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_val == 0 || cfs_fail_val == opc) &&
 803              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 804                 GOTO(out, rc = 0);
 805
 806         rc = lustre_msg_check_version(msg, h->th_version);
 807         if (unlikely(rc)) {
 808                 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
 809                           " %08x, expecting %08x\n", tgt_name(tgt),
 810                           lustre_msg_get_version(msg), h->th_version);
 811                 req->rq_status = -EINVAL;
 812                 rc = ptlrpc_error(req);
 813                 GOTO(out, rc);
 814         }
 815
 816 handle_recov:
 817         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 818         if (likely(rc == 1)) {
 819                 rc = tgt_handle_request0(tsi, h, req);
 820                 if (rc)
 821                         GOTO(out, rc);
 822         }
 823         EXIT;
 824 out:
 825         req_capsule_fini(tsi->tsi_pill);
 826         if (tsi->tsi_corpus != NULL) {
 827                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 828                 tsi->tsi_corpus = NULL;
 829         }
 830         return rc;
 831 }
 832 EXPORT_SYMBOL(tgt_request_handle);
 833
 834 /** Assign high priority operations to the request if needed. */
 835 int tgt_hpreq_handler(struct ptlrpc_request *req)
 836 {
 837         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 838         struct tgt_handler      *h;
 839         int                      rc;
 840
 841         ENTRY;
 842
 843         if (req->rq_export == NULL)
 844                 RETURN(0);
 845
 846         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 847         tsi->tsi_pill = &req->rq_pill;
 848         tsi->tsi_env = req->rq_svc_thread->t_env;
 849         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 850         tsi->tsi_exp = req->rq_export;
 851
 852         h = tgt_handler_find_check(req);
 853         if (IS_ERR(h)) {
 854                 rc = PTR_ERR(h);
 855                 RETURN(rc);
 856         }
 857
 858         rc = tgt_request_preprocess(tsi, h, req);
 859         if (unlikely(rc != 0))
 860                 RETURN(rc);
 861
 862         if (h->th_hp != NULL)
 863                 h->th_hp(tsi);
 864         RETURN(0);
 865 }
 866 EXPORT_SYMBOL(tgt_hpreq_handler);
 867
 868 void tgt_counter_incr(struct obd_export *exp, int opcode)
 869 {
 870         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 871         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 872                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 873 }
 874 EXPORT_SYMBOL(tgt_counter_incr);
 875
 876 /*
 877  * Unified target generic handlers.
 878  */
 879
 880 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 881 {
 882         struct lu_target        *tgt = class_exp2tgt(exp);
 883         struct sptlrpc_flavor    flvr;
 884         int                      rc = 0;
 885
 886         LASSERT(tgt);
 887         LASSERT(tgt->lut_obd);
 888         LASSERT(tgt->lut_slice);
 889
 890         /* always allow ECHO client */
 891         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 892                             LUSTRE_ECHO_NAME) == 0)) {
 893                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 894                 return 0;
 895         }
 896
 897         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 898                 read_lock(&tgt->lut_sptlrpc_lock);
 899                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 900                                              req->rq_sp_from,
 901                                              req->rq_peer.nid,
 902                                              &flvr);
 903                 read_unlock(&tgt->lut_sptlrpc_lock);
 904
 905                 spin_lock(&exp->exp_lock);
 906                 exp->exp_sp_peer = req->rq_sp_from;
 907                 exp->exp_flvr = flvr;
 908
 909                 /* when on mgs, if no restriction is set, or if client
 910                  * is loopback, allow any flavor */
 911                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 912                            LUSTRE_MGS_NAME) == 0) &&
 913                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 914                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 915                       == LOLND))
 916                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 917
 918                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 919                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 920                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 921                                "expect %x\n", tgt_name(tgt),
 922                                req->rq_flvr.sf_rpc,
 923                                libcfs_nid2str(req->rq_peer.nid),
 924                                exp->exp_flvr.sf_rpc);
 925                         rc = -EACCES;
 926                 }
 927                 spin_unlock(&exp->exp_lock);
 928         } else {
 929                 if (exp->exp_sp_peer != req->rq_sp_from) {
 930                         CERROR("%s: RPC source %s doesn't match %s\n",
 931                                tgt_name(tgt),
 932                                sptlrpc_part2name(req->rq_sp_from),
 933                                sptlrpc_part2name(exp->exp_sp_peer));
