(struct ptlrpc_request *req)
 630 {
 631         __u64   last_xid;
 632         ENTRY;
 633
 634         /* check request's xid is consistent with export's last_xid */
 635         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 636         if (last_xid > req->rq_export->exp_last_xid)
 637                 req->rq_export->exp_last_xid = last_xid;
 638
 639         if (req->rq_xid == 0 ||
 640             (req->rq_xid <= req->rq_export->exp_last_xid)) {
 641                 DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
 642                           "last_xid %llx\n", req->rq_xid,
 643                           req->rq_export->exp_last_xid);
 644                 /* Some request is allowed to be sent during replay,
 645                  * such as OUT update requests, FLD requests, so it
 646                  * is possible that replay requests has smaller XID
 647                  * than the exp_last_xid.
 648                  *
 649                  * Some non-replay requests may have smaller XID as
 650                  * well:
 651                  *
 652                  * - Client send a no_resend RPC, like statfs;
 653                  * - The RPC timedout (or some other error) on client,
 654                  *   then it's removed from the unreplied list;
 655                  * - Client send some other request to bump the
 656                  *   exp_last_xid on server;
 657                  * - The former RPC got chance to be processed;
 658                  */
 659                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 660                         RETURN(-EPROTO);
 661         }
 662
 663         /* The "last_xid" is the minimum xid among unreplied requests,
 664          * if the request is from the previous connection, its xid can
 665          * still be larger than "exp_last_xid", then the above check of
 666          * xid is not enough to determine whether the request is delayed.
 667          *
 668          * For example, if some replay request was delayed and caused
 669          * timeout at client and the replay is restarted, the delayed
 670          * replay request will have the larger xid than "exp_last_xid"
 671          */
 672         if (req->rq_export->exp_conn_cnt >
 673             lustre_msg_get_conn_cnt(req->rq_reqmsg))
 674                 RETURN(-ESTALE);
 675
 676         /* try to release in-memory reply data */
 677         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 678                 tgt_handle_received_xid(req->rq_export,
 679                                 lustre_msg_get_last_xid(req->rq_reqmsg));
 680                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 681                       (MSG_RESENT | MSG_REPLAY)))
 682                         tgt_handle_tag(req->rq_export,
 683                                        lustre_msg_get_tag(req->rq_reqmsg));
 684         }
 685         RETURN(0);
 686 }
 687
 688 int tgt_request_handle(struct ptlrpc_request *req)
 689 {
 690         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 691
 692         struct lustre_msg       *msg = req->rq_reqmsg;
 693         struct tgt_handler      *h;
 694         struct lu_target        *tgt;
 695         int                      request_fail_id = 0;
 696         __u32                    opc = lustre_msg_get_opc(msg);
 697         struct obd_device       *obd;
 698         int                      rc;
 699         bool                     is_connect = false;
 700         ENTRY;
 701
 702         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 703                 if (cfs_fail_val == 0 &&
 704                     lustre_msg_get_opc(msg) != OBD_PING &&
 705                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 706                         struct l_wait_info lwi =  { 0 };
 707
 708                         cfs_fail_val = 1;
 709                         cfs_race_state = 0;
 710                         l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
 711                                      &lwi);
 712                 }
 713         }
 714
 715         /* Refill the context, to make sure all thread keys are allocated */
 716         lu_env_refill(req->rq_svc_thread->t_env);
 717
 718         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 719         tsi->tsi_pill = &req->rq_pill;
 720         tsi->tsi_env = req->rq_svc_thread->t_env;
 721
 722         /* if request has export then get handlers slice from corresponding
 723          * target, otherwise that should be connect operation */
 724         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 725             opc == MGS_CONNECT) {
 726                 is_connect = true;
 727                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 728                 rc = target_handle_connect(req);
 729                 if (rc != 0) {
 730                         rc = ptlrpc_error(req);
 731                         GOTO(out, rc);
 732                 }
 733                 /* recovery-small test 18c asks to drop connect reply */
 734                 if (unlikely(opc == OST_CONNECT &&
 735                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 736                         GOTO(out, rc = 0);
 737         }
 738
 739         if (unlikely(!class_connected_export(req->rq_export))) {
 740                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 741                     opc == SEC_CTX_FINI) {
 742                         /* sec context initialization has to be handled
 743                          * by hand in tgt_handle_request0() */
 744                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 745                         h = NULL;
 746                         GOTO(handle_recov, rc = 0);
 747                 }
 748                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 749                        opc, libcfs_id2str(req->rq_peer));
 750                 req->rq_status = -ENOTCONN;
 751                 rc = ptlrpc_error(req);
 752                 GOTO(out, rc);
 753         }
 754
 755         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 756         tsi->tsi_exp = req->rq_export;
 757         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 758                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 759         else
 760                 tsi->tsi_jobid = NULL;
 761
 762         if (tgt == NULL) {
 763                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
 764                           class_exp2obd(req->rq_export)->obd_name);
 765                 req->rq_status = -EINVAL;
 766                 rc = ptlrpc_error(req);
 767                 GOTO(out, rc);
 768         }
 769
 770         /* Skip last_xid processing for the recovery thread, otherwise, the
 771          * last_xid on same request could be processed twice: first time when
 772          * processing the incoming request, second time when the request is
 773          * being processed by recovery thread. */
 774         obd = class_exp2obd(req->rq_export);
 775         if (is_connect) {
 776                 /* reset the exp_last_xid on each connection. */
 777                 req->rq_export->exp_last_xid = 0;
 778         } else if (obd->obd_recovery_data.trd_processing_task !=
 779                    current_pid()) {
 780                 rc = process_req_last_xid(req);
 781                 if (rc) {
 782                         req->rq_status = rc;
 783                         rc = ptlrpc_error(req);
 784                         GOTO(out, rc);
 785                 }
 786         }
 787
 788         request_fail_id = tgt->lut_request_fail_id;
 789         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 790
 791         h = tgt_handler_find_check(req);
 792         if (IS_ERR(h)) {
 793                 req->rq_status = PTR_ERR(h);
 794                 rc = ptlrpc_error(req);
 795                 GOTO(out, rc);
 796         }
 797
 798         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 799                  h->th_opc, opc);
 800
 801         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 802              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 803                 GOTO(out, rc = 0);
 804
 805         rc = lustre_msg_check_version(msg, h->th_version);
 806         if (unlikely(rc)) {
 807                 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
 808                           " %08x, expecting %08x\n", tgt_name(tgt),
 809                           lustre_msg_get_version(msg), h->th_version);
 810                 req->rq_status = -EINVAL;
 811                 rc = ptlrpc_error(req);
 812                 GOTO(out, rc);
 813         }
 814
 815 handle_recov:
 816         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 817         if (likely(rc == 1)) {
 818                 rc = tgt_handle_request0(tsi, h, req);
 819                 if (rc)
 820                         GOTO(out, rc);
 821         }
 822         EXIT;
 823 out:
 824         req_capsule_fini(tsi->tsi_pill);
 825         if (tsi->tsi_corpus != NULL) {
 826                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 827                 tsi->tsi_corpus = NULL;
 828         }
 829         return rc;
 830 }
 831 EXPORT_SYMBOL(tgt_request_handle);
 832
 833 /** Assign high priority operations to the request if needed. */
 834 int tgt_hpreq_handler(struct ptlrpc_request *req)
 835 {
 836         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 837         struct tgt_handler      *h;
 838         int                      rc;
 839
 840         ENTRY;
 841
 842         if (req->rq_export == NULL)
 843                 RETURN(0);
 844
 845         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 846         tsi->tsi_pill = &req->rq_pill;
 847         tsi->tsi_env = req->rq_svc_thread->t_env;
 848         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 849         tsi->tsi_exp = req->rq_export;
 850
 851         h = tgt_handler_find_check(req);
 852         if (IS_ERR(h)) {
 853                 rc = PTR_ERR(h);
 854                 RETURN(rc);
 855         }
 856
 857         rc = tgt_request_preprocess(tsi, h, req);
 858         if (unlikely(rc != 0))
 859                 RETURN(rc);
 860
 861         if (h->th_hp != NULL)
 862                 h->th_hp(tsi);
 863         RETURN(0);
 864 }
 865 EXPORT_SYMBOL(tgt_hpreq_handler);
 866
 867 void tgt_counter_incr(struct obd_export *exp, int opcode)
 868 {
 869         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 870         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 871                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 872 }
 873 EXPORT_SYMBOL(tgt_counter_incr);
 874
 875 /*
 876  * Unified target generic handlers.
