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 HABEO_CORPUS 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 & HABEO_CORPUS) && !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 & HABEO_CORPUS) {
 281                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
 282                                "but OID/FID is mandatory with HABEO_CORPUS\n",
 283                                tgt_name(tsi->tsi_tgt));
 284                         RETURN(-EPROTO);
 285                 } else {
 286                         RETURN(0);
 287                 }
 288         }
 289
 290         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 291
 292         /*
 293          * OST doesn't get object in advance for further use to prevent
 294          * situations with nested object_find which is potential deadlock.
 295          */
 296         tsi->tsi_corpus = NULL;
 297         RETURN(rc);
 298 }
 299
 300 /*
 301  * Do necessary preprocessing according to handler ->th_flags.
 302  */
 303 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 304                                   struct tgt_handler *h,
 305                                   struct ptlrpc_request *req)
 306 {
 307         struct req_capsule      *pill = tsi->tsi_pill;
 308         __u32                    flags = h->th_flags;
 309         int                      rc = 0;
 310
 311         ENTRY;
 312
 313         if (tsi->tsi_preprocessed)
 314                 RETURN(0);
 315
 316         LASSERT(h->th_act != NULL);
 317         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 318         LASSERT(current->journal_info == NULL);
 319
 320         LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
 321                      h->th_fmt != NULL));
 322         if (h->th_fmt != NULL) {
 323                 req_capsule_set(pill, h->th_fmt);
 324                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
 325                         rc = tgt_mdt_body_unpack(tsi, flags);
 326                         if (rc < 0)
 327                                 RETURN(rc);
 328                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 329                                                  RCL_CLIENT)) {
 330                         rc = tgt_ost_body_unpack(tsi, flags);
 331                         if (rc < 0)
 332                                 RETURN(rc);
 333                 }
 334         }
 335
 336         if (flags & MUTABOR && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 337                 RETURN(-EROFS);
 338
 339         if (flags & HABEO_CLAVIS) {
 340                 struct ldlm_request *dlm_req;
 341
 342                 LASSERT(h->th_fmt != NULL);
 343
 344                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 345                 if (dlm_req != NULL) {
 346                         union ldlm_wire_policy_data *policy =
 347                                         &dlm_req->lock_desc.l_policy_data;
 348
 349                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 350                                      LDLM_IBITS &&
 351                                      (policy->l_inodebits.bits |
 352                                       policy->l_inodebits.try_bits) == 0)) {
 353                                 /*
 354                                  * Lock without inodebits makes no sense and
 355                                  * will oops later in ldlm. If client miss to
 356                                  * set such bits, do not trigger ASSERTION.
 357                                  *
 358                                  * For liblustre flock case, it maybe zero.
 359                                  */
 360                                 rc = -EPROTO;
 361                         } else {
 362                                 tsi->tsi_dlm_req = dlm_req;
 363                         }
 364                 } else {
 365                         rc = -EFAULT;
 366                 }
 367         }
 368         tsi->tsi_preprocessed = 1;
 369         RETURN(rc);
 370 }
 371
 372 /*
 373  * Invoke handler for this request opc. Also do necessary preprocessing
 374  * (according to handler ->th_flags), and post-processing (setting of
 375  * ->last_{xid,committed}).
 376  */
 377 static int tgt_handle_request0(struct tgt_session_info *tsi,
 378                                struct tgt_handler *h,
 379                                struct ptlrpc_request *req)
 380 {
 381         int      serious = 0;
 382         int      rc;
 383         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 384
 385         ENTRY;
 386
 387
 388         /* When dealing with sec context requests, no export is associated yet,
 389          * because these requests are sent before *_CONNECT requests.
 390          * A NULL req->rq_export means the normal *_common_slice handlers will
 391          * not be called, because there is no reference to the target.
 392          * So deal with them by hand and jump directly to target_send_reply().
 393          */
 394         switch (opc) {
 395         case SEC_CTX_INIT:
 396         case SEC_CTX_INIT_CONT:
 397         case SEC_CTX_FINI:
 398                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 399                 GOTO(out, rc = 0);
 400         }
 401
 402         /*
 403          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 404          * to put same checks into handlers like mdt_close(), mdt_reint(),
 405          * etc., without talking to mdt authors first. Checking same thing
 406          * there again is useless and returning 0 error without packing reply
 407          * is buggy! Handlers either pack reply or return error.
 408          *
 409          * We return 0 here and do not send any reply in order to emulate
 410          * network failure. Do not send any reply in case any of NET related
 411          * fail_id has occured.
 412          */
 413         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 414                 RETURN(0);
 415         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 416                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 417                 RETURN(0);
 418
 419         rc = tgt_request_preprocess(tsi, h, req);
 420         /* pack reply if reply format is fixed */
 421         if (rc == 0 && h->th_flags & HABEO_REFERO) {
 422                 /* Pack reply */
 423                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 424                                           RCL_SERVER))
 425                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 426                                              RCL_SERVER,
 427                                              tsi->tsi_mdt_body->mbo_eadatasize);
 428                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 429                                           RCL_SERVER))
 430                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 431                                              RCL_SERVER, 0);
 432                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 433                         req_capsule_set_size(tsi->tsi_pill,
 434                                              &RMF_ACL, RCL_SERVER,
 435                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 436
 437                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 438                                           RCL_SERVER)) {
 439                         struct niobuf_remote *remote_nb =
 440                                 req_capsule_client_get(tsi->tsi_pill,
 441                                                        &RMF_NIOBUF_REMOTE);
 442                         struct ost_body *body = tsi->tsi_ost_body;
 443
 444                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 445                                          RCL_SERVER,
 446                                          (body->oa.o_flags & OBD_FL_SHORT_IO) ?
 447                                          remote_nb[0].rnb_len : 0);
 448                 }
 449
 450                 rc = req_capsule_server_pack(tsi->tsi_pill);
 451         }
 452
 453         if (likely(rc == 0)) {
 454                 /*
 455                  * Process request, there can be two types of rc:
 456                  * 1) errors with msg unpack/pack, other failures outside the
 457                  * operation itself. This is counted as serious errors;
 458                  * 2) errors during fs operation, should be placed in rq_status
 459                  * only
 460                  */
 461                 rc = h->th_act(tsi);
 462                 if (!is_serious(rc) &&
 463                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 464                         DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
 465                                   "pack reply and returned 0 error\n",
 466                                   tgt_name(tsi->tsi_tgt), h->th_name);
 467                         LBUG();
 468                 }
 469                 serious = is_serious(rc);
 470                 rc = clear_serious(rc);
 471         } else {
 472                 serious = 1;
 473         }
 474
 475         req->rq_status = rc;
 476
 477         /*
 478          * ELDLM_* codes which > 0 should be in rq_status only as well as
 479          * all non-serious errors.