 934                         rc = -EACCES;
 935                 } else {
 936                         rc = sptlrpc_target_export_check(exp, req);
 937                 }
 938         }
 939
 940         return rc;
 941 }
 942
 943 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 944 {
 945         struct sptlrpc_rule_set  tmp_rset;
 946         int                      rc;
 947
 948         if (unlikely(tgt == NULL)) {
 949                 CERROR("No target passed");
 950                 return -EINVAL;
 951         }
 952
 953         sptlrpc_rule_set_init(&tmp_rset);
 954         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 955         if (rc) {
 956                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 957                        tgt_name(tgt), rc);
 958                 return rc;
 959         }
 960
 961         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 962
 963         write_lock(&tgt->lut_sptlrpc_lock);
 964         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 965         tgt->lut_sptlrpc_rset = tmp_rset;
 966         write_unlock(&tgt->lut_sptlrpc_lock);
 967
 968         return 0;
 969 }
 970 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 971
 972 int tgt_connect(struct tgt_session_info *tsi)
 973 {
 974         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 975         struct obd_connect_data *reply;
 976         int                      rc;
 977
 978         ENTRY;
 979
 980         /* XXX: better to call this check right after getting new export but
 981          * before last_rcvd slot allocation to avoid server load upon insecure
 982          * connects. This is to be fixed after unifiyng all targets.
 983          */
 984         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 985         if (rc)
 986                 GOTO(out, rc);
 987
 988         /* To avoid exposing partially initialized connection flags, changes up
 989          * to this point have been staged in reply->ocd_connect_flags. Now that
 990          * connection handling has completed successfully, atomically update
 991          * the connect flags in the shared export data structure. LU-1623 */
 992         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
 993         spin_lock(&tsi->tsi_exp->exp_lock);
 994         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
 995         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
 996                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
 997                         reply->ocd_connect_flags2;
 998         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
 999         spin_unlock(&tsi->tsi_exp->exp_lock);
1000
1001         RETURN(0);
1002 out:
1003         obd_disconnect(class_export_get(tsi->tsi_exp));
1004         return rc;
1005 }
1006 EXPORT_SYMBOL(tgt_connect);
1007
1008 int tgt_disconnect(struct tgt_session_info *tsi)
1009 {
1010         int rc;
1011
1012         ENTRY;
1013
1014         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1015
1016         rc = target_handle_disconnect(tgt_ses_req(tsi));
1017         if (rc)
1018                 RETURN(err_serious(rc));
1019
1020         RETURN(rc);
1021 }
1022 EXPORT_SYMBOL(tgt_disconnect);
1023
1024 /*
1025  * Unified target OBD handlers
1026  */
1027 int tgt_obd_ping(struct tgt_session_info *tsi)
1028 {
1029         int rc;
1030
1031         ENTRY;
1032
1033         rc = target_handle_ping(tgt_ses_req(tsi));
1034         if (rc)
1035                 RETURN(err_serious(rc));
1036
1037         RETURN(rc);
1038 }
1039 EXPORT_SYMBOL(tgt_obd_ping);
1040
1041 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1042 {
1043         return err_serious(-EOPNOTSUPP);
1044 }
1045
1046 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1047 {
1048         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1049         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1050         struct obd_export       *exp = req->rq_export;
1051         struct ptlrpc_bulk_desc *desc;
1052         struct l_wait_info      *lwi = &tti->tti_u.update.tti_wait_info;
1053         int                      i;
1054         int                      rc;
1055
1056         ENTRY;
1057
1058         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1059                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1060                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1061         if (desc == NULL)
1062                 RETURN(-ENOMEM);
1063
1064         for (i = 0; i < rdbuf->rb_nbufs; i++)
1065                 desc->bd_frag_ops->add_iov_frag(desc,
1066                                         rdbuf->rb_bufs[i].lb_buf,
1067                                         rdbuf->rb_bufs[i].lb_len);
1068
1069         rc = target_bulk_io(exp, desc, lwi);
1070         ptlrpc_free_bulk(desc);
1071         RETURN(rc);
1072 }
1073 EXPORT_SYMBOL(tgt_send_buffer);
1074
1075 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1076 {
1077         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1078         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1079         struct obd_export       *exp = req->rq_export;
1080         struct ptlrpc_bulk_desc *desc;
1081         struct l_wait_info      *lwi = &tti->tti_u.rdpg.tti_wait_info;
1082         int                      tmpcount;
1083         int                      tmpsize;
1084         int                      i;
1085         int                      rc;