 877  */
 878
 879 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 880 {
 881         struct lu_target        *tgt = class_exp2tgt(exp);
 882         struct sptlrpc_flavor    flvr;
 883         int                      rc = 0;
 884
 885         LASSERT(tgt);
 886         LASSERT(tgt->lut_obd);
 887         LASSERT(tgt->lut_slice);
 888
 889         /* always allow ECHO client */
 890         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 891                             LUSTRE_ECHO_NAME) == 0)) {
 892                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 893                 return 0;
 894         }
 895
 896         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 897                 read_lock(&tgt->lut_sptlrpc_lock);
 898                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 899                                              req->rq_sp_from,
 900                                              req->rq_peer.nid,
 901                                              &flvr);
 902                 read_unlock(&tgt->lut_sptlrpc_lock);
 903
 904                 spin_lock(&exp->exp_lock);
 905                 exp->exp_sp_peer = req->rq_sp_from;
 906                 exp->exp_flvr = flvr;
 907
 908                 /* when on mgs, if no restriction is set, or if client
 909                  * is loopback, allow any flavor */
 910                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 911                            LUSTRE_MGS_NAME) == 0) &&
 912                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 913                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 914                       == LOLND))
 915                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 916
 917                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 918                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 919                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 920                                "expect %x\n", tgt_name(tgt),
 921                                req->rq_flvr.sf_rpc,
 922                                libcfs_nid2str(req->rq_peer.nid),
 923                                exp->exp_flvr.sf_rpc);
 924                         rc = -EACCES;
 925                 }
 926                 spin_unlock(&exp->exp_lock);
 927         } else {
 928                 if (exp->exp_sp_peer != req->rq_sp_from) {
 929                         CERROR("%s: RPC source %s doesn't match %s\n",
 930                                tgt_name(tgt),
 931                                sptlrpc_part2name(req->rq_sp_from),
 932                                sptlrpc_part2name(exp->exp_sp_peer));
 933                         rc = -EACCES;
 934                 } else {
 935                         rc = sptlrpc_target_export_check(exp, req);
 936                 }
 937         }
 938
 939         return rc;
 940 }
 941
 942 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 943 {
 944         struct sptlrpc_rule_set  tmp_rset;
 945         int                      rc;
 946
 947         if (unlikely(tgt == NULL)) {
 948                 CERROR("No target passed");
 949                 return -EINVAL;
 950         }
 951
 952         sptlrpc_rule_set_init(&tmp_rset);
 953         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 954         if (rc) {
 955                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 956                        tgt_name(tgt), rc);
 957                 return rc;
 958         }
 959
 960         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 961
 962         write_lock(&tgt->lut_sptlrpc_lock);
 963         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 964         tgt->lut_sptlrpc_rset = tmp_rset;
 965         write_unlock(&tgt->lut_sptlrpc_lock);
 966
 967         return 0;
 968 }
 969 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 970
 971 int tgt_connect(struct tgt_session_info *tsi)
 972 {
 973         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 974         struct obd_connect_data *reply;
 975         int                      rc;
 976
 977         ENTRY;
 978
 979         /* XXX: better to call this check right after getting new export but
 980          * before last_rcvd slot allocation to avoid server load upon insecure
 981          * connects. This is to be fixed after unifiyng all targets.
 982          */
 983         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 984         if (rc)
 985                 GOTO(out, rc);
 986
 987         /* To avoid exposing partially initialized connection flags, changes up
 988          * to this point have been staged in reply->ocd_connect_flags. Now that
 989          * connection handling has completed successfully, atomically update
 990          * the connect flags in the shared export data structure. LU-1623 */
 991         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
 992         spin_lock(&tsi->tsi_exp->exp_lock);
 993         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
 994         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
 995                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
 996                         reply->ocd_connect_flags2;
 997         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
 998         spin_unlock(&tsi->tsi_exp->exp_lock);
 999
1000         RETURN(0);
1001 out:
1002         obd_disconnect(class_export_get(tsi->tsi_exp));
1003         return rc;
1004 }
1005 EXPORT_SYMBOL(tgt_connect);
1006
1007 int tgt_disconnect(struct tgt_session_info *tsi)
1008 {
1009         int rc;
1010
1011         ENTRY;
1012
1013         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1014
1015         rc = target_handle_disconnect(tgt_ses_req(tsi));
1016         if (rc)
1017                 RETURN(err_serious(rc));
1018
1019         RETURN(rc);
1020 }
1021 EXPORT_SYMBOL(tgt_disconnect);
1022
1023 /*
1024  * Unified target OBD handlers
1025  */
1026 int tgt_obd_ping(struct tgt_session_info *tsi)
1027 {
1028         int rc;
1029
1030         ENTRY;
1031
1032         /* The target-specific part of OBD_PING request handling.
1033          * It controls Filter Modification Data (FMD) expiration each time
1034          * PING is received.
1035          *
1036          * Valid only for replayable targets, e.g. MDT and OFD
1037          */
1038         if (tsi->tsi_exp->exp_obd->obd_replayable)
1039                 tgt_fmd_expire(tsi->tsi_exp);
1040
1041         rc = req_capsule_server_pack(tsi->tsi_pill);
1042         if (rc)
1043                 RETURN(err_serious(rc));
1044
1045         RETURN(rc);
1046 }
1047 EXPORT_SYMBOL(tgt_obd_ping);
1048
1049 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1050 {
1051         return err_serious(-EOPNOTSUPP);
1052 }
1053
1054 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1055 {
1056         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1057         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1058         struct obd_export       *exp = req->rq_export;
1059         struct ptlrpc_bulk_desc *desc;
1060         struct l_wait_info      *lwi = &tti->tti_u.update.tti_wait_info;
1061         int                      i;
1062         int                      rc;
1063
1064         ENTRY;
1065
1066         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1067                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1068                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1069         if (desc == NULL)
1070                 RETURN(-ENOMEM);
1071
1072         for (i = 0; i < rdbuf->rb_nbufs; i++)
1073                 desc->bd_frag_ops->add_iov_frag(desc,
1074                                         rdbuf->rb_bufs[i].lb_buf,
1075                                         rdbuf->rb_bufs[i].lb_len);
1076
1077         rc = target_bulk_io(exp, desc, lwi);
1078         ptlrpc_free_bulk(desc);
1079         RETURN(rc);
1080 }
1081 EXPORT_SYMBOL(tgt_send_buffer);
1082
1083 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1084 {
1085         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1086         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1087         struct obd_export       *exp = req->rq_export;
1088         struct ptlrpc_bulk_desc *desc;
1089         struct l_wait_info      *lwi = &tti->tti_u.rdpg.tti_wait_info;
1090         int                      tmpcount;
1091         int                      tmpsize;
1092         int                      i;
1093         int                      rc;
1094
1095         ENTRY;
1096
1097         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1098                                     PTLRPC_BULK_PUT_SOURCE |
1099                                         PTLRPC_BULK_BUF_KIOV,
1100                                     MDS_BULK_PORTAL,
1101                                     &ptlrpc_bulk_kiov_pin_ops);
1102         if (desc == NULL)
1103                 RETURN(-ENOMEM);
1104
1105         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1106                 /* old client requires reply size in it's PAGE_SIZE,
1107                  * which is rdpg->rp_count */
1108                 nob = rdpg->rp_count;
1109
1110         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1111              i++, tmpcount -= tmpsize) {
1112                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1113                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1114                                                  tmpsize);
1115         }
1116
1117         LASSERT(desc->bd_nob == nob);
1118         rc = target_bulk_io(exp, desc, lwi);
1119         ptlrpc_free_bulk(desc);
1120         RETURN(rc);
1121 }
1122 EXPORT_SYMBOL(tgt_sendpage);
1123
1124 /*
1125  * OBD_IDX_READ handler
1126  */
1127 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1128 {