 480          */
 481         if (rc > 0 || !serious)
 482                 rc = 0;
 483
 484         LASSERT(current->journal_info == NULL);
 485
 486         if (likely(rc == 0 && req->rq_export))
 487                 target_committed_to_req(req);
 488
 489 out:
 490         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 491         RETURN(0);
 492 }
 493
 494 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 495                                        struct obd_device *obd, int *process)
 496 {
 497         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 498         case MDS_DISCONNECT:
 499         case OST_DISCONNECT:
 500         case OBD_IDX_READ:
 501                 *process = 1;
 502                 RETURN(0);
 503         case MDS_CLOSE:
 504         case MDS_SYNC: /* used in unmounting */
 505         case OBD_PING:
 506         case MDS_REINT:
 507         case OUT_UPDATE:
 508         case SEQ_QUERY:
 509         case FLD_QUERY:
 510         case FLD_READ:
 511         case LDLM_ENQUEUE:
 512         case OST_CREATE:
 513         case OST_DESTROY:
 514         case OST_PUNCH:
 515         case OST_SETATTR:
 516         case OST_SYNC:
 517         case OST_WRITE:
 518         case MDS_HSM_PROGRESS:
 519         case MDS_HSM_STATE_SET:
 520         case MDS_HSM_REQUEST:
 521                 *process = target_queue_recovery_request(req, obd);
 522                 RETURN(0);
 523
 524         default:
 525                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 526                 *process = -EAGAIN;
 527                 RETURN(0);
 528         }
 529 }
 530
 531 /*
 532  * Handle recovery. Return:
 533  *        +1: continue request processing;
 534  *       -ve: abort immediately with the given error code;
 535  *         0: send reply with error code in req->rq_status;
 536  */
 537 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 538 {
 539         ENTRY;
 540
 541         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 542         case MDS_CONNECT:
 543         case OST_CONNECT:
 544         case MGS_CONNECT:
 545         case SEC_CTX_INIT:
 546         case SEC_CTX_INIT_CONT:
 547         case SEC_CTX_FINI:
 548                 RETURN(+1);
 549         }
 550
 551         if (!req->rq_export->exp_obd->obd_replayable)
 552                 RETURN(+1);
 553
 554         /* sanity check: if the xid matches, the request must be marked as a
 555          * resent or replayed */
 556         if (req_can_reconstruct(req, NULL)) {
 557                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 558                       (MSG_RESENT | MSG_REPLAY))) {
 559                         DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches "
 560                                   "saved xid, expected REPLAY or RESENT flag "
 561                                   "(%x)", req->rq_xid,
 562                                   lustre_msg_get_flags(req->rq_reqmsg));
 563                         req->rq_status = -ENOTCONN;
 564                         RETURN(-ENOTCONN);
 565                 }
 566         }
 567         /* else: note the opposite is not always true; a RESENT req after a
 568          * failover will usually not match the last_xid, since it was likely
 569          * never committed. A REPLAYed request will almost never match the
 570          * last xid, however it could for a committed, but still retained,
 571          * open. */
 572
 573         /* Check for aborted recovery... */
 574         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 575                 int rc;
 576                 int should_process;
 577
 578                 DEBUG_REQ(D_INFO, req, "Got new replay");
 579                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 580                                                  &should_process);
 581                 if (rc != 0 || !should_process)
 582                         RETURN(rc);
 583                 else if (should_process < 0) {
 584                         req->rq_status = should_process;
 585                         rc = ptlrpc_error(req);
 586                         RETURN(rc);
 587                 }
 588         }
 589         RETURN(+1);
 590 }
 591
 592 /* Initial check for request, it is validation mostly */
 593 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 594 {
 595         struct tgt_handler      *h;
 596         struct tgt_opc_slice    *s;
 597         struct lu_target        *tgt;
 598         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 599
 600         ENTRY;
 601
 602         tgt = class_exp2tgt(req->rq_export);
 603         if (unlikely(tgt == NULL)) {
 604                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
 605                           class_exp2obd(req->rq_export)->obd_name);
 606                 RETURN(ERR_PTR(-EINVAL));
 607         }
 608
 609         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 610                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 611                         break;
 612
 613         /* opcode was not found in slice */
 614         if (unlikely(s->tos_hs == NULL)) {
 615                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 616                        opc);
 617                 RETURN(ERR_PTR(-ENOTSUPP));
 618         }
 619
 620         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 621         h = s->tos_hs + (opc - s->tos_opc_start);
 622         if (unlikely(h->th_opc == 0)) {
 623                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 624                 RETURN(ERR_PTR(-ENOTSUPP));
 625         }
 626
 627         RETURN(h);
 628 }
 629
 630 static int process_req_last_xid(struct ptlrpc_request *req)
 631 {
 632         __u64   last_xid;
 633         ENTRY;
 634
 635         /* check request's xid is consistent with export's last_xid */
 636         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 637         if (last_xid > req->rq_export->exp_last_xid)
 638                 req->rq_export->exp_last_xid = last_xid;
 639
 640         if (req->rq_xid == 0 ||
 641             (req->rq_xid <= req->rq_export->exp_last_xid)) {
 642                 DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
 643                           "last_xid %llx\n", req->rq_xid,
 644                           req->rq_export->exp_last_xid);
 645                 /* Some request is allowed to be sent during replay,
 646                  * such as OUT update requests, FLD requests, so it
 647                  * is possible that replay requests has smaller XID
 648                  * than the exp_last_xid.
 649                  *
 650                  * Some non-replay requests may have smaller XID as
 651                  * well:
 652                  *
 653                  * - Client send a no_resend RPC, like statfs;
 654                  * - The RPC timedout (or some other error) on client,
 655                  *   then it's removed from the unreplied list;
 656                  * - Client send some other request to bump the
 657                  *   exp_last_xid on server;
 658                  * - The former RPC got chance to be processed;
 659                  */
 660                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 661                         RETURN(-EPROTO);
 662         }
 663
 664         /* The "last_xid" is the minimum xid among unreplied requests,
 665          * if the request is from the previous connection, its xid can
 666          * still be larger than "exp_last_xid", then the above check of
 667          * xid is not enough to determine whether the request is delayed.
 668          *
 669          * For example, if some replay request was delayed and caused
 670          * timeout at client and the replay is restarted, the delayed
 671          * replay request will have the larger xid than "exp_last_xid"
 672          */
 673         if (req->rq_export->exp_conn_cnt >
 674             lustre_msg_get_conn_cnt(req->rq_reqmsg))
 675                 RETURN(-ESTALE);
 676
 677         /* try to release in-memory reply data */
 678         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 679                 tgt_handle_received_xid(req->rq_export,
 680                                 lustre_msg_get_last_xid(req->rq_reqmsg));
 681                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 682                       (MSG_RESENT | MSG_REPLAY)))
 683                         tgt_handle_tag(req->rq_export,
 684                                        lustre_msg_get_tag(req->rq_reqmsg));
 685         }
 686         RETURN(0);
 687 }
 688
 689 int tgt_request_handle(struct ptlrpc_request *req)
 690 {
 691         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 692
 693         struct lustre_msg       *msg = req->rq_reqmsg;
 694         struct tgt_handler      *h;
 695         struct lu_target        *tgt;
 696         int                      request_fail_id = 0;
 697         __u32                    opc = lustre_msg_get_opc(msg);
 698         struct obd_device       *obd;
 699         int                      rc;
 700         bool                     is_connect = false;
 701         ENTRY;
 702
 703         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 704                 if (cfs_fail_val == 0 &&
 705                     lustre_msg_get_opc(msg) != OBD_PING &&
 706                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 707                         struct l_wait_info lwi =  { 0 };
 708
 709                         cfs_fail_val = 1;
 710                         cfs_race_state = 0;
 711                         l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
 712                                      &lwi);
 713                 }
 714         }
 715
 716         /* Refill the context, to make sure all thread keys are allocated */
 717         lu_env_refill(req->rq_svc_thread->t_env);
 718
 719         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 720         tsi->tsi_pill = &req->rq_pill;
 721         tsi->tsi_env = req->rq_svc_thread->t_env;
 722
 723         /* if request has export then get handlers slice from corresponding
 724          * target, otherwise that should be connect operation */
 725         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 726             opc == MGS_CONNECT) {
 727                 is_connect = true;
 728                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 729                 rc = target_handle_connect(req);
 730                 if (rc != 0) {
 731                         rc = ptlrpc_error(req);
 732                         GOTO(out, rc);
 733                 }
 734                 /* recovery-small test 18c asks to drop connect reply */
 735                 if (unlikely(opc == OST_CONNECT &&
 736                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 737                         GOTO(out, rc = 0);
 738         }
 739
 740         if (unlikely(!class_connected_export(req->rq_export))) {
 741                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 742                     opc == SEC_CTX_FINI) {
 743                         /* sec context initialization has to be handled