1086
1087         ENTRY;
1088
1089         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1090                                     PTLRPC_BULK_PUT_SOURCE |
1091                                         PTLRPC_BULK_BUF_KIOV,
1092                                     MDS_BULK_PORTAL,
1093                                     &ptlrpc_bulk_kiov_pin_ops);
1094         if (desc == NULL)
1095                 RETURN(-ENOMEM);
1096
1097         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1098                 /* old client requires reply size in it's PAGE_SIZE,
1099                  * which is rdpg->rp_count */
1100                 nob = rdpg->rp_count;
1101
1102         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1103              i++, tmpcount -= tmpsize) {
1104                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1105                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1106                                                  tmpsize);
1107         }
1108
1109         LASSERT(desc->bd_nob == nob);
1110         rc = target_bulk_io(exp, desc, lwi);
1111         ptlrpc_free_bulk(desc);
1112         RETURN(rc);
1113 }
1114 EXPORT_SYMBOL(tgt_sendpage);
1115
1116 /*
1117  * OBD_IDX_READ handler
1118  */
1119 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1120 {
1121         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1122         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1123         struct idx_info         *req_ii, *rep_ii;
1124         int                      rc, i;
1125
1126         ENTRY;
1127
1128         memset(rdpg, 0, sizeof(*rdpg));
1129         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1130
1131         /* extract idx_info buffer from request & reply */
1132         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1133         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1134                 RETURN(err_serious(-EPROTO));
1135
1136         rc = req_capsule_server_pack(tsi->tsi_pill);
1137         if (rc)
1138                 RETURN(err_serious(rc));
1139
1140         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1141         if (rep_ii == NULL)
1142                 RETURN(err_serious(-EFAULT));
1143         rep_ii->ii_magic = IDX_INFO_MAGIC;
1144
1145         /* extract hash to start with */
1146         rdpg->rp_hash = req_ii->ii_hash_start;
1147
1148         /* extract requested attributes */
1149         rdpg->rp_attrs = req_ii->ii_attrs;
1150
1151         /* check that fid packed in request is valid and supported */
1152         if (!fid_is_sane(&req_ii->ii_fid))
1153                 RETURN(-EINVAL);
1154         rep_ii->ii_fid = req_ii->ii_fid;
1155
1156         /* copy flags */
1157         rep_ii->ii_flags = req_ii->ii_flags;
1158
1159         /* compute number of pages to allocate, ii_count is the number of 4KB
1160          * containers */
1161         if (req_ii->ii_count <= 0)
1162                 GOTO(out, rc = -EFAULT);
1163         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1164                                exp_max_brw_size(tsi->tsi_exp));
1165         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1166
1167         /* allocate pages to store the containers */
1168         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1169         if (rdpg->rp_pages == NULL)
1170                 GOTO(out, rc = -ENOMEM);
1171         for (i = 0; i < rdpg->rp_npages; i++) {
1172                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1173                 if (rdpg->rp_pages[i] == NULL)
1174                         GOTO(out, rc = -ENOMEM);
1175         }
1176
1177         /* populate pages with key/record pairs */
1178         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1179         if (rc < 0)
1180                 GOTO(out, rc);
1181
1182         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1183                  "asked %d > %d\n", rc, rdpg->rp_count);
1184
1185         /* send pages to client */
1186         rc = tgt_sendpage(tsi, rdpg, rc);
1187         if (rc)
1188                 GOTO(out, rc);
1189         EXIT;
1190 out:
1191         if (rdpg->rp_pages) {
1192                 for (i = 0; i < rdpg->rp_npages; i++)
1193                         if (rdpg->rp_pages[i])
1194                                 __free_page(rdpg->rp_pages[i]);
1195                 OBD_FREE(rdpg->rp_pages,
1196                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1197         }
1198         return rc;
1199 }
1200
1201 struct tgt_handler tgt_obd_handlers[] = {
1202 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1203 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1204 };
1205 EXPORT_SYMBOL(tgt_obd_handlers);
1206
1207 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1208              struct dt_object *obj, __u64 start, __u64 end)
1209 {
1210         int rc = 0;
1211
1212         ENTRY;
1213
1214         /* if no objid is specified, it means "sync whole filesystem" */
1215         if (obj == NULL) {
1216                 rc = dt_sync(env, tgt->lut_bottom);
1217         } else if (dt_version_get(env, obj) >
1218                    tgt->lut_obd->obd_last_committed) {
1219                 rc = dt_object_sync(env, obj, start, end);
1220         }
1221         atomic_inc(&tgt->lut_sync_count);
1222
1223         RETURN(rc);
1224 }
1225 EXPORT_SYMBOL(tgt_sync);
1226 /*
1227  * Unified target DLM handlers.