1129         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1130         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1131         struct idx_info         *req_ii, *rep_ii;
1132         int                      rc, i;
1133
1134         ENTRY;
1135
1136         memset(rdpg, 0, sizeof(*rdpg));
1137         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1138
1139         /* extract idx_info buffer from request & reply */
1140         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1141         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1142                 RETURN(err_serious(-EPROTO));
1143
1144         rc = req_capsule_server_pack(tsi->tsi_pill);
1145         if (rc)
1146                 RETURN(err_serious(rc));
1147
1148         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1149         if (rep_ii == NULL)
1150                 RETURN(err_serious(-EFAULT));
1151         rep_ii->ii_magic = IDX_INFO_MAGIC;
1152
1153         /* extract hash to start with */
1154         rdpg->rp_hash = req_ii->ii_hash_start;
1155
1156         /* extract requested attributes */
1157         rdpg->rp_attrs = req_ii->ii_attrs;
1158
1159         /* check that fid packed in request is valid and supported */
1160         if (!fid_is_sane(&req_ii->ii_fid))
1161                 RETURN(-EINVAL);
1162         rep_ii->ii_fid = req_ii->ii_fid;
1163
1164         /* copy flags */
1165         rep_ii->ii_flags = req_ii->ii_flags;
1166
1167         /* compute number of pages to allocate, ii_count is the number of 4KB
1168          * containers */
1169         if (req_ii->ii_count <= 0)
1170                 GOTO(out, rc = -EFAULT);
1171         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1172                                exp_max_brw_size(tsi->tsi_exp));
1173         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1174
1175         /* allocate pages to store the containers */
1176         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1177         if (rdpg->rp_pages == NULL)
1178                 GOTO(out, rc = -ENOMEM);
1179         for (i = 0; i < rdpg->rp_npages; i++) {
1180                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1181                 if (rdpg->rp_pages[i] == NULL)
1182                         GOTO(out, rc = -ENOMEM);
1183         }
1184
1185         /* populate pages with key/record pairs */
1186         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1187         if (rc < 0)
1188                 GOTO(out, rc);
1189
1190         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1191                  "asked %d > %d\n", rc, rdpg->rp_count);
1192
1193         /* send pages to client */
1194         rc = tgt_sendpage(tsi, rdpg, rc);
1195         if (rc)
1196                 GOTO(out, rc);
1197         EXIT;
1198 out:
1199         if (rdpg->rp_pages) {
1200                 for (i = 0; i < rdpg->rp_npages; i++)
1201                         if (rdpg->rp_pages[i])
1202                                 __free_page(rdpg->rp_pages[i]);
1203                 OBD_FREE(rdpg->rp_pages,
1204                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1205         }
1206         return rc;
1207 }
1208
1209 struct tgt_handler tgt_obd_handlers[] = {
1210 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1211 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1212 };
1213 EXPORT_SYMBOL(tgt_obd_handlers);
1214
1215 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1216              struct dt_object *obj, __u64 start, __u64 end)
1217 {
1218         int rc = 0;
1219
1220         ENTRY;
1221
1222         /* if no objid is specified, it means "sync whole filesystem" */
1223         if (obj == NULL) {
1224                 rc = dt_sync(env, tgt->lut_bottom);
1225         } else if (dt_version_get(env, obj) >
1226                    tgt->lut_obd->obd_last_committed) {
1227                 rc = dt_object_sync(env, obj, start, end);
1228         }
1229         atomic_inc(&tgt->lut_sync_count);
1230
1231         RETURN(rc);
1232 }
1233 EXPORT_SYMBOL(tgt_sync);
1234 /*
1235  * Unified target DLM handlers.
1236  */
1237
1238 /**
1239  * Unified target BAST
1240  *
1241  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1242  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1243  * but if it's cross-MDT lock, because remote object version is not set, a
1244  * filesystem sync is needed.
1245  *
1246  * \param lock server side lock
1247  * \param desc lock desc
1248  * \param data ldlm_cb_set_arg
1249  * \param flag  indicates whether this cancelling or blocking callback
1250  * \retval      0 on success
1251  * \retval      negative number on error
1252  */
1253 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1254                             void *data, int flag)
1255 {
1256         struct lu_env            env;
1257         struct lu_target        *tgt;
1258         struct dt_object        *obj = NULL;
1259         struct lu_fid            fid;
1260         int                      rc = 0;
1261
1262         ENTRY;
1263
1264         tgt = class_exp2tgt(lock->l_export);
1265
1266         if (unlikely(tgt == NULL)) {
1267                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1268                        class_exp2obd(lock->l_export)->obd_name);
1269                 RETURN(-EINVAL);
1270         }
1271
1272         if (flag == LDLM_CB_CANCELING &&
1273             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1274             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1275              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1276               ldlm_is_cbpending(lock))) &&
1277             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1278              lock->l_resource->lr_type == LDLM_EXTENT)) {
1279                 __u64 start = 0;
1280                 __u64 end = OBD_OBJECT_EOF;
1281
1282                 rc = lu_env_init(&env, LCT_DT_THREAD);
1283                 if (unlikely(rc != 0))
1284                         GOTO(err, rc);
1285
1286                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1287                                    tgt->lut_lsd.lsd_osd_index);
1288
1289                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1290                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1291                         if (IS_ERR(obj))
1292                                 GOTO(err_env, rc = PTR_ERR(obj));
1293
1294                         if (!dt_object_exists(obj))
1295                                 GOTO(err_put, rc = -ENOENT);
1296
1297                         start = lock->l_policy_data.l_extent.start;
1298                         end = lock->l_policy_data.l_extent.end;
1299                 }
1300
1301                 rc = tgt_sync(&env, tgt, obj, start, end);
1302                 if (rc < 0) {
1303                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1304                                "cancel: rc = %d\n",
1305                                tgt_name(tgt), PFID(&fid),
1306                                lock->l_policy_data.l_extent.start,
1307                                lock->l_policy_data.l_extent.end, rc);
1308                 }
1309 err_put:
1310                 if (obj != NULL)
1311                         dt_object_put(&env, obj);
1312 err_env:
1313                 lu_env_fini(&env);
1314         }
1315 err:
1316         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1317         RETURN(rc);
1318 }
1319
1320 static struct ldlm_callback_suite tgt_dlm_cbs = {
1321         .lcs_completion = ldlm_server_completion_ast,
1322         .lcs_blocking   = tgt_blocking_ast,
1323         .lcs_glimpse    = ldlm_server_glimpse_ast
1324 };
1325
1326 int tgt_enqueue(struct tgt_session_info *tsi)
1327 {
1328         struct ptlrpc_request *req = tgt_ses_req(tsi);
1329         int rc;
1330
1331         ENTRY;
1332         /*
1333          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1334          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1335          */
1336         LASSERT(tsi->tsi_dlm_req != NULL);
1337         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1338                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1339         if (rc)
1340                 RETURN(err_serious(rc));
1341
1342         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1343         case LUT_FAIL_MDT:
1344                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1345                 break;
1346         case LUT_FAIL_OST:
1347                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1348                 break;
1349         case LUT_FAIL_MGT:
1350                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1351                 break;
1352         default:
1353                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1354                 break;
1355         }
1356         RETURN(req->rq_status);
1357 }
1358 EXPORT_SYMBOL(tgt_enqueue);
1359
1360 int tgt_convert(struct tgt_session_info *tsi)
1361 {
1362         struct ptlrpc_request *req = tgt_ses_req(tsi);
1363         int rc;
1364
1365         ENTRY;
1366         LASSERT(tsi->tsi_dlm_req);
1367         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1368         if (rc)
1369                 RETURN(err_serious(rc));
1370
1371         RETURN(req->rq_status);
1372 }
1373
1374 int tgt_bl_callback(struct tgt_session_info *tsi)
1375 {
1376         return err_serious(-EOPNOTSUPP);
1377 }
1378
1379 int tgt_cp_callback(struct tgt_session_info *tsi)
1380 {
1381         return err_serious(-EOPNOTSUPP);
1382 }
1383
1384 /* generic LDLM target handler */
1385 struct tgt_handler tgt_dlm_handlers[] = {
1386 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1387 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1388 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1389 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1390 };
1391 EXPORT_SYMBOL(tgt_dlm_handlers);
1392
1393 /*
1394  * Unified target LLOG handlers.