 744                          * by hand in tgt_handle_request0() */
 745                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 746                         h = NULL;
 747                         GOTO(handle_recov, rc = 0);
 748                 }
 749                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 750                        opc, libcfs_id2str(req->rq_peer));
 751                 req->rq_status = -ENOTCONN;
 752                 rc = ptlrpc_error(req);
 753                 GOTO(out, rc);
 754         }
 755
 756         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 757         tsi->tsi_exp = req->rq_export;
 758         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 759                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 760         else
 761                 tsi->tsi_jobid = NULL;
 762
 763         if (tgt == NULL) {
 764                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
 765                           class_exp2obd(req->rq_export)->obd_name);
 766                 req->rq_status = -EINVAL;
 767                 rc = ptlrpc_error(req);
 768                 GOTO(out, rc);
 769         }
 770
 771         /* Skip last_xid processing for the recovery thread, otherwise, the
 772          * last_xid on same request could be processed twice: first time when
 773          * processing the incoming request, second time when the request is
 774          * being processed by recovery thread. */
 775         obd = class_exp2obd(req->rq_export);
 776         if (is_connect) {
 777                 /* reset the exp_last_xid on each connection. */
 778                 req->rq_export->exp_last_xid = 0;
 779         } else if (obd->obd_recovery_data.trd_processing_task !=
 780                    current_pid()) {
 781                 rc = process_req_last_xid(req);
 782                 if (rc) {
 783                         req->rq_status = rc;
 784                         rc = ptlrpc_error(req);
 785                         GOTO(out, rc);
 786                 }
 787         }
 788
 789         request_fail_id = tgt->lut_request_fail_id;
 790         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 791
 792         h = tgt_handler_find_check(req);
 793         if (IS_ERR(h)) {
 794                 req->rq_status = PTR_ERR(h);
 795                 rc = ptlrpc_error(req);
 796                 GOTO(out, rc);
 797         }
 798
 799         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 800                  h->th_opc, opc);
 801
 802         if (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         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
 995         spin_unlock(&tsi->tsi_exp->exp_lock);
 996
 997         RETURN(0);
 998 out:
 999         obd_disconnect(class_export_get(tsi->tsi_exp));
1000         return rc;
1001 }
1002 EXPORT_SYMBOL(tgt_connect);
1003
1004 int tgt_disconnect(struct tgt_session_info *tsi)
1005 {
1006         int rc;
1007
1008         ENTRY;
1009
1010         rc = target_handle_disconnect(tgt_ses_req(tsi));
1011         if (rc)
1012                 RETURN(err_serious(rc));
1013
1014         RETURN(rc);
1015 }
1016 EXPORT_SYMBOL(tgt_disconnect);
1017
1018 /*
1019  * Unified target OBD handlers
1020  */
1021 int tgt_obd_ping(struct tgt_session_info *tsi)
1022 {
1023         int rc;
1024
1025         ENTRY;
1026
1027         rc = target_handle_ping(tgt_ses_req(tsi));
1028         if (rc)
1029                 RETURN(err_serious(rc));
1030
1031         RETURN(rc);
1032 }
1033 EXPORT_SYMBOL(tgt_obd_ping);
1034
1035 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1036 {
1037         return err_serious(-EOPNOTSUPP);
1038 }
1039
1040 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1041 {
1042         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1043         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1044         struct obd_export       *exp = req->rq_export;
1045         struct ptlrpc_bulk_desc *desc;
1046         struct l_wait_info      *lwi = &tti->tti_u.update.tti_wait_info;
1047         int                      i;
1048         int                      rc;
1049
1050         ENTRY;
1051
1052         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1053                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1054                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1055         if (desc == NULL)
1056                 RETURN(-ENOMEM);
1057
1058         for (i = 0; i < rdbuf->rb_nbufs; i++)
1059                 desc->bd_frag_ops->add_iov_frag(desc,
1060                                         rdbuf->rb_bufs[i].lb_buf,
1061                                         rdbuf->rb_bufs[i].lb_len);
1062
1063         rc = target_bulk_io(exp, desc, lwi);
1064         ptlrpc_free_bulk(desc);
1065         RETURN(rc);
1066 }
1067 EXPORT_SYMBOL(tgt_send_buffer);
1068
1069 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1070 {
1071         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1072         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1073         struct obd_export       *exp = req->rq_export;
1074         struct ptlrpc_bulk_desc *desc;
1075         struct l_wait_info      *lwi = &tti->tti_u.rdpg.tti_wait_info;
1076         int                      tmpcount;
1077         int                      tmpsize;
1078         int                      i;
1079         int                      rc;
1080
1081         ENTRY;
1082
1083         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1084                                     PTLRPC_BULK_PUT_SOURCE |
1085                                         PTLRPC_BULK_BUF_KIOV,
1086                                     MDS_BULK_PORTAL,
1087                                     &ptlrpc_bulk_kiov_pin_ops);
1088         if (desc == NULL)
1089                 RETURN(-ENOMEM);
1090
1091         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1092                 /* old client requires reply size in it's PAGE_SIZE,
1093                  * which is rdpg->rp_count */
1094                 nob = rdpg->rp_count;
1095
1096         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1097              i++, tmpcount -= tmpsize) {
1098                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1099                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1100                                                  tmpsize);
1101         }
1102
1103         LASSERT(desc->bd_nob == nob);
1104         rc = target_bulk_io(exp, desc, lwi);
1105         ptlrpc_free_bulk(desc);
1106         RETURN(rc);
1107 }
1108 EXPORT_SYMBOL(tgt_sendpage);
1109
1110 /*
1111  * OBD_IDX_READ handler
1112  */
1113 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1114 {
1115         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1116         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1117         struct idx_info         *req_ii, *rep_ii;
1118         int                      rc, i;
1119
1120         ENTRY;
1121
1122         memset(rdpg, 0, sizeof(*rdpg));
1123         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1124
1125         /* extract idx_info buffer from request & reply */
1126         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1127         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1128                 RETURN(err_serious(-EPROTO));
1129
1130         rc = req_capsule_server_pack(tsi->tsi_pill);
1131         if (rc)
1132                 RETURN(err_serious(rc));
1133
1134         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1135         if (rep_ii == NULL)
1136                 RETURN(err_serious(-EFAULT));
1137         rep_ii->ii_magic = IDX_INFO_MAGIC;
1138
1139         /* extract hash to start with */
1140         rdpg->rp_hash = req_ii->ii_hash_start;
1141
1142         /* extract requested attributes */
1143         rdpg->rp_attrs = req_ii->ii_attrs;
1144
1145         /* check that fid packed in request is valid and supported */
1146         if (!fid_is_sane(&req_ii->ii_fid))
1147                 RETURN(-EINVAL);
1148         rep_ii->ii_fid = req_ii->ii_fid;
1149
1150         /* copy flags */
1151         rep_ii->ii_flags = req_ii->ii_flags;
1152
1153         /* compute number of pages to allocate, ii_count is the number of 4KB
1154          * containers */
1155         if (req_ii->ii_count <= 0)
1156                 GOTO(out, rc = -EFAULT);
1157         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1158                                exp_max_brw_size(tsi->tsi_exp));
1159         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1160
1161         /* allocate pages to store the containers */
1162         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1163         if (rdpg->rp_pages == NULL)
1164                 GOTO(out, rc = -ENOMEM);
1165         for (i = 0; i < rdpg->rp_npages; i++) {
1166                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1167                 if (rdpg->rp_pages[i] == NULL)
1168                         GOTO(out, rc = -ENOMEM);
1169         }
1170
1171         /* populate pages with key/record pairs */
1172         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1173         if (rc < 0)
1174                 GOTO(out, rc);
1175
1176         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1177                  "asked %d > %d\n", rc, rdpg->rp_count);
1178
1179         /* send pages to client */
1180         rc = tgt_sendpage(tsi, rdpg, rc);
1181         if (rc)
1182                 GOTO(out, rc);
1183         EXIT;
1184 out:
1185         if (rdpg->rp_pages) {
1186                 for (i = 0; i < rdpg->rp_npages; i++)
1187                         if (rdpg->rp_pages[i])
1188                                 __free_page(rdpg->rp_pages[i]);
1189                 OBD_FREE(rdpg->rp_pages,
1190                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1191         }
1192         return rc;
1193 }
1194
1195 struct tgt_handler tgt_obd_handlers[] = {
1196 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1197 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1198 };
1199 EXPORT_SYMBOL(tgt_obd_handlers);
1200
1201 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1202              struct dt_object *obj, __u64 start, __u64 end)
1203 {
1204         int rc = 0;
1205
1206         ENTRY;
1207
1208         /* if no objid is specified, it means "sync whole filesystem" */
1209         if (obj == NULL) {
1210                 rc = dt_sync(env, tgt->lut_bottom);
1211         } else if (dt_version_get(env, obj) >
1212                    tgt->lut_obd->obd_last_committed) {
1213                 rc = dt_object_sync(env, obj, start, end);
1214         }
1215         atomic_inc(&tgt->lut_sync_count);
1216
1217         RETURN(rc);
1218 }
1219 EXPORT_SYMBOL(tgt_sync);
1220 /*
1221  * Unified target DLM handlers.
1222  */
1223
1224 /**
1225  * Unified target BAST
1226  *
1227  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1228  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1229  * but if it's cross-MDT lock, because remote object version is not set, a
1230  * filesystem sync is needed.