1228  */
1229
1230 /**
1231  * Unified target BAST
1232  *
1233  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1234  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1235  * but if it's cross-MDT lock, because remote object version is not set, a
1236  * filesystem sync is needed.
1237  *
1238  * \param lock server side lock
1239  * \param desc lock desc
1240  * \param data ldlm_cb_set_arg
1241  * \param flag  indicates whether this cancelling or blocking callback
1242  * \retval      0 on success
1243  * \retval      negative number on error
1244  */
1245 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1246                             void *data, int flag)
1247 {
1248         struct lu_env            env;
1249         struct lu_target        *tgt;
1250         struct dt_object        *obj = NULL;
1251         struct lu_fid            fid;
1252         int                      rc = 0;
1253
1254         ENTRY;
1255
1256         tgt = class_exp2tgt(lock->l_export);
1257
1258         if (unlikely(tgt == NULL)) {
1259                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1260                        class_exp2obd(lock->l_export)->obd_name);
1261                 RETURN(-EINVAL);
1262         }
1263
1264         if (flag == LDLM_CB_CANCELING &&
1265             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1266             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1267              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1268               ldlm_is_cbpending(lock))) &&
1269             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1270              lock->l_resource->lr_type == LDLM_EXTENT)) {
1271                 __u64 start = 0;
1272                 __u64 end = OBD_OBJECT_EOF;
1273
1274                 rc = lu_env_init(&env, LCT_DT_THREAD);
1275                 if (unlikely(rc != 0))
1276                         GOTO(err, rc);
1277
1278                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1279                                    tgt->lut_lsd.lsd_osd_index);
1280
1281                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1282                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1283                         if (IS_ERR(obj))
1284                                 GOTO(err_env, rc = PTR_ERR(obj));
1285
1286                         if (!dt_object_exists(obj))
1287                                 GOTO(err_put, rc = -ENOENT);
1288
1289                         start = lock->l_policy_data.l_extent.start;
1290                         end = lock->l_policy_data.l_extent.end;
1291                 }
1292
1293                 rc = tgt_sync(&env, tgt, obj, start, end);
1294                 if (rc < 0) {
1295                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1296                                "cancel: rc = %d\n",
1297                                tgt_name(tgt), PFID(&fid),
1298                                lock->l_policy_data.l_extent.start,
1299                                lock->l_policy_data.l_extent.end, rc);
1300                 }
1301 err_put:
1302                 if (obj != NULL)
1303                         dt_object_put(&env, obj);
1304 err_env:
1305                 lu_env_fini(&env);
1306         }
1307 err:
1308         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1309         RETURN(rc);
1310 }
1311
1312 static struct ldlm_callback_suite tgt_dlm_cbs = {
1313         .lcs_completion = ldlm_server_completion_ast,
1314         .lcs_blocking   = tgt_blocking_ast,
1315         .lcs_glimpse    = ldlm_server_glimpse_ast
1316 };
1317
1318 int tgt_enqueue(struct tgt_session_info *tsi)
1319 {
1320         struct ptlrpc_request *req = tgt_ses_req(tsi);
1321         int rc;
1322
1323         ENTRY;
1324         /*
1325          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1326          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1327          */
1328         LASSERT(tsi->tsi_dlm_req != NULL);
1329         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1330                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1331         if (rc)
1332                 RETURN(err_serious(rc));
1333
1334         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1335         case LUT_FAIL_MDT:
1336                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1337                 break;
1338         case LUT_FAIL_OST:
1339                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1340                 break;
1341         case LUT_FAIL_MGT:
1342                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1343                 break;
1344         default:
1345                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1346                 break;
1347         }
1348         RETURN(req->rq_status);
1349 }
1350 EXPORT_SYMBOL(tgt_enqueue);
1351
1352 int tgt_convert(struct tgt_session_info *tsi)
1353 {
1354         struct ptlrpc_request *req = tgt_ses_req(tsi);
1355         int rc;
1356
1357         ENTRY;
1358         LASSERT(tsi->tsi_dlm_req);
1359         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1360         if (rc)
1361                 RETURN(err_serious(rc));
1362
1363         RETURN(req->rq_status);
1364 }
1365
1366 int tgt_bl_callback(struct tgt_session_info *tsi)
1367 {
1368         return err_serious(-EOPNOTSUPP);
1369 }
1370
1371 int tgt_cp_callback(struct tgt_session_info *tsi)
1372 {
1373         return err_serious(-EOPNOTSUPP);
1374 }
1375
1376 /* generic LDLM target handler */
1377 struct tgt_handler tgt_dlm_handlers[] = {
1378 TGT_DLM_HDL    (HABEO_CLAVIS,   LDLM_ENQUEUE,           tgt_enqueue),
1379 TGT_DLM_HDL    (HABEO_CLAVIS,   LDLM_CONVERT,           tgt_convert),
1380 TGT_DLM_HDL_VAR(0,              LDLM_BL_CALLBACK,       tgt_bl_callback),
1381 TGT_DLM_HDL_VAR(0,              LDLM_CP_CALLBACK,       tgt_cp_callback)
1382 };
1383 EXPORT_SYMBOL(tgt_dlm_handlers);
1384
1385 /*
1386  * Unified target LLOG handlers.
1387  */
1388 int tgt_llog_open(struct tgt_session_info *tsi)
1389 {
1390         int rc;
1391
1392         ENTRY;
1393
1394         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1395
1396         RETURN(rc);
1397 }
1398 EXPORT_SYMBOL(tgt_llog_open);
1399
1400 int tgt_llog_read_header(struct tgt_session_info *tsi)
1401 {
1402         int rc;
1403
1404         ENTRY;
1405
1406         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1407
1408         RETURN(rc);
1409 }
1410 EXPORT_SYMBOL(tgt_llog_read_header);
1411
1412 int tgt_llog_next_block(struct tgt_session_info *tsi)
1413 {
1414         int rc;
1415
1416         ENTRY;
1417
1418         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1419
1420         RETURN(rc);
1421 }
1422 EXPORT_SYMBOL(tgt_llog_next_block);
1423
1424 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1425 {
1426         int rc;
1427
1428         ENTRY;
1429
1430         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1431
1432         RETURN(rc);
1433 }
1434 EXPORT_SYMBOL(tgt_llog_prev_block);
1435
1436 /* generic llog target handler */
1437 struct tgt_handler tgt_llog_handlers[] = {
1438 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1439 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1440 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1441 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1442 };
1443 EXPORT_SYMBOL(tgt_llog_handlers);
1444
1445 /*
1446  * sec context handlers
1447  */
1448 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1449 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1450 {
1451         return 0;
1452 }
1453
1454 struct tgt_handler tgt_sec_ctx_handlers[] = {
1455 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1456 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1457 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1458 };
1459 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1460
1461 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1462                                  struct dt_device *key,
1463                                  struct lfsck_req_local *lrl,
1464                                  struct thandle *th) = NULL;
1465
1466 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1467                                                       struct dt_device *,
1468                                                       struct lfsck_req_local *,
1469                                                       struct thandle *))
1470 {
1471         tgt_lfsck_in_notify_local = notify;
1472 }
1473 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1474
1475 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1476                            struct dt_device *key,
1477                            struct lfsck_request *lr) = NULL;
1478
1479 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1480                                                 struct dt_device *,
1481                                                 struct lfsck_request *))
1482 {
1483         tgt_lfsck_in_notify = notify;
1484 }
1485 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1486