1395  */
1396 int tgt_llog_open(struct tgt_session_info *tsi)
1397 {
1398         int rc;
1399
1400         ENTRY;
1401
1402         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1403
1404         RETURN(rc);
1405 }
1406 EXPORT_SYMBOL(tgt_llog_open);
1407
1408 int tgt_llog_read_header(struct tgt_session_info *tsi)
1409 {
1410         int rc;
1411
1412         ENTRY;
1413
1414         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1415
1416         RETURN(rc);
1417 }
1418 EXPORT_SYMBOL(tgt_llog_read_header);
1419
1420 int tgt_llog_next_block(struct tgt_session_info *tsi)
1421 {
1422         int rc;
1423
1424         ENTRY;
1425
1426         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1427
1428         RETURN(rc);
1429 }
1430 EXPORT_SYMBOL(tgt_llog_next_block);
1431
1432 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1433 {
1434         int rc;
1435
1436         ENTRY;
1437
1438         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1439
1440         RETURN(rc);
1441 }
1442 EXPORT_SYMBOL(tgt_llog_prev_block);
1443
1444 /* generic llog target handler */
1445 struct tgt_handler tgt_llog_handlers[] = {
1446 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1447 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1448 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1449 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1450 };
1451 EXPORT_SYMBOL(tgt_llog_handlers);
1452
1453 /*
1454  * sec context handlers
1455  */
1456 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1457 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1458 {
1459         return 0;
1460 }
1461
1462 struct tgt_handler tgt_sec_ctx_handlers[] = {
1463 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1464 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1465 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1466 };
1467 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1468
1469 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1470                                  struct dt_device *key,
1471                                  struct lfsck_req_local *lrl,
1472                                  struct thandle *th) = NULL;
1473
1474 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1475                                                       struct dt_device *,
1476                                                       struct lfsck_req_local *,
1477                                                       struct thandle *))
1478 {
1479         tgt_lfsck_in_notify_local = notify;
1480 }
1481 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1482
1483 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1484                            struct dt_device *key,
1485                            struct lfsck_request *lr) = NULL;
1486
1487 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1488                                                 struct dt_device *,
1489                                                 struct lfsck_request *))
1490 {
1491         tgt_lfsck_in_notify = notify;
1492 }
1493 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1494
1495 static int (*tgt_lfsck_query)(const struct lu_env *env,
1496                               struct dt_device *key,
1497                               struct lfsck_request *req,
1498                               struct lfsck_reply *rep,
1499                               struct lfsck_query *que) = NULL;
1500
1501 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1502                                            struct dt_device *,
1503                                            struct lfsck_request *,
1504                                            struct lfsck_reply *,
1505                                            struct lfsck_query *))
1506 {
1507         tgt_lfsck_query = query;
1508 }
1509 EXPORT_SYMBOL(tgt_register_lfsck_query);
1510
1511 /* LFSCK request handlers */
1512 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1513 {
1514         const struct lu_env     *env = tsi->tsi_env;
1515         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1516         struct lfsck_request    *lr;
1517         int                      rc;
1518         ENTRY;
1519
1520         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1521         if (lr == NULL)
1522                 RETURN(-EPROTO);
1523
1524         rc = tgt_lfsck_in_notify(env, key, lr);
1525
1526         RETURN(rc);
1527 }
1528
1529 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1530 {
1531         struct lfsck_request    *request;
1532         struct lfsck_reply      *reply;
1533         int                      rc;
1534         ENTRY;
1535
1536         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1537         if (request == NULL)
1538                 RETURN(-EPROTO);
1539
1540         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1541         if (reply == NULL)
1542                 RETURN(-ENOMEM);
1543
1544         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1545                              request, reply, NULL);
1546
1547         RETURN(rc < 0 ? rc : 0);
1548 }
1549
1550 struct tgt_handler tgt_lfsck_handlers[] = {
1551 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1552 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1553 };
1554 EXPORT_SYMBOL(tgt_lfsck_handlers);
1555
1556 /*
1557  * initialize per-thread page pool (bug 5137).
1558  */
1559 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1560 {
1561         struct tgt_thread_big_cache *tbc;
1562
1563         ENTRY;
1564
1565         LASSERT(thread != NULL);
1566         LASSERT(thread->t_data == NULL);
1567
1568         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1569         if (tbc == NULL)
1570                 RETURN(-ENOMEM);
1571         thread->t_data = tbc;
1572         RETURN(0);
1573 }
1574 EXPORT_SYMBOL(tgt_io_thread_init);
1575
1576 /*
1577  * free per-thread pool created by tgt_thread_init().
1578  */
1579 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1580 {
1581         struct tgt_thread_big_cache *tbc;
1582
1583         ENTRY;
1584
1585         LASSERT(thread != NULL);
1586
1587         /*
1588          * be prepared to handle partially-initialized pools (because this is
1589          * called from ost_io_thread_init() for cleanup.
1590          */
1591         tbc = thread->t_data;
1592         if (tbc != NULL) {
1593                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1594                 thread->t_data = NULL;
1595         }
1596         EXIT;
1597 }
1598 EXPORT_SYMBOL(tgt_io_thread_done);
1599
1600 /**
1601  * Helper function for getting Data-on-MDT file server DLM lock
1602  * if asked by client.
1603  */
1604 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1605                       struct lustre_handle *lh, int mode, __u64 *flags)
1606 {
1607         union ldlm_policy_data policy = {
1608                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1609         };
1610         int rc;
1611
1612         ENTRY;
1613
1614         LASSERT(lh != NULL);
1615         LASSERT(ns != NULL);
1616         LASSERT(!lustre_handle_is_used(lh));
1617
1618         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1619                                     flags, ldlm_blocking_ast,
1620                                     ldlm_completion_ast, ldlm_glimpse_ast,
1621                                     NULL, 0, LVB_T_NONE, NULL, lh);
1622
1623         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1624 }
1625 EXPORT_SYMBOL(tgt_mdt_data_lock);
1626
1627 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1628 {
1629         LASSERT(lustre_handle_is_used(lh));
1630         ldlm_lock_decref(lh, mode);
1631 }
1632 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1633
1634 /**
1635  * Helper function for getting server side [start, start+count] DLM lock
1636  * if asked by client.
1637  */
1638 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1639                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1640                     struct lustre_handle *lh, int mode, __u64 *flags)
1641 {
1642         union ldlm_policy_data policy;
1643         int rc;
1644
1645         ENTRY;
1646
1647         LASSERT(lh != NULL);
1648         LASSERT(ns != NULL);
1649         LASSERT(!lustre_handle_is_used(lh));
1650
1651         policy.l_extent.gid = 0;
1652         policy.l_extent.start = start & PAGE_MASK;
1653
1654         /*
1655          * If ->o_blocks is EOF it means "lock till the end of the file".
1656          * Otherwise, it's size of an extent or hole being punched (in bytes).