1231  *
1232  * \param lock server side lock
1233  * \param desc lock desc
1234  * \param data ldlm_cb_set_arg
1235  * \param flag  indicates whether this cancelling or blocking callback
1236  * \retval      0 on success
1237  * \retval      negative number on error
1238  */
1239 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1240                             void *data, int flag)
1241 {
1242         struct lu_env            env;
1243         struct lu_target        *tgt;
1244         struct dt_object        *obj = NULL;
1245         struct lu_fid            fid;
1246         int                      rc = 0;
1247
1248         ENTRY;
1249
1250         tgt = class_exp2tgt(lock->l_export);
1251
1252         if (unlikely(tgt == NULL)) {
1253                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1254                        class_exp2obd(lock->l_export)->obd_name);
1255                 RETURN(-EINVAL);
1256         }
1257
1258         if (flag == LDLM_CB_CANCELING &&
1259             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1260             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1261              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1262               ldlm_is_cbpending(lock))) &&
1263             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1264              lock->l_resource->lr_type == LDLM_EXTENT)) {
1265                 __u64 start = 0;
1266                 __u64 end = OBD_OBJECT_EOF;
1267
1268                 rc = lu_env_init(&env, LCT_DT_THREAD);
1269                 if (unlikely(rc != 0))
1270                         RETURN(rc);
1271
1272                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1273                                    tgt->lut_lsd.lsd_osd_index);
1274
1275                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1276                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1277                         if (IS_ERR(obj))
1278                                 GOTO(err_env, rc = PTR_ERR(obj));
1279
1280                         if (!dt_object_exists(obj))
1281                                 GOTO(err_put, rc = -ENOENT);
1282
1283                         start = lock->l_policy_data.l_extent.start;
1284                         end = lock->l_policy_data.l_extent.end;
1285                 }
1286
1287                 rc = tgt_sync(&env, tgt, obj, start, end);
1288                 if (rc < 0) {
1289                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1290                                "cancel: rc = %d\n",
1291                                tgt_name(tgt), PFID(&fid),
1292                                lock->l_policy_data.l_extent.start,
1293                                lock->l_policy_data.l_extent.end, rc);
1294                 }
1295 err_put:
1296                 if (obj != NULL)
1297                         dt_object_put(&env, obj);
1298 err_env:
1299                 lu_env_fini(&env);
1300         }
1301
1302         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1303         RETURN(rc);
1304 }
1305
1306 static struct ldlm_callback_suite tgt_dlm_cbs = {
1307         .lcs_completion = ldlm_server_completion_ast,
1308         .lcs_blocking   = tgt_blocking_ast,
1309         .lcs_glimpse    = ldlm_server_glimpse_ast
1310 };
1311
1312 int tgt_enqueue(struct tgt_session_info *tsi)
1313 {
1314         struct ptlrpc_request *req = tgt_ses_req(tsi);
1315         int rc;
1316
1317         ENTRY;
1318         /*
1319          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1320          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1321          */
1322         LASSERT(tsi->tsi_dlm_req != NULL);
1323         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1324                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1325         if (rc)
1326                 RETURN(err_serious(rc));
1327
1328         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1329         case LUT_FAIL_MDT:
1330                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1331                 break;
1332         case LUT_FAIL_OST:
1333                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1334                 break;
1335         case LUT_FAIL_MGT:
1336                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1337                 break;
1338         default:
1339                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1340                 break;
1341         }
1342         RETURN(req->rq_status);
1343 }
1344 EXPORT_SYMBOL(tgt_enqueue);
1345
1346 int tgt_convert(struct tgt_session_info *tsi)
1347 {
1348         struct ptlrpc_request *req = tgt_ses_req(tsi);
1349         int rc;
1350
1351         ENTRY;
1352         LASSERT(tsi->tsi_dlm_req);
1353         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1354         if (rc)
1355                 RETURN(err_serious(rc));
1356
1357         RETURN(req->rq_status);
1358 }
1359
1360 int tgt_bl_callback(struct tgt_session_info *tsi)
1361 {
1362         return err_serious(-EOPNOTSUPP);
1363 }
1364
1365 int tgt_cp_callback(struct tgt_session_info *tsi)
1366 {
1367         return err_serious(-EOPNOTSUPP);
1368 }
1369
1370 /* generic LDLM target handler */
1371 struct tgt_handler tgt_dlm_handlers[] = {
1372 TGT_DLM_HDL    (HABEO_CLAVIS,   LDLM_ENQUEUE,           tgt_enqueue),
1373 TGT_DLM_HDL    (HABEO_CLAVIS,   LDLM_CONVERT,           tgt_convert),
1374 TGT_DLM_HDL_VAR(0,              LDLM_BL_CALLBACK,       tgt_bl_callback),
1375 TGT_DLM_HDL_VAR(0,              LDLM_CP_CALLBACK,       tgt_cp_callback)
1376 };
1377 EXPORT_SYMBOL(tgt_dlm_handlers);
1378
1379 /*
1380  * Unified target LLOG handlers.
1381  */
1382 int tgt_llog_open(struct tgt_session_info *tsi)
1383 {
1384         int rc;
1385
1386         ENTRY;
1387
1388         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1389
1390         RETURN(rc);
1391 }
1392 EXPORT_SYMBOL(tgt_llog_open);
1393
1394 int tgt_llog_read_header(struct tgt_session_info *tsi)
1395 {
1396         int rc;
1397
1398         ENTRY;
1399
1400         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1401
1402         RETURN(rc);
1403 }
1404 EXPORT_SYMBOL(tgt_llog_read_header);
1405
1406 int tgt_llog_next_block(struct tgt_session_info *tsi)
1407 {
1408         int rc;
1409
1410         ENTRY;
1411
1412         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1413
1414         RETURN(rc);
1415 }
1416 EXPORT_SYMBOL(tgt_llog_next_block);
1417
1418 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1419 {
1420         int rc;
1421
1422         ENTRY;
1423
1424         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1425
1426         RETURN(rc);
1427 }
1428 EXPORT_SYMBOL(tgt_llog_prev_block);
1429
1430 /* generic llog target handler */
1431 struct tgt_handler tgt_llog_handlers[] = {
1432 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1433 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1434 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1435 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1436 };
1437 EXPORT_SYMBOL(tgt_llog_handlers);
1438
1439 /*
1440  * sec context handlers
1441  */
1442 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1443 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1444 {
1445         return 0;
1446 }
1447
1448 struct tgt_handler tgt_sec_ctx_handlers[] = {
1449 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1450 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1451 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1452 };
1453 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1454
1455 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1456                                  struct dt_device *key,
1457                                  struct lfsck_req_local *lrl,
1458                                  struct thandle *th) = NULL;
1459
1460 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1461                                                       struct dt_device *,
1462                                                       struct lfsck_req_local *,
1463                                                       struct thandle *))
1464 {
1465         tgt_lfsck_in_notify_local = notify;
1466 }
1467 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1468
1469 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1470                            struct dt_device *key,
1471                            struct lfsck_request *lr) = NULL;
1472
1473 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1474                                                 struct dt_device *,
1475                                                 struct lfsck_request *))
1476 {
1477         tgt_lfsck_in_notify = notify;
1478 }
1479 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1480
1481 static int (*tgt_lfsck_query)(const struct lu_env *env,
1482                               struct dt_device *key,
1483                               struct lfsck_request *req,
1484                               struct lfsck_reply *rep,
1485                               struct lfsck_query *que) = NULL;
1486
1487 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1488                                            struct dt_device *,
1489                                            struct lfsck_request *,
1490                                            struct lfsck_reply *,
1491                                            struct lfsck_query *))
1492 {
1493         tgt_lfsck_query = query;
1494 }
1495 EXPORT_SYMBOL(tgt_register_lfsck_query);
1496
1497 /* LFSCK request handlers */
1498 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1499 {
1500         const struct lu_env     *env = tsi->tsi_env;
1501         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1502         struct lfsck_request    *lr;
1503         int                      rc;
1504         ENTRY;
1505
1506         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1507         if (lr == NULL)
1508                 RETURN(-EPROTO);
1509
1510         rc = tgt_lfsck_in_notify(env, key, lr);
1511
1512         RETURN(rc);
1513 }
1514
1515 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1516 {
1517         struct lfsck_request    *request;
1518         struct lfsck_reply      *reply;
1519         int                      rc;
1520         ENTRY;
1521
1522         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1523         if (request == NULL)
1524                 RETURN(-EPROTO);
1525
1526         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1527         if (reply == NULL)
1528                 RETURN(-ENOMEM);
1529
1530         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1531                              request, reply, NULL);
1532
1533         RETURN(rc < 0 ? rc : 0);
1534 }
1535
1536 struct tgt_handler tgt_lfsck_handlers[] = {
1537 TGT_LFSCK_HDL(HABEO_REFERO,     LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1538 TGT_LFSCK_HDL(HABEO_REFERO,     LFSCK_QUERY,    tgt_handle_lfsck_query),
1539 };
1540 EXPORT_SYMBOL(tgt_lfsck_handlers);
1541
1542 /*
1543  * initialize per-thread page pool (bug 5137).
1544  */
1545 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1546 {
1547         struct tgt_thread_big_cache *tbc;
1548
1549         ENTRY;
1550
1551         LASSERT(thread != NULL);
1552         LASSERT(thread->t_data == NULL);
1553
1554         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1555         if (tbc == NULL)
1556                 RETURN(-ENOMEM);
1557         thread->t_data = tbc;
1558         RETURN(0);
1559 }
1560 EXPORT_SYMBOL(tgt_io_thread_init);
1561
1562 /*
1563  * free per-thread pool created by tgt_thread_init().
1564  */
1565 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1566 {
1567         struct tgt_thread_big_cache *tbc;
1568
1569         ENTRY;
1570
1571         LASSERT(thread != NULL);
1572
1573         /*
1574          * be prepared to handle partially-initialized pools (because this is
1575          * called from ost_io_thread_init() for cleanup.
1576          */
1577         tbc = thread->t_data;
1578         if (tbc != NULL) {
1579                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1580                 thread->t_data = NULL;
1581         }
1582         EXIT;
1583 }
1584 EXPORT_SYMBOL(tgt_io_thread_done);
1585
1586 /**
1587  * Helper function for getting Data-on-MDT file server DLM lock
1588  * if asked by client.
1589  */
1590 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1591                       struct lustre_handle *lh, int mode, __u64 *flags)
1592 {
1593         union ldlm_policy_data policy = {
1594                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1595         };
1596         int rc;
1597
1598         ENTRY;
1599
1600         LASSERT(lh != NULL);
1601         LASSERT(ns != NULL);
1602         LASSERT(!lustre_handle_is_used(lh));
1603
1604         rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, &policy, mode,
1605                                     flags, ldlm_blocking_ast,
1606                                     ldlm_completion_ast, ldlm_glimpse_ast,
1607                                     NULL, 0, LVB_T_NONE, NULL, lh);
1608
1609         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1610 }
1611 EXPORT_SYMBOL(tgt_mdt_data_lock);
1612
1613 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1614 {
1615         LASSERT(lustre_handle_is_used(lh));
1616         ldlm_lock_decref(lh, mode);
1617 }
1618 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1619
1620 /**
1621  * Helper function for getting server side [start, start+count] DLM lock
1622  * if asked by client.