1487 static int (*tgt_lfsck_query)(const struct lu_env *env,
1488                               struct dt_device *key,
1489                               struct lfsck_request *req,
1490                               struct lfsck_reply *rep,
1491                               struct lfsck_query *que) = NULL;
1492
1493 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1494                                            struct dt_device *,
1495                                            struct lfsck_request *,
1496                                            struct lfsck_reply *,
1497                                            struct lfsck_query *))
1498 {
1499         tgt_lfsck_query = query;
1500 }
1501 EXPORT_SYMBOL(tgt_register_lfsck_query);
1502
1503 /* LFSCK request handlers */
1504 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1505 {
1506         const struct lu_env     *env = tsi->tsi_env;
1507         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1508         struct lfsck_request    *lr;
1509         int                      rc;
1510         ENTRY;
1511
1512         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1513         if (lr == NULL)
1514                 RETURN(-EPROTO);
1515
1516         rc = tgt_lfsck_in_notify(env, key, lr);
1517
1518         RETURN(rc);
1519 }
1520
1521 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1522 {
1523         struct lfsck_request    *request;
1524         struct lfsck_reply      *reply;
1525         int                      rc;
1526         ENTRY;
1527
1528         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1529         if (request == NULL)
1530                 RETURN(-EPROTO);
1531
1532         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1533         if (reply == NULL)
1534                 RETURN(-ENOMEM);
1535
1536         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1537                              request, reply, NULL);
1538
1539         RETURN(rc < 0 ? rc : 0);
1540 }
1541
1542 struct tgt_handler tgt_lfsck_handlers[] = {
1543 TGT_LFSCK_HDL(HABEO_REFERO,     LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1544 TGT_LFSCK_HDL(HABEO_REFERO,     LFSCK_QUERY,    tgt_handle_lfsck_query),
1545 };
1546 EXPORT_SYMBOL(tgt_lfsck_handlers);
1547
1548 /*
1549  * initialize per-thread page pool (bug 5137).
1550  */
1551 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1552 {
1553         struct tgt_thread_big_cache *tbc;
1554
1555         ENTRY;
1556
1557         LASSERT(thread != NULL);
1558         LASSERT(thread->t_data == NULL);
1559
1560         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1561         if (tbc == NULL)
1562                 RETURN(-ENOMEM);
1563         thread->t_data = tbc;
1564         RETURN(0);
1565 }
1566 EXPORT_SYMBOL(tgt_io_thread_init);
1567
1568 /*
1569  * free per-thread pool created by tgt_thread_init().
1570  */
1571 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1572 {
1573         struct tgt_thread_big_cache *tbc;
1574
1575         ENTRY;
1576
1577         LASSERT(thread != NULL);
1578
1579         /*
1580          * be prepared to handle partially-initialized pools (because this is
1581          * called from ost_io_thread_init() for cleanup.
1582          */
1583         tbc = thread->t_data;
1584         if (tbc != NULL) {
1585                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1586                 thread->t_data = NULL;
1587         }
1588         EXIT;
1589 }
1590 EXPORT_SYMBOL(tgt_io_thread_done);
1591
1592 /**
1593  * Helper function for getting Data-on-MDT file server DLM lock
1594  * if asked by client.
1595  */
1596 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1597                       struct lustre_handle *lh, int mode, __u64 *flags)
1598 {
1599         union ldlm_policy_data policy = {
1600                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1601         };
1602         int rc;
1603
1604         ENTRY;
1605
1606         LASSERT(lh != NULL);
1607         LASSERT(ns != NULL);
1608         LASSERT(!lustre_handle_is_used(lh));
1609
1610         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1611                                     flags, ldlm_blocking_ast,
1612                                     ldlm_completion_ast, ldlm_glimpse_ast,
1613                                     NULL, 0, LVB_T_NONE, NULL, lh);
1614
1615         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1616 }
1617 EXPORT_SYMBOL(tgt_mdt_data_lock);
1618
1619 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1620 {
1621         LASSERT(lustre_