1657          */
1658         if (end == OBD_OBJECT_EOF || end < start)
1659                 policy.l_extent.end = OBD_OBJECT_EOF;
1660         else
1661                 policy.l_extent.end = end | ~PAGE_MASK;
1662
1663         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1664                                     mode, flags, ldlm_blocking_ast,
1665                                     ldlm_completion_ast, ldlm_glimpse_ast,
1666                                     NULL, 0, LVB_T_NONE, NULL, lh);
1667         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1668 }
1669 EXPORT_SYMBOL(tgt_extent_lock);
1670
1671 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1672 {
1673         LASSERT(lustre_handle_is_used(lh));
1674         ldlm_lock_decref(lh, mode);
1675 }
1676 EXPORT_SYMBOL(tgt_extent_unlock);
1677
1678 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1679                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1680                         struct niobuf_remote *nb, struct lustre_handle *lh,
1681                         enum ldlm_mode mode)
1682 {
1683         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1684         __u64                    flags = 0;
1685         int                      nrbufs = obj->ioo_bufcnt;
1686         int                      i;
1687         int                      rc;
1688
1689         ENTRY;
1690
1691         LASSERT(mode == LCK_PR || mode == LCK_PW);
1692         LASSERT(!lustre_handle_is_used(lh));
1693
1694         if (ns->ns_obd->obd_recovering)
1695                 RETURN(0);
1696
1697         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1698                 RETURN(0);
1699
1700         for (i = 1; i < nrbufs; i++)
1701                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1702                         RETURN(-EFAULT);
1703
1704         /* MDT IO for data-on-mdt */
1705         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1706                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1707         else
1708                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1709                                      nb[nrbufs - 1].rnb_offset +
1710                                      nb[nrbufs - 1].rnb_len - 1,
1711                                      lh, mode, &flags);
1712         RETURN(rc);
1713 }
1714
1715 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1716                            struct lustre_handle *lh, enum ldlm_mode mode)
1717 {
1718         ENTRY;
1719
1720         LASSERT(mode == LCK_PR || mode == LCK_PW);
1721         LASSERT((obj->ioo_bufcnt > 0 &&
1722                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1723                 lustre_handle_is_used(lh));
1724
1725         if (lustre_handle_is_used(lh))
1726                 tgt_extent_unlock(lh, mode);
1727         EXIT;
1728 }
1729 static int tgt_checksum_niobuf(struct lu_target *tgt,
1730                                  struct niobuf_local *local_nb, int npages,
1731                                  int opc, enum cksum_types cksum_type,
1732                                  __u32 *cksum)
1733 {
1734         struct ahash_request           *req;
1735         unsigned int                    bufsize;
1736         int                             i, err;
1737         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1738
1739         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1740         if (IS_ERR(req)) {
1741                 CERROR("%s: unable to initialize checksum hash %s\n",
1742                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1743                 return PTR_ERR(req);
1744         }
1745
1746         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1747         for (i = 0; i < npages; i++) {
1748                 /* corrupt the data before we compute the checksum, to
1749                  * simulate a client->OST data error */
1750                 if (i == 0 && opc == OST_WRITE &&
1751                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1752                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1753                         int len = local_nb[i].lnb_len;
1754                         struct page *np = tgt_page_to_corrupt;
1755
1756                         if (np) {
1757                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1758                                                         KM_USER0);
1759                                 char *ptr2 = page_address(np);
1760
1761                                 memcpy(ptr2 + off, ptr + off, len);
1762                                 memcpy(ptr2 + off, "bad3", min(4, len));
1763                                 ll_kunmap_atomic(ptr, KM_USER0);
1764
1765                                 /* LU-8376 to preserve original index for
1766                                  * display in dump_all_bulk_pages() */
1767                                 np->index = i;
1768
1769                                 cfs_crypto_hash_update_page(req, np, off,
1770                                                             len);
1771                                 continue;
1772                         } else {
1773                                 CERROR("%s: can't alloc page for corruption\n",
1774                                        tgt_name(tgt));
1775                         }
1776                 }
1777                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1778                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1779                                   local_nb[i].lnb_len);
1780
1781                  /* corrupt the data after we compute the checksum, to
1782                  * simulate an OST->client data error */
1783                 if (i == 0 && opc == OST_READ &&
1784                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1785                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1786                         int len = local_nb[i].lnb_len;
1787                         struct page *np = tgt_page_to_corrupt;
1788
1789                         if (np) {
1790                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1791                                                         KM_USER0);
1792                                 char *ptr2 = page_address(np);
1793
1794                                 memcpy(ptr2 + off, ptr + off, len);
1795                                 memcpy(ptr2 + off, "bad4", min(4, len));
1796                                 ll_kunmap_atomic(ptr, KM_USER0);
1797
1798                                 /* LU-8376 to preserve original index for
1799                                  * display in dump_all_bulk_pages() */
1800                                 np->index = i;
1801
1802                                 cfs_crypto_hash_update_page(req, np, off,
1803                                                             len);
1804                                 continue;
1805                         } else {
1806                                 CERROR("%s: can't alloc page for corruption\n",
1807                                        tgt_name(tgt));
1808                         }
1809                 }
1810         }
1811
1812         bufsize = sizeof(*cksum);
1813         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1814
1815         return 0;
1816 }
1817
1818 char dbgcksum_file_name[PATH_MAX];
1819
1820 static void dump_all_bulk_pages(struct obdo *oa, int count,
1821                                 struct niobuf_local *local_nb,
1822                                 __u32 server_cksum, __u32 client_cksum)
1823 {
1824         struct file *filp;
1825         int rc, i;
1826         unsigned int len;
1827         char *buf;
1828
1829         /* will only keep dump of pages on first error for the same range in
1830          * file/fid, not during the resends/retries. */
1831         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1832                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1833                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1834                   libcfs_debug_file_path_arr :
1835                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1836                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1837                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1838                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1839                  local_nb[0].lnb_file_offset,
1840                  local_nb[count-1].lnb_file_offset +
1841                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1842         filp = filp_open(dbgcksum_file_name,
1843                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1844         if (IS_ERR(filp)) {
1845                 rc = PTR_ERR(filp);
1846                 if (rc == -EEXIST)
1847                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1848                                "checksum error: rc = %d\n", dbgcksum_file_name,
1849                                rc);
1850                 else
1851                         CERROR("%s: can't open to dump pages with checksum "
1852                                "error: rc = %d\n", dbgcksum_file_name, rc);
1853                 return;
1854         }
1855
1856         for (i = 0; i < count; i++) {
1857                 len = local_nb[i].lnb_len;
1858                 buf = kmap(local_nb[i].lnb_page);
1859                 while (len != 0) {
1860                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1861                         if (rc < 0) {
1862                                 CERROR("%s: wanted to write %u but got %d "
1863                                        "error\n", dbgcksum_file_name, len, rc);
1864                                 break;
1865                         }
1866                         len -= rc;
1867                         buf += rc;
1868                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1869                                dbgcksum_file_name, rc);
1870                 }
1871                 kunmap(local_nb[i].lnb_page);
1872         }
1873
1874         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1875         if (rc)
1876                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1877         filp_close(filp, NULL);
1878         return;
1879 }
1880
1881 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1882                                struct obd_export *exp, struct obdo *oa,
1883                                const struct lnet_process_id *peer,
1884                                __u32 client_cksum, __u32 server_cksum,
1885                                enum cksum_types server_cksum_type)
1886 {
1887         char *msg;
1888         enum cksum_types cksum_type;
1889         loff_t start, end;
1890
1891         /* unlikely to happen and only if resend does not occur due to cksum
1892          * control failure on Client */
1893         if (unlikely(server_cksum == client_cksum)) {
1894                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1895                        client_cksum);
1896                 return 0;
1897         }
1898
1899         if (exp->exp_obd->obd_checksum_dump)
1900                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1901                                     client_cksum);
1902
1903         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1904                                            oa->o_flags : 0);
1905
1906         if (cksum_type != server_cksum_type)
1907                 msg = "the server may have not used the checksum type specified"
1908                       " in the original request - likely a protocol problem";
1909         else
1910                 msg = "should have changed on the client or in transit";
1911
1912         start = local_nb[0].lnb_file_offset;
1913         end = local_nb[npages-1].lnb_file_offset +
1914                                         local_nb[npages-1].lnb_len - 1;
1915
1916         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1917                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1918                 " %x (type %x), server csum %x (type %x)\n",