1623  */
1624 int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1625                     __u64 start, __u64 end, struct lustre_handle *lh,
1626                     int mode, __u64 *flags)
1627 {
1628         union ldlm_policy_data policy;
1629         int rc;
1630
1631         ENTRY;
1632
1633         LASSERT(lh != NULL);
1634         LASSERT(ns != NULL);
1635         LASSERT(!lustre_handle_is_used(lh));
1636
1637         policy.l_extent.gid = 0;
1638         policy.l_extent.start = start & PAGE_MASK;
1639
1640         /*
1641          * If ->o_blocks is EOF it means "lock till the end of the file".
1642          * Otherwise, it's size of an extent or hole being punched (in bytes).
1643          */
1644         if (end == OBD_OBJECT_EOF || end < start)
1645                 policy.l_extent.end = OBD_OBJECT_EOF;
1646         else
1647                 policy.l_extent.end = end | ~PAGE_MASK;
1648
1649         rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_EXTENT, &policy, mode,
1650                                     flags, ldlm_blocking_ast,
1651                                     ldlm_completion_ast, ldlm_glimpse_ast,
1652                                     NULL, 0, LVB_T_NONE, NULL, lh);
1653         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1654 }
1655 EXPORT_SYMBOL(tgt_extent_lock);
1656
1657 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1658 {
1659         LASSERT(lustre_handle_is_used(lh));
1660         ldlm_lock_decref(lh, mode);
1661 }
1662 EXPORT_SYMBOL(tgt_extent_unlock);
1663
1664 static int tgt_brw_lock(struct obd_export *exp, struct ldlm_res_id *res_id,
1665                         struct obd_ioobj *obj, struct niobuf_remote *nb,
1666                         struct lustre_handle *lh, enum ldlm_mode mode)
1667 {
1668         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1669         __u64                    flags = 0;
1670         int                      nrbufs = obj->ioo_bufcnt;
1671         int                      i;
1672         int                      rc;
1673
1674         ENTRY;
1675
1676         LASSERT(mode == LCK_PR || mode == LCK_PW);
1677         LASSERT(!lustre_handle_is_used(lh));
1678
1679         if (ns->ns_obd->obd_recovering)
1680                 RETURN(0);
1681
1682         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1683                 RETURN(0);
1684
1685         for (i = 1; i < nrbufs; i++)
1686                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1687                         RETURN(-EFAULT);
1688
1689         /* MDT IO for data-on-mdt */
1690         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1691                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1692         else
1693                 rc = tgt_extent_lock(ns, res_id, nb[0].rnb_offset,
1694                                      nb[nrbufs - 1].rnb_offset +
1695                                      nb[nrbufs - 1].rnb_len - 1,
1696                                      lh, mode, &flags);
1697         RETURN(rc);
1698 }
1699
1700 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1701                            struct lustre_handle *lh, enum ldlm_mode mode)
1702 {
1703         ENTRY;
1704
1705         LASSERT(mode == LCK_PR || mode == LCK_PW);
1706         LASSERT((obj->ioo_bufcnt > 0 &&
1707                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1708                 lustre_handle_is_used(lh));
1709
1710         if (lustre_handle_is_used(lh))
1711                 tgt_extent_unlock(lh, mode);
1712         EXIT;
1713 }
1714 static int tgt_checksum_niobuf(struct lu_target *tgt,
1715                                  struct niobuf_local *local_nb, int npages,
1716                                  int opc, enum cksum_types cksum_type,
1717                                  __u32 *cksum)
1718 {
1719         struct cfs_crypto_hash_desc     *hdesc;
1720         unsigned int                    bufsize;
1721         int                             i, err;
1722         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1723
1724         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1725         if (IS_ERR(hdesc)) {
1726                 CERROR("%s: unable to initialize checksum hash %s\n",
1727                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1728                 return PTR_ERR(hdesc);
1729         }
1730
1731         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1732         for (i = 0; i < npages; i++) {
1733                 /* corrupt the data before we compute the checksum, to
1734                  * simulate a client->OST data error */
1735                 if (i == 0 && opc == OST_WRITE &&
1736                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1737                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1738                         int len = local_nb[i].lnb_len;
1739                         struct page *np = tgt_page_to_corrupt;
1740
1741                         if (np) {
1742                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1743                                                         KM_USER0);
1744                                 char *ptr2 = page_address(np);
1745
1746                                 memcpy(ptr2 + off, ptr + off, len);
1747                                 memcpy(ptr2 + off, "bad3", min(4, len));
1748                                 ll_kunmap_atomic(ptr, KM_USER0);
1749
1750                                 /* LU-8376 to preserve original index for
1751                                  * display in dump_all_bulk_pages() */
1752                                 np->index = i;
1753
1754                                 cfs_crypto_hash_update_page(hdesc, np, off,
1755                                                             len);
1756                                 continue;
1757                         } else {
1758                                 CERROR("%s: can't alloc page for corruption\n",
1759                                        tgt_name(tgt));
1760                         }
1761                 }
1762                 cfs_crypto_hash_update_page(hdesc, local_nb[i].lnb_page,
1763                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1764                                   local_nb[i].lnb_len);
1765
1766                  /* corrupt the data after we compute the checksum, to
1767                  * simulate an OST->client data error */
1768                 if (i == 0 && opc == OST_READ &&
1769                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1770                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1771                         int len = local_nb[i].lnb_len;
1772                         struct page *np = tgt_page_to_corrupt;
1773
1774                         if (np) {
1775                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1776                                                         KM_USER0);
1777                                 char *ptr2 = page_address(np);
1778
1779                                 memcpy(ptr2 + off, ptr + off, len);
1780                                 memcpy(ptr2 + off, "bad4", min(4, len));
1781                                 ll_kunmap_atomic(ptr, KM_USER0);
1782
1783                                 /* LU-8376 to preserve original index for
1784                                  * display in dump_all_bulk_pages() */
1785                                 np->index = i;
1786
1787                                 cfs_crypto_hash_update_page(hdesc, np, off,
1788                                                             len);
1789                                 continue;
1790                         } else {
1791                                 CERROR("%s: can't alloc page for corruption\n",
1792                                        tgt_name(tgt));
1793                         }
1794                 }
1795         }
1796
1797         bufsize = sizeof(*cksum);
1798         err = cfs_crypto_hash_final(hdesc, (unsigned char *)cksum, &bufsize);
1799
1800         return 0;
1801 }
1802
1803 char dbgcksum_file_name[PATH_MAX];
1804
1805 static void dump_all_bulk_pages(struct obdo *oa, int count,
1806                                 struct niobuf_local *local_nb,
1807                                 __u32 server_cksum, __u32 client_cksum)
1808 {
1809         struct file *filp;
1810         int rc, i;
1811         unsigned int len;
1812         char *buf;
1813
1814         /* will only keep dump of pages on first error for the same range in
1815          * file/fid, not during the resends/retries. */
1816         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1817                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1818                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1819                   libcfs_debug_file_path_arr :
1820                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1821                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1822                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1823                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1824                  local_nb[0].lnb_file_offset,
1825                  local_nb[count-1].lnb_file_offset +
1826                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1827         filp = filp_open(dbgcksum_file_name,
1828                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1829         if (IS_ERR(filp)) {
1830                 rc = PTR_ERR(filp);
1831                 if (rc == -EEXIST)
1832                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1833                                "checksum error: rc = %d\n", dbgcksum_file_name,
1834                                rc);
1835                 else
1836                         CERROR("%s: can't open to dump pages with checksum "
1837                                "error: rc = %d\n", dbgcksum_file_name, rc);
1838                 return;
1839         }
1840
1841         for (i = 0; i < count; i++) {
1842                 len = local_nb[i].lnb_len;
1843                 buf = kmap(local_nb[i].lnb_page);
1844                 while (len != 0) {
1845                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1846                         if (rc < 0) {
1847                                 CERROR("%s: wanted to write %u but got %d "
1848                                        "error\n", dbgcksum_file_name, len, rc);
1849                                 break;
1850                         }
1851                         len -= rc;
1852                         buf += rc;
1853                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1854                                dbgcksum_file_name, rc);
1855                 }
1856                 kunmap(local_nb[i].lnb_page);
1857         }
1858
1859         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1860         if (rc)
1861                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1862         filp_close(filp, NULL);
1863         return;
1864 }
1865
1866 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1867                                struct obd_export *exp, struct obdo *oa,
1868                                const struct lnet_process_id *peer,
1869                                __u32 client_cksum, __u32 server_cksum,
1870                                enum cksum_types server_cksum_type)
1871 {
1872         char *msg;
1873         enum cksum_types cksum_type;
1874         loff_t start, end;
1875
1876         /* unlikely to happen and only if resend does not occur due to cksum
1877          * control failure on Client */
1878         if (unlikely(server_cksum == client_cksum)) {
1879                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1880                        client_cksum);
1881                 return 0;
1882         }
1883
1884         if (exp->exp_obd->obd_checksum_dump)
1885                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1886                                     client_cksum);
1887
1888         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1889                                            oa->o_flags : 0);
1890
1891         if (cksum_type != server_cksum_type)
1892                 msg = "the server may have not used the checksum type specified"
1893                       " in the original request - likely a protocol problem";
1894         else
1895                 msg = "should have changed on the client or in transit";
1896
1897         start = local_nb[0].lnb_file_offset;
1898         end = local_nb[npages-1].lnb_file_offset +
1899                                         local_nb[npages-1].lnb_len - 1;
1900
1901         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1902                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1903                 " %x (type %x), server csum %x (type %x)\n",
1904                 exp->exp_obd->obd_name,
1905                 msg, libcfs_nid2str(peer->nid),
1906                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1907                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1908                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1909                 POSTID(&oa->o_oi),
1910                 start, end, client_cksum, cksum_type, server_cksum,
1911                 server_cksum_type);
1912
1913         return 1;
1914 }
1915
1916 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1917                              unsigned char *buf, int size)
1918 {
1919         int     i, off, len, copied = size;
1920         char    *ptr;
1921
1922         for (i = 0; i < npages; i++) {
1923                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1924                 len = local[i].lnb_len;
1925
1926                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1927                        i, off, len, size);
1928                 if (len > size)
1929                         return -EINVAL;
1930
1931                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1932                 memcpy(buf + off, ptr, len);
1933                 ll_kunmap_atomic(ptr, KM_USER0);
1934                 buf += len;
1935                 size -= len;
1936         }
1937         return copied - size;
1938 }
1939
1940 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1941                                      struct niobuf_local *local_nb,
1942                                      int npages, int opc,
1943                                      obd_dif_csum_fn *fn,
1944                                      int sector_size,
1945                                      u32 *check_sum)
1946 {
1947         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1948         const char *obd_name = tgt->lut_obd->obd_name;
1949         struct cfs_crypto_hash_desc *hdesc;
1950         unsigned int bufsize;
1951         unsigned char *buffer;
1952         struct page *__page;
1953         __u16 *guard_start;
1954         int guard_number;
1955         int used_number = 0;
1956         __u32 cksum;
1957         int rc = 0;
1958         int used;
1959         int i;
1960
1961         __page = alloc_page(GFP_KERNEL);
1962         if (__page == NULL)
1963                 return -ENOMEM;
1964
1965         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1966         if (IS_ERR(hdesc)) {
1967                 CERROR("%s: unable to initialize checksum hash %s\n",
1968                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1969                 return PTR_ERR(hdesc);
1970         }
1971
1972         buffer = kmap(__page);
1973         guard_start = (__u16 *)buffer;
1974         guard_number = PAGE_SIZE / sizeof(*guard_start);
1975         for (i = 0; i < npages; i++) {
1976                 /* corrupt the data before we compute the checksum, to
1977                  * simulate a client->OST data error */
1978                 if (i == 0 && opc == OST_WRITE &&
1979                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1980                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1981                         int len = local_nb[i].lnb_len;
1982                         struct page *np = tgt_page_to_corrupt;
1983
1984                         if (np) {
1985                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1986                                                         KM_USER0);
1987                                 char *ptr2 = page_address(np);
1988
1989                                 memcpy(ptr2 + off, ptr + off, len);
1990                                 memcpy(ptr2 + off, "bad3", min(4, len));
1991                                 ll_kunmap_atomic(ptr, KM_USER0);
1992
1993                                 /* LU-8376 to preserve original index for
1994                                  * display in dump_all_bulk_pages() */
1995                                 np->index = i;
1996
1997                                 cfs_crypto_hash_update_page(hdesc, np, off,
1998                                                             len);
1999                                 continue;
2000                         } else {
2001                                 CERROR("%s: can't alloc page for corruption\n",
2002                                        tgt_name(tgt));
2003                         }
2004                 }
2005
2006                 /*
2007                  * The left guard number should be able to hold checksums of a
2008                  * whole page
2009                  */
2010                 rc = obd_page_dif_generate_buffer(obd_name,
2011                         local_nb[i].lnb_page,
2012                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2013                         local_nb[i].lnb_len, guard_start + used_number,
2014                         guard_number - used_number, &used, sector_size,
2015                         fn);
2016                 if (rc)
2017                         break;
2018
2019                 used_number += used;
2020                 if (used_number == guard_number) {
2021                         cfs_crypto_hash_update_page(hdesc, __page, 0,
2022                                 used_number * sizeof(*guard_start));
2023                         used_number = 0;
2024                 }
2025
2026                  /* corrupt the data after we compute the checksum, to
2027                  * simulate an OST->client data error */
2028                 if (unlikely(i == 0 && opc == OST_READ &&
2029                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2030                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2031                         int len = local_nb[i].lnb_len;
2032                         struct page *np = tgt_page_to_corrupt;
2033
2034                         if (np) {
2035                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2036                                                         KM_USER0);
2037                                 char *ptr2 = page_address(np);
2038
2039                                 memcpy(ptr2 + off, ptr + off, len);
2040                                 memcpy(ptr2 + off, "bad4", min(4, len));
2041                                 ll_kunmap_atomic(ptr, KM_USER0);
2042
2043                                 /* LU-8376 to preserve original index for
2044                                  * display in dump_all_bulk_pages() */
2045                                 np->index = i;
2046
2047                                 cfs_crypto_hash_update_page(hdesc, np, off,
2048                                                             len);
2049                                 continue;
2050                         } else {
2051                                 CERROR("%s: can't alloc page for corruption\n",
2052                                        tgt_name(tgt));
2053                         }
2054                 }
2055         }
2056         kunmap(__page);
2057         if (rc)
2058                 GOTO(out, rc);
2059
2060         if (used_number != 0)
2061                 cfs_crypto_hash_update_page(hdesc, __page, 0,
2062                         used_number * sizeof(*guard_start));
2063
2064         bufsize = sizeof(cksum);
2065         rc = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
2066
2067         if (rc == 0)
2068                 *check_sum = cksum;
2069 out:
2070         __free_page(__page);
2071         return rc;
2072 }
2073
2074 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2075                                   enum cksum_types cksum_type,
2076                                   struct niobuf_local *local_nb,
2077                                   int npages, int opc, u32 *check_sum)
2078 {
2079         obd_dif_csum_fn *fn = NULL;
2080         int sector_size = 0;
2081         int rc;
2082
2083         ENTRY;
2084         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2085
2086         if (fn)
2087                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2088                                                opc, fn, sector_size,
2089                                                check_sum);
2090         else
2091                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2092                                          cksum_type, check_sum);
2093         RETURN(rc);
2094 }
2095
2096 int tgt_brw_read(struct tgt_session_info *tsi)
2097 {
2098         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2099         struct ptlrpc_bulk_desc *desc = NULL;
2100         struct obd_export       *exp = tsi->tsi_exp;
2101         struct niobuf_remote    *remote_nb;
2102         struct niobuf_local     *local_nb;
2103         struct obd_ioobj        *ioo;
2104         struct ost_body         *body, *repbody;
2105         struct l_wait_info       lwi;
2106         struct lustre_handle     lockh = { 0 };
2107         int                      npages, nob = 0, rc, i, no_reply = 0,
2108                                  npages_read;
2109         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2110         const char *obd_name = exp->exp_obd->obd_name;
2111
2112         ENTRY;
2113
2114         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2115             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2116                 CERROR("%s: deny read request from %s to portal %u\n",
2117                        tgt_name(tsi->tsi_tgt),
2118                        obd_export_nid2str(req->rq_export),
2119                        ptlrpc_req2svc(req)->srv_req_portal);
2120                 RETURN(-EPROTO);
2121         }
2122
2123         req->rq_bulk_read = 1;
2124
2125         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2126                 RETURN(-EIO);
2127
2128         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2129                          cfs_fail_val : (obd_timeout + 1) / 4);
2130
2131         /* Check if there is eviction in progress, and if so, wait for it to
2132          * finish */
2133         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2134                 /* We do not care how long it takes */
2135                 lwi = LWI_INTR(NULL, NULL);
2136                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2137                          !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2138                          &lwi);
2139         }
2140
2141         /* There must be big cache in current thread to process this request
2142          * if it is NULL then something went wrong and it wasn't allocated,