1919                 exp->exp_obd->obd_name,
1920                 msg, libcfs_nid2str(peer->nid),
1921                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1922                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1923                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1924                 POSTID(&oa->o_oi),
1925                 start, end, client_cksum, cksum_type, server_cksum,
1926                 server_cksum_type);
1927
1928         return 1;
1929 }
1930
1931 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1932                              unsigned char *buf, int size)
1933 {
1934         int     i, off, len, copied = size;
1935         char    *ptr;
1936
1937         for (i = 0; i < npages; i++) {
1938                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1939                 len = local[i].lnb_len;
1940
1941                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1942                        i, off, len, size);
1943                 if (len > size)
1944                         return -EINVAL;
1945
1946                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1947                 memcpy(buf + off, ptr, len);
1948                 ll_kunmap_atomic(ptr, KM_USER0);
1949                 buf += len;
1950                 size -= len;
1951         }
1952         return copied - size;
1953 }
1954
1955 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1956                                      struct niobuf_local *local_nb,
1957                                      int npages, int opc,
1958                                      obd_dif_csum_fn *fn,
1959                                      int sector_size,
1960                                      u32 *check_sum)
1961 {
1962         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1963         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1964         const char *obd_name = tgt->lut_obd->obd_name;
1965         struct ahash_request *req;
1966         unsigned int bufsize;
1967         unsigned char *buffer;
1968         struct page *__page;
1969         __u16 *guard_start;
1970         int guard_number;
1971         int used_number = 0;
1972         __u32 cksum;
1973         int rc = 0;
1974         int used;
1975         int i;
1976
1977         __page = alloc_page(GFP_KERNEL);
1978         if (__page == NULL)
1979                 return -ENOMEM;
1980
1981         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1982         if (IS_ERR(req)) {
1983                 CERROR("%s: unable to initialize checksum hash %s\n",
1984                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1985                 return PTR_ERR(req);
1986         }
1987
1988         buffer = kmap(__page);
1989         guard_start = (__u16 *)buffer;
1990         guard_number = PAGE_SIZE / sizeof(*guard_start);
1991         for (i = 0; i < npages; i++) {
1992                 /* corrupt the data before we compute the checksum, to
1993                  * simulate a client->OST data error */
1994                 if (i == 0 && opc == OST_WRITE &&
1995                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1996                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1997                         int len = local_nb[i].lnb_len;
1998                         struct page *np = tgt_page_to_corrupt;
1999
2000                         if (np) {
2001                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2002                                                         KM_USER0);
2003                                 char *ptr2 = page_address(np);
2004
2005                                 memcpy(ptr2 + off, ptr + off, len);
2006                                 memcpy(ptr2 + off, "bad3", min(4, len));
2007                                 ll_kunmap_atomic(ptr, KM_USER0);
2008
2009                                 /* LU-8376 to preserve original index for
2010                                  * display in dump_all_bulk_pages() */
2011                                 np->index = i;
2012
2013                                 cfs_crypto_hash_update_page(req, np, off,
2014                                                             len);
2015                                 continue;
2016                         } else {
2017                                 CERROR("%s: can't alloc page for corruption\n",
2018                                        tgt_name(tgt));
2019                         }
2020                 }
2021
2022                 /*
2023                  * The left guard number should be able to hold checksums of a
2024                  * whole page
2025                  */
2026                 if (t10_cksum_type && opc == OST_READ &&
2027                     local_nb[i].lnb_guard_disk) {
2028                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2029                         if (used > (guard_number - used_number)) {
2030                                 rc = -E2BIG;
2031                                 break;
2032                         }
2033                         memcpy(guard_start + used_number,
2034                                local_nb[i].lnb_guards,
2035                                used * sizeof(*local_nb[i].lnb_guards));
2036                 } else {
2037                         rc = obd_page_dif_generate_buffer(obd_name,
2038                                 local_nb[i].lnb_page,
2039                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2040                                 local_nb[i].lnb_len, guard_start + used_number,
2041                                 guard_number - used_number, &used, sector_size,
2042                                 fn);
2043                         if (rc)
2044                                 break;
2045                 }
2046
2047                 LASSERT(used <= MAX_GUARD_NUMBER);
2048                 /*
2049                  * If disk support T10PI checksum, copy guards to local_nb.
2050                  * If the write is partial page, do not use the guards for bio
2051                  * submission since the data might not be full-sector. The bio
2052                  * guards will be generated later based on the full sectors. If
2053                  * the sector size is 512B rather than 4 KB, or the page size
2054                  * is larger than 4KB, this might drop some useful guards for
2055                  * partial page write, but it will only add minimal extra time
2056                  * of checksum calculation.
2057                  */
2058                 if (t10_cksum_type && opc == OST_WRITE &&
2059                     local_nb[i].lnb_len == PAGE_SIZE) {
2060                         local_nb[i].lnb_guard_rpc = 1;
2061                         memcpy(local_nb[i].lnb_guards,
2062                                guard_start + used_number,
2063                                used * sizeof(*local_nb[i].lnb_guards));
2064                 }
2065
2066                 used_number += used;
2067                 if (used_number == guard_number) {
2068                         cfs_crypto_hash_update_page(req, __page, 0,
2069                                 used_number * sizeof(*guard_start));
2070                         used_number = 0;
2071                 }
2072
2073                  /* corrupt the data after we compute the checksum, to
2074                  * simulate an OST->client data error */
2075                 if (unlikely(i == 0 && opc == OST_READ &&
2076                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2077                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2078                         int len = local_nb[i].lnb_len;
2079                         struct page *np = tgt_page_to_corrupt;
2080
2081                         if (np) {
2082                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2083                                                         KM_USER0);
2084                                 char *ptr2 = page_address(np);
2085
2086                                 memcpy(ptr2 + off, ptr + off, len);
2087                                 memcpy(ptr2 + off, "bad4", min(4, len));
2088                                 ll_kunmap_atomic(ptr, KM_USER0);
2089
2090                                 /* LU-8376 to preserve original index for
2091                                  * display in dump_all_bulk_pages() */
2092                                 np->index = i;
2093
2094                                 cfs_crypto_hash_update_page(req, np, off,
2095                                                             len);
2096                                 continue;
2097                         } else {
2098                                 CERROR("%s: can't alloc page for corruption\n",
2099                                        tgt_name(tgt));
2100                         }
2101                 }
2102         }
2103         kunmap(__page);
2104         if (rc)
2105                 GOTO(out, rc);
2106
2107         if (used_number != 0)
2108                 cfs_crypto_hash_update_page(req, __page, 0,
2109                         used_number * sizeof(*guard_start));
2110
2111         bufsize = sizeof(cksum);
2112         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2113
2114         if (rc == 0)
2115                 *check_sum = cksum;
2116 out:
2117         __free_page(__page);
2118         return rc;
2119 }
2120
2121 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2122                                   enum cksum_types cksum_type,
2123                                   struct niobuf_local *local_nb,
2124                                   int npages, int opc, u32 *check_sum)
2125 {
2126         obd_dif_csum_fn *fn = NULL;
2127         int sector_size = 0;
2128         int rc;
2129
2130         ENTRY;
2131         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2132
2133         if (fn)
2134                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2135                                                opc, fn, sector_size,
2136                                                check_sum);
2137         else
2138                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2139                                          cksum_type, check_sum);
2140         RETURN(rc);
2141 }
2142
2143 int tgt_brw_read(struct tgt_session_info *tsi)
2144 {
2145         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2146         struct ptlrpc_bulk_desc *desc = NULL;
2147         struct obd_export       *exp = tsi->tsi_exp;
2148         struct niobuf_remote    *remote_nb;
2149         struct niobuf_local     *local_nb;
2150         struct obd_ioobj        *ioo;
2151         struct ost_body         *body, *repbody;
2152         struct l_wait_info       lwi;
2153         struct lustre_handle     lockh = { 0 };
2154         int                      npages, nob = 0, rc, i, no_reply = 0,
2155                                  npages_read;
2156         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2157         const char *obd_name = exp->exp_obd->obd_name;
2158
2159         ENTRY;
2160
2161         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2162             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2163                 CERROR("%s: deny read request from %s to portal %u\n",
2164                        tgt_name(tsi->tsi_tgt),
2165                        obd_export_nid2str(req->rq_export),
2166                        ptlrpc_req2svc(req)->srv_req_portal);
2167                 RETURN(-EPROTO);
2168         }
2169
2170         req->rq_bulk_read = 1;
2171
2172         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2173                 RETURN(-EIO);
2174
2175         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2176                          cfs_fail_val : (obd_timeout + 1) / 4);
2177
2178         /* Check if there is eviction in progress, and if so, wait for it to
2179          * finish */
2180         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2181                 /* We do not care how long it takes */
2182                 lwi = LWI_INTR(NULL, NULL);