2143          * report -ENOMEM in that case */
2144         if (tbc == NULL)
2145                 RETURN(-ENOMEM);
2146
2147         body = tsi->tsi_ost_body;
2148         LASSERT(body != NULL);
2149
2150         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2151         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2152
2153         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2154         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2155
2156         local_nb = tbc->local;
2157
2158         rc = tgt_brw_lock(exp, &tsi->tsi_resid, ioo, remote_nb, &lockh,
2159                           LCK_PR);
2160         if (rc != 0)
2161                 RETURN(rc);
2162
2163         /*
2164          * If getting the lock took more time than
2165          * client was willing to wait, drop it. b=11330
2166          */
2167         if (ktime_get_real_seconds() > req->rq_deadline ||
2168             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2169                 no_reply = 1;
2170                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2171                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2172                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2173                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2174                 GOTO(out_lock, rc = -ETIMEDOUT);
2175         }
2176
2177         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2178         repbody->oa = body->oa;
2179
2180         npages = PTLRPC_MAX_BRW_PAGES;
2181         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2182                         ioo, remote_nb, &npages, local_nb);
2183         if (rc != 0)
2184                 GOTO(out_lock, rc);
2185
2186         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2187                 desc = NULL;
2188         } else {
2189                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2190                                             PTLRPC_BULK_PUT_SOURCE |
2191                                                 PTLRPC_BULK_BUF_KIOV,
2192                                             OST_BULK_PORTAL,
2193                                             &ptlrpc_bulk_kiov_nopin_ops);
2194                 if (desc == NULL)
2195                         GOTO(out_commitrw, rc = -ENOMEM);
2196         }
2197
2198         nob = 0;
2199         npages_read = npages;
2200         for (i = 0; i < npages; i++) {
2201                 int page_rc = local_nb[i].lnb_rc;
2202
2203                 if (page_rc < 0) {
2204                         rc = page_rc;
2205                         npages_read = i;
2206                         break;
2207                 }
2208
2209                 nob += page_rc;
2210                 if (page_rc != 0 && desc != NULL) { /* some data! */
2211                         LASSERT(local_nb[i].lnb_page != NULL);
2212                         desc->bd_frag_ops->add_kiov_frag
2213                           (desc, local_nb[i].lnb_page,
2214                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2215                            page_rc);
2216                 }
2217
2218                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2219                         local_nb[i].lnb_len = page_rc;
2220                         npages_read = i + (page_rc != 0 ? 1 : 0);
2221                         /* All subsequent pages should be 0 */
2222                         while (++i < npages)
2223                                 LASSERT(local_nb[i].lnb_rc == 0);
2224                         break;
2225                 }
2226         }
2227         if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2228             nob != cfs_fail_val)
2229                 rc = -E2BIG;
2230
2231         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2232                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2233                            body->oa.o_flags : 0;
2234                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2235
2236                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2237                                                           cksum_type);
2238                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2239
2240                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2241                                             local_nb, npages_read, OST_READ,
2242                                             &repbody->oa.o_cksum);
2243                 if (rc < 0)
2244                         GOTO(out_commitrw, rc);
2245                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2246                        repbody->oa.o_cksum);
2247
2248                 /* if a resend it could be for a cksum error, so check Server
2249                  * cksum with returned Client cksum (this should even cover
2250                  * zero-cksum case) */
2251                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2252                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2253                         check_read_checksum(local_nb, npages_read, exp,
2254                                             &body->oa, &req->rq_peer,
2255                                             body->oa.o_cksum,
2256                                             repbody->oa.o_cksum, cksum_type);
2257         } else {
2258                 repbody->oa.o_valid = 0;
2259         }
2260         /* We're finishing using body->oa as an input variable */
2261
2262         /* Check if client was evicted while we were doing i/o before touching
2263          * network */
2264         if (rc == 0) {
2265                 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2266                         unsigned char *short_io_buf;
2267                         int short_io_size;
2268
2269                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2270                                                               &RMF_SHORT_IO);
2271                         short_io_size = req_capsule_get_size(&req->rq_pill,
2272                                                              &RMF_SHORT_IO,
2273                                                              RCL_SERVER);
2274                         rc = tgt_pages2shortio(local_nb, npages_read,
2275                                                short_io_buf, short_io_size);
2276                         if (rc >= 0)
2277                                 req_capsule_shrink(&req->rq_pill,
2278                                                    &RMF_SHORT_IO, rc,
2279                                                    RCL_SERVER);
2280                         rc = rc > 0 ? 0 : rc;
2281                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2282                         rc = target_bulk_io(exp, desc, &lwi);
2283                 }
2284                 no_reply = rc != 0;
2285         } else {
2286                 if (body->oa.o_flags & OBD_FL_SHORT_IO)
2287                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2288                                            RCL_SERVER);
2289         }
2290
2291 out_commitrw:
2292         /* Must commit after prep above in all cases */
2293         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2294                           remote_nb, npages, local_nb, rc);
2295 out_lock:
2296         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2297
2298         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2299                 ptlrpc_free_bulk(desc);
2300
2301         LASSERT(rc <= 0);
2302         if (rc == 0) {
2303                 rc = nob;
2304                 ptlrpc_lprocfs_brw(req, nob);
2305         } else if (no_reply) {
2306                 req->rq_no_reply = 1;
2307                 /* reply out callback would free */
2308                 ptlrpc_req_drop_rs(req);
2309                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2310                               "client will retry: rc %d\n",
2311                               obd_name,
2312                               obd_uuid2str(&exp->exp_client_uuid),
2313                               obd_export_nid2str(exp), rc);
2314         }
2315         /* send a bulk after reply to simulate a network delay or reordering
2316          * by a router - Note that !desc implies short io, so there is no bulk
2317          * to reorder. */
2318         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2319             desc) {
2320                 wait_queue_head_t        waitq;
2321                 struct l_wait_info       lwi1;
2322
2323                 CDEBUG(D_INFO, "reorder BULK\n");
2324                 init_waitqueue_head(&waitq);
2325
2326                 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2327                 l_wait_event(waitq, 0, &lwi1);
2328                 target_bulk_io(exp, desc, &lwi);
2329                 ptlrpc_free_bulk(desc);
2330         }
2331
2332         RETURN(rc);
2333 }
2334 EXPORT_SYMBOL(tgt_brw_read);
2335
2336 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2337                              unsigned char *buf, int size)
2338 {
2339         int     i, off, len;
2340         char    *ptr;
2341
2342         for (i = 0; i < npages; i++) {
2343                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2344                 len = local[i].lnb_len;
2345
2346                 if (len == 0)
2347                         continue;
2348
2349                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2350                        i, off, len, size);
2351                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2352                 if (ptr == NULL)
2353                         return -EINVAL;
2354                 memcpy(ptr + off, buf, len < size ? len : size);
2355                 ll_kunmap_atomic(ptr, KM_USER0);
2356                 buf += len;
2357                 size -= len;
2358         }
2359         return 0;
2360 }
2361
2362 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2363                               struct ptlrpc_bulk_desc *desc,
2364                               struct niobuf_local *local_nb, int npages,
2365                               u32 client_cksum, u32 server_cksum,
2366                               bool mmap)
2367 {
2368         struct obd_export *exp = req->rq_export;
2369         struct ost_body *body;
2370         char *router = "";
2371         char *via = "";
2372
2373         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2374         LASSERT(body != NULL);
2375
2376         if (desc && req->rq_peer.nid != desc->bd_sender) {
2377                 via = " via ";
2378                 router = libcfs_nid2str(desc->bd_sender);
2379         }
2380
2381         if (exp->exp_obd->obd_checksum_dump)
2382                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2383                                     client_cksum);
2384
2385         if (mmap) {
2386                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2387                              client_cksum, server_cksum);
2388                 return;
2389         }
2390
2391         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2392                            DFID" object "DOSTID" extent [%llu-%llu"
2393                            "]: client csum %x, server csum %x\n",
2394                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2395                            via, router,
2396                            body->oa.o_valid & OBD_MD_FLFID ?
2397                            body->oa.o_parent_seq : (__u64)0,
2398                            body->oa.o_valid & OBD_MD_FLFID ?
2399                            body->oa.o_parent_oid : 0,
2400                            body->oa.o_valid & OBD_MD_FLFID ?
2401                            body->oa.o_parent_ver : 0,
2402                            POSTID(&body->oa.o_oi),
2403                            local_nb[0].lnb_file_offset,
2404                            local_nb[npages-1].lnb_file_offset +
2405                            local_nb[npages - 1].lnb_len - 1,
2406                            client_cksum, server_cksum);
2407 }
2408
2409 int tgt_brw_write(struct tgt_session_info *tsi)
2410 {
2411         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2412         struct ptlrpc_bulk_desc *desc = NULL;
2413         struct obd_export       *exp = req->rq_export;
2414         struct niobuf_remote    *remote_nb;
2415         struct niobuf_local     *local_nb;
2416         struct obd_ioobj        *ioo;
2417         struct ost_body         *body, *repbody;
2418         struct l_wait_info       lwi;
2419         struct lustre_handle     lockh = {0};
2420         __u32                   *rcs;
2421         int                      objcount, niocount, npages;
2422         int                      rc, i, j;
2423         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2424         bool                     no_reply = false, mmap;
2425         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2426         bool wait_sync = false;
2427         const char *obd_name = exp->exp_obd->obd_name;
2428
2429         ENTRY;
2430
2431         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2432             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2433                 CERROR("%s: deny write request from %s to portal %u\n",
2434                        tgt_name(tsi->tsi_tgt),
2435                        obd_export_nid2str(req->rq_export),
2436                        ptlrpc_req2svc(req)->srv_req_portal);
2437                 RETURN(err_serious(-EPROTO));
2438         }
2439
2440         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2441                 RETURN(err_serious(-ENOSPC));
2442         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2443                 RETURN(err_serious(-EROFS));
2444
2445         req->rq_bulk_write = 1;
2446
2447         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2448                 RETURN(err_serious(-EIO));
2449         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2450                 RETURN(err_serious(-EFAULT));
2451
2452         /* pause before transaction has been started */
2453         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2454                          cfs_fail_val : (obd_timeout + 1) / 4);
2455
2456         /* There must be big cache in current thread to process this request
2457          * if it is NULL then something went wrong and it wasn't allocated,
2458          * report -ENOMEM in that case */
2459         if (tbc == NULL)
2460                 RETURN(-ENOMEM);
2461
2462         body = tsi->tsi_ost_body;
2463         LASSERT(body != NULL);
2464
2465         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2466         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2467
2468         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2469                                         RCL_CLIENT) / sizeof(*ioo);
2470
2471         for (niocount = i = 0; i < objcount; i++)
2472                 niocount += ioo[i].ioo_bufcnt;
2473
2474         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2475         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2476         if (niocount != req_capsule_get_size(&req->rq_pill,
2477                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2478                         sizeof(*remote_nb))
2479                 RETURN(err_serious(-EPROTO));
2480
2481         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2482             ptlrpc_connection_is_local(exp->exp_connection))
2483                 memory_pressure_set();
2484
2485         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2486                              niocount * sizeof(*rcs));
2487         rc = req_capsule_server_pack(&req->rq_pill);
2488         if (rc != 0)
2489                 GOTO(out, rc = err_serious(rc));
2490
2491         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2492         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2493
2494         local_nb = tbc->local;
2495
2496         rc = tgt_brw_lock(exp, &tsi->tsi_resid, ioo, remote_nb, &lockh,
2497                           LCK_PW);
2498         if (rc != 0)
2499                 GOTO(out, rc);
2500
2501         /*
2502          * If getting the lock took more time than
2503          * client was willing to wait, drop it. b=11330
2504          */
2505         if (ktime_get_real_seconds() > req->rq_deadline ||
2506             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2507                 no_reply = true;
2508                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2509                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2510                        POSTID(&ioo->ioo_oid),
2511                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2512                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2513                 GOTO(out_lock, rc = -ETIMEDOUT);
2514         }
2515
2516         /* Because we already sync grant info with client when reconnect,
2517          * grant info will be cleared for resent req, then fed_grant and
2518          * total_grant will not be modified in following preprw_write */
2519         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2520                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2521                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2522         }
2523
2524         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2525         if (repbody == NULL)
2526                 GOTO(out_lock, rc = -ENOMEM);
2527         repbody->oa = body->oa;
2528
2529         npages = PTLRPC_MAX_BRW_PAGES;
2530         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2531                         objcount, ioo, remote_nb, &npages, local_nb);
2532         if (rc < 0)
2533                 GOTO(out_lock, rc);
2534         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2535                 int short_io_size;
2536                 unsigned char *short_io_buf;
2537
2538                 short_io_size = req_capsule_get_size(&req->rq_pill,
2539                                                      &RMF_SHORT_IO,
2540                                                      RCL_CLIENT);
2541                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2542                                                       &RMF_SHORT_IO);
2543                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2544                                " size = %d\n", short_io_size);
2545
2546                 /* Copy short io buf to pages */
2547                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2548                                        short_io_size);
2549                 desc = NULL;
2550         } else {
2551                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2552                                             PTLRPC_BULK_GET_SINK |
2553                                             PTLRPC_BULK_BUF_KIOV,
2554                                             OST_BULK_PORTAL,
2555                                             &ptlrpc_bulk_kiov_nopin_ops);
2556                 if (desc == NULL)
2557                         GOTO(skip_transfer, rc = -ENOMEM);
2558
2559                 /* NB Having prepped, we must commit... */
2560                 for (i = 0; i < npages; i++)
2561                         desc->bd_frag_ops->add_kiov_frag(desc,
2562                                         local_nb[i].lnb_page,
2563                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2564                                         local_nb[i].lnb_len);
2565
2566                 rc = sptlrpc_svc_prep_bulk(req, desc);
2567                 if (rc != 0)
2568                         GOTO(skip_transfer, rc);
2569
2570                 rc = target_bulk_io(exp, desc, &lwi);
2571         }
2572
2573         no_reply = rc != 0;
2574
2575 skip_transfer:
2576         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2577                 static int cksum_counter;
2578
2579                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2580                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2581
2582                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2583                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2584                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2585                                                            cksum_type);
2586
2587                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2588                                             local_nb, npages, OST_WRITE,
2589                                             &repbody->oa.o_cksum);
2590                 if (rc < 0)
2591                         GOTO(out_commitrw, rc);
2592
2593                 cksum_counter++;
2594
2595                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2596                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2597                                 body->oa.o_flags & OBD_FL_MMAP);
2598
2599                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2600                                           body->oa.o_cksum,
2601                                           repbody->oa.o_cksum, mmap);
2602                         cksum_counter = 0;
2603                 } else if ((cksum_counter & (-cksum_counter)) ==
2604                            cksum_counter) {
2605                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2606                                cksum_counter, libcfs_id2str(req->rq_peer),
2607                                repbody->oa.o_cksum);
2608                 }
2609         }
2610
2611 out_commitrw:
2612         /* Must commit after prep above in all cases */
2613         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2614                           objcount, ioo, remote_nb, npages, local_nb, rc);
2615         if (rc == -ENOTCONN)
2616                 /* quota acquire process has been given up because
2617                  * either the client has been evicted or the client
2618                  * has timed out the request already */
2619                 no_reply = true;
2620
2621         for (i = 0; i < niocount; i++) {
2622                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2623                         wait_sync = true;
2624                         break;
2625                 }
2626         }
2627         /*
2628          * Disable sending mtime back to the client. If the client locked the
2629          * whole object, then it has already updated the mtime on its side,
2630          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2631          */
2632         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2633
2634         if (rc == 0) {
2635                 int nob = 0;
2636
2637                 /* set per-requested niobuf return codes */
2638                 for (i = j = 0; i < niocount; i++) {
2639                         int len = remote_nb[i].rnb_len;
2640
2641                         nob += len;
2642                         rcs[i] = 0;
2643                         do {
2644                                 LASSERT(j < npages);
2645                                 if (local_nb[j].lnb_rc < 0)
2646                                         rcs[i] = local_nb[j].lnb_rc;
2647                                 len -= local_nb[j].lnb_len;
2648                                 j++;
2649                         } while (len > 0);
2650                         LASSERT(len == 0);
2651                 }
2652                 LASSERT(j == npages);
2653                 ptlrpc_lprocfs_brw(req, nob);
2654         }
2655 out_lock:
2656         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2657         if (desc)
2658                 ptlrpc_free_bulk(desc);
2659 out:
2660         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2661                 req->rq_no_reply = 1;
2662                 /* reply out callback would free */
2663                 ptlrpc_req_drop_rs(req);
2664                 if (!exp->exp_obd->obd_no_transno)
2665                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2666                                       " client will retry: rc = %d\n",
2667                                       obd_name,
2668                                       obd_uuid2str(&exp->exp_client_uuid),
2669                                       obd_export_nid2str(exp), rc);
2670         }
2671         memory_pressure_clr();
2672         RETURN(rc);
2673 }
2674 EXPORT_SYMBOL(tgt_brw_write);
2675
2676 /* Check if request can be reconstructed from saved reply data
2677  * A copy of the reply data is returned in @trd if the pointer is not NULL
2678  */
2679 bool req_can_reconstruct(struct ptlrpc_request *req,
2680                          struct tg_reply_data *trd)
2681 {
2682         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2683         struct lsd_client_data *lcd = ted->ted_lcd;
2684         bool found;
2685
2686         if (tgt_is_multimodrpcs_client(req->rq_export))
2687                 return tgt_lookup_reply(req, trd);
2688
2689         mutex_lock(&ted->ted_lcd_lock);
2690         found = req->rq_xid == lcd->lcd_last_xid ||
2691                 req->rq_xid == lcd->lcd_last_close_xid;
2692
2693         if (found && trd != NULL) {
2694                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2695                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2696                         trd->trd_reply.lrd_transno =
2697                                                 lcd->lcd_last_close_transno;
2698                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2699                 } else {
2700                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2701                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2702                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2703                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2704                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2705                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2706                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2707                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2708                 }
2709         }
2710         mutex_unlock(&ted->ted_lcd_lock);
2711
2712         return found;
2713 }
2714 EXPORT_SYMBOL(req_can_reconstruct);
2715