2183                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2184                          !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2185                          &lwi);
2186         }
2187
2188         /* There must be big cache in current thread to process this request
2189          * if it is NULL then something went wrong and it wasn't allocated,
2190          * report -ENOMEM in that case */
2191         if (tbc == NULL)
2192                 RETURN(-ENOMEM);
2193
2194         body = tsi->tsi_ost_body;
2195         LASSERT(body != NULL);
2196
2197         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2198         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2199
2200         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2201         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2202
2203         local_nb = tbc->local;
2204
2205         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2206                           &lockh, LCK_PR);
2207         if (rc != 0)
2208                 RETURN(rc);
2209
2210         /*
2211          * If getting the lock took more time than
2212          * client was willing to wait, drop it. b=11330
2213          */
2214         if (ktime_get_real_seconds() > req->rq_deadline ||
2215             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2216                 no_reply = 1;
2217                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2218                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2219                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2220                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2221                 GOTO(out_lock, rc = -ETIMEDOUT);
2222         }
2223
2224         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2225         repbody->oa = body->oa;
2226
2227         npages = PTLRPC_MAX_BRW_PAGES;
2228         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2229                         ioo, remote_nb, &npages, local_nb);
2230         if (rc != 0)
2231                 GOTO(out_lock, rc);
2232
2233         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2234                 desc = NULL;
2235         } else {
2236                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2237                                             PTLRPC_BULK_PUT_SOURCE |
2238                                                 PTLRPC_BULK_BUF_KIOV,
2239                                             OST_BULK_PORTAL,
2240                                             &ptlrpc_bulk_kiov_nopin_ops);
2241                 if (desc == NULL)
2242                         GOTO(out_commitrw, rc = -ENOMEM);
2243         }
2244
2245         nob = 0;
2246         npages_read = npages;
2247         for (i = 0; i < npages; i++) {
2248                 int page_rc = local_nb[i].lnb_rc;
2249
2250                 if (page_rc < 0) {
2251                         rc = page_rc;
2252                         npages_read = i;
2253                         break;
2254                 }
2255
2256                 nob += page_rc;
2257                 if (page_rc != 0 && desc != NULL) { /* some data! */
2258                         LASSERT(local_nb[i].lnb_page != NULL);
2259                         desc->bd_frag_ops->add_kiov_frag
2260                           (desc, local_nb[i].lnb_page,
2261                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2262                            page_rc);
2263                 }
2264
2265                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2266                         local_nb[i].lnb_len = page_rc;
2267                         npages_read = i + (page_rc != 0 ? 1 : 0);
2268                         /* All subsequent pages should be 0 */
2269                         while (++i < npages)
2270                                 LASSERT(local_nb[i].lnb_rc == 0);
2271                         break;
2272                 }
2273         }
2274         if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2275             nob != cfs_fail_val)
2276                 rc = -E2BIG;
2277
2278         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2279                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2280                            body->oa.o_flags : 0;
2281                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2282
2283                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2284                                                           cksum_type);
2285                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2286
2287                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2288                                             local_nb, npages_read, OST_READ,
2289                                             &repbody->oa.o_cksum);
2290                 if (rc < 0)
2291                         GOTO(out_commitrw, rc);
2292                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2293                        repbody->oa.o_cksum);
2294
2295                 /* if a resend it could be for a cksum error, so check Server
2296                  * cksum with returned Client cksum (this should even cover
2297                  * zero-cksum case) */
2298                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2299                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2300                         check_read_checksum(local_nb, npages_read, exp,
2301                                             &body->oa, &req->rq_peer,
2302                                             body->oa.o_cksum,
2303                                             repbody->oa.o_cksum, cksum_type);
2304         } else {
2305                 repbody->oa.o_valid = 0;
2306         }
2307         /* We're finishing using body->oa as an input variable */
2308
2309         /* Check if client was evicted while we were doing i/o before touching
2310          * network */
2311         if (rc == 0) {
2312                 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2313                         unsigned char *short_io_buf;
2314                         int short_io_size;
2315
2316                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2317                                                               &RMF_SHORT_IO);
2318                         short_io_size = req_capsule_get_size(&req->rq_pill,
2319                                                              &RMF_SHORT_IO,
2320                                                              RCL_SERVER);
2321                         rc = tgt_pages2shortio(local_nb, npages_read,
2322                                                short_io_buf, short_io_size);
2323                         if (rc >= 0)
2324                                 req_capsule_shrink(&req->rq_pill,
2325                                                    &RMF_SHORT_IO, rc,
2326                                                    RCL_SERVER);
2327                         rc = rc > 0 ? 0 : rc;
2328                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2329                         rc = target_bulk_io(exp, desc, &lwi);
2330                 }
2331                 no_reply = rc != 0;
2332         } else {
2333                 if (body->oa.o_flags & OBD_FL_SHORT_IO)
2334                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2335                                            RCL_SERVER);
2336         }
2337
2338 out_commitrw:
2339         /* Must commit after prep above in all cases */
2340         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2341                           remote_nb, npages, local_nb, rc);
2342 out_lock:
2343         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2344
2345         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2346                 ptlrpc_free_bulk(desc);
2347
2348         LASSERT(rc <= 0);
2349         if (rc == 0) {
2350                 rc = nob;
2351                 ptlrpc_lprocfs_brw(req, nob);
2352         } else if (no_reply) {
2353                 req->rq_no_reply = 1;
2354                 /* reply out callback would free */
2355                 ptlrpc_req_drop_rs(req);
2356                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2357                               "client will retry: rc %d\n",
2358                               obd_name,
2359                               obd_uuid2str(&exp->exp_client_uuid),
2360                               obd_export_nid2str(exp), rc);
2361         }
2362         /* send a bulk after reply to simulate a network delay or reordering
2363          * by a router - Note that !desc implies short io, so there is no bulk
2364          * to reorder. */
2365         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2366             desc) {
2367                 wait_queue_head_t        waitq;
2368                 struct l_wait_info       lwi1;
2369
2370                 CDEBUG(D_INFO, "reorder BULK\n");
2371                 init_waitqueue_head(&waitq);
2372
2373                 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2374                 l_wait_event(waitq, 0, &lwi1);
2375                 target_bulk_io(exp, desc, &lwi);
2376                 ptlrpc_free_bulk(desc);
2377         }
2378
2379         RETURN(rc);
2380 }
2381 EXPORT_SYMBOL(tgt_brw_read);
2382
2383 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2384                              unsigned char *buf, int size)
2385 {
2386         int     i, off, len;
2387         char    *ptr;
2388
2389         for (i = 0; i < npages; i++) {
2390                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2391                 len = local[i].lnb_len;
2392
2393                 if (len == 0)
2394                         continue;
2395
2396                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2397                        i, off, len, size);
2398                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2399                 if (ptr == NULL)
2400                         return -EINVAL;
2401                 memcpy(ptr + off, buf, len < size ? len : size);
2402                 ll_kunmap_atomic(ptr, KM_USER0);
2403                 buf += len;
2404                 size -= len;
2405         }
2406         return 0;
2407 }
2408
2409 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2410                               struct ptlrpc_bulk_desc *desc,
2411                               struct niobuf_local *local_nb, int npages,
2412                               u32 client_cksum, u32 server_cksum,
2413                               bool mmap)
2414 {
2415         struct obd_export *exp = req->rq_export;
2416         struct ost_body *body;
2417         char *router = "";
2418         char *via = "";
2419
2420         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2421         LASSERT(body != NULL);
2422
2423         if (desc && req->rq_peer.nid != desc->bd_sender) {
2424                 via = " via ";
2425                 router = libcfs_nid2str(desc->bd_sender);
2426         }
2427
2428         if (exp->exp_obd->obd_checksum_dump)
2429                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2430                                     client_cksum);
2431
2432         if (mmap) {
2433                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2434                              client_cksum, server_cksum);
2435                 return;
2436         }
2437
2438         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2439                            DFID" object "DOSTID" extent [%llu-%llu"
2440                            "]: client csum %x, server csum %x\n",
2441                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2442                            via, router,
2443                            body->oa.o_valid & OBD_MD_FLFID ?
2444                            body->oa.o_parent_seq : (__u64)0,
2445                            body->oa.o_valid & OBD_MD_FLFID ?
2446                            body->oa.o_parent_oid : 0,
2447                            body->oa.o_valid & OBD_MD_FLFID ?
2448                            body->oa.o_parent_ver : 0,
2449                            POSTID(&body->oa.o_oi),
2450                            local_nb[0].lnb_file_offset,
2451                            local_nb[npages-1].lnb_file_offset +
2452                            local_nb[npages - 1].lnb_len - 1,
2453                            client_cksum, server_cksum);
2454 }
2455
2456 int tgt_brw_write(struct tgt_session_info *tsi)
2457 {
2458         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2459         struct ptlrpc_bulk_desc *desc = NULL;
2460         struct obd_export       *exp = req->rq_export;
2461         struct niobuf_remote    *remote_nb;
2462         struct niobuf_local     *local_nb;
2463         struct obd_ioobj        *ioo;
2464         struct ost_body         *body, *repbody;
2465         struct l_wait_info       lwi;
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         /* There must be big cache in current thread to process this request
2504          * if it is NULL then something went wrong and it wasn't allocated,
2505          * report -ENOMEM in that case */
2506         if (tbc == NULL)
2507                 RETURN(-ENOMEM);
2508
2509         body = tsi->tsi_ost_body;
2510         LASSERT(body != NULL);
2511
2512         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2513         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2514
2515         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2516                                         RCL_CLIENT) / sizeof(*ioo);
2517
2518         for (niocount = i = 0; i < objcount; i++)
2519                 niocount += ioo[i].ioo_bufcnt;
2520
2521         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2522         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2523         if (niocount != req_capsule_get_size(&req->rq_pill,
2524                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2525                         sizeof(*remote_nb))
2526                 RETURN(err_serious(-EPROTO));
2527
2528         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2529             ptlrpc_connection_is_local(exp->exp_connection))
2530                 memory_pressure_set();
2531
2532         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2533                              niocount * sizeof(*rcs));
2534         rc = req_capsule_server_pack(&req->rq_pill);
2535         if (rc != 0)
2536                 GOTO(out, rc = err_serious(rc));
2537
2538         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2539         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2540
2541         local_nb = tbc->local;
2542
2543         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2544                           &lockh, LCK_PW);
2545         if (rc != 0)
2546                 GOTO(out, rc);
2547
2548         /*
2549          * If getting the lock took more time than
2550          * client was willing to wait, drop it. b=11330
2551          */
2552         if (ktime_get_real_seconds() > req->rq_deadline ||
2553             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2554                 no_reply = true;
2555                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2556                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2557                        POSTID(&ioo->ioo_oid),
2558                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2559                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2560                 GOTO(out_lock, rc = -ETIMEDOUT);
2561         }
2562
2563         /* Because we already sync grant info with client when reconnect,
2564          * grant info will be cleared for resent req, then fed_grant and
2565          * total_grant will not be modified in following preprw_write */
2566         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2567                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2568                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2569         }
2570
2571         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2572         if (repbody == NULL)
2573                 GOTO(out_lock, rc = -ENOMEM);
2574         repbody->oa = body->oa;
2575
2576         npages = PTLRPC_MAX_BRW_PAGES;
2577         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2578                         objcount, ioo, remote_nb, &npages, local_nb);
2579         if (rc < 0)
2580                 GOTO(out_lock, rc);
2581         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2582                 int short_io_size;
2583                 unsigned char *short_io_buf;
2584
2585                 short_io_size = req_capsule_get_size(&req->rq_pill,
2586                                                      &RMF_SHORT_IO,
2587                                                      RCL_CLIENT);
2588                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2589                                                       &RMF_SHORT_IO);
2590                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2591                                " size = %d\n", short_io_size);
2592
2593                 /* Copy short io buf to pages */
2594                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2595                                        short_io_size);
2596                 desc = NULL;
2597         } else {
2598                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2599                                             PTLRPC_BULK_GET_SINK |
2600                                             PTLRPC_BULK_BUF_KIOV,
2601                                             OST_BULK_PORTAL,
2602                                             &ptlrpc_bulk_kiov_nopin_ops);
2603                 if (desc == NULL)
2604                         GOTO(skip_transfer, rc = -ENOMEM);
2605
2606                 /* NB Having prepped, we must commit... */
2607                 for (i = 0; i < npages; i++)
2608                         desc->bd_frag_ops->add_kiov_frag(desc,
2609                                         local_nb[i].lnb_page,
2610                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2611                                         local_nb[i].lnb_len);
2612
2613                 rc = sptlrpc_svc_prep_bulk(req, desc);
2614                 if (rc != 0)
2615                         GOTO(skip_transfer, rc);
2616
2617                 rc = target_bulk_io(exp, desc, &lwi);
2618         }
2619
2620         no_reply = rc != 0;
2621
2622 skip_transfer:
2623         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2624                 static int cksum_counter;
2625
2626                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2627                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2628
2629                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2630                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2631                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2632                                                            cksum_type);
2633
2634                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2635                                             local_nb, npages, OST_WRITE,
2636                                             &repbody->oa.o_cksum);
2637                 if (rc < 0)
2638                         GOTO(out_commitrw, rc);
2639
2640                 cksum_counter++;
2641
2642                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2643                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2644                                 body->oa.o_flags & OBD_FL_MMAP);
2645
2646                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2647                                           body->oa.o_cksum,
2648                                           repbody->oa.o_cksum, mmap);
2649                         cksum_counter = 0;
2650                 } else if ((cksum_counter & (-cksum_counter)) ==
2651                            cksum_counter) {
2652                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2653                                cksum_counter, libcfs_id2str(req->rq_peer),
2654                                repbody->oa.o_cksum);
2655                 }
2656         }
2657
2658 out_commitrw:
2659         /* Must commit after prep above in all cases */
2660         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2661                           objcount, ioo, remote_nb, npages, local_nb, rc);
2662         if (rc == -ENOTCONN)
2663                 /* quota acquire process has been given up because
2664                  * either the client has been evicted or the client
2665                  * has timed out the request already */
2666                 no_reply = true;
2667
2668         for (i = 0; i < niocount; i++) {
2669                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2670                         wait_sync = true;
2671                         break;
2672                 }
2673         }
2674         /*
2675          * Disable sending mtime back to the client. If the client locked the
2676          * whole object, then it has already updated the mtime on its side,
2677          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2678          */
2679         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2680
2681         if (rc == 0) {
2682                 int nob = 0;
2683
2684                 /* set per-requested niobuf return codes */
2685                 for (i = j = 0; i < niocount; i++) {
2686                         int len = remote_nb[i].rnb_len;
2687
2688                         nob += len;
2689                         rcs[i] = 0;
2690                         do {
2691                                 LASSERT(j < npages);
2692                                 if (local_nb[j].lnb_rc < 0)
2693                                         rcs[i] = local_nb[j].lnb_rc;
2694                                 len -= local_nb[j].lnb_len;
2695                                 j++;
2696                         } while (len > 0);
2697                         LASSERT(len == 0);
2698                 }
2699                 LASSERT(j == npages);
2700                 ptlrpc_lprocfs_brw(req, nob);
2701         }
2702 out_lock:
2703         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2704         if (desc)
2705                 ptlrpc_free_bulk(desc);
2706 out:
2707         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2708                 req->rq_no_reply = 1;
2709                 /* reply out callback would free */
2710                 ptlrpc_req_drop_rs(req);
2711                 if (!exp->exp_obd->obd_no_transno)
2712                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2713                                       " client will retry: rc = %d\n",
2714                                       obd_name,
2715                                       obd_uuid2str(&exp->exp_client_uuid),
2716                                       obd_export_nid2str(exp), rc);
2717         }
2718         memory_pressure_clr();
2719         RETURN(rc);
2720 }
2721 EXPORT_SYMBOL(tgt_brw_write);
2722
2723 /* Check if request can be reconstructed from saved reply data
2724  * A copy of the reply data is returned in @trd if the pointer is not NULL
2725  */
2726 bool req_can_reconstruct(struct ptlrpc_request *req,
2727                          struct tg_reply_data *trd)
2728 {
2729         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2730         struct lsd_client_data *lcd = ted->ted_lcd;
2731         bool found;
2732
2733         if (tgt_is_multimodrpcs_client(req->rq_export))
2734                 return tgt_lookup_reply(req, trd);
2735
2736         mutex_lock(&ted->ted_lcd_lock);
2737         found = req->rq_xid == lcd->lcd_last_xid ||
2738                 req->rq_xid == lcd->lcd_last_close_xid;
2739
2740         if (found && trd != NULL) {
2741                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2742                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2743                         trd->trd_reply.lrd_transno =
2744                                                 lcd->lcd_last_close_transno;
2745                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2746                 } else {
2747                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2748                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2749                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2750                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2751                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2752                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2753                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2754                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2755                 }
2756         }
2757         mutex_unlock(&ted->ted_lcd_lock);
2758
2759         return found;
2760 }
2761 EXPORT_SYMBOL(req_can_reconstruct);
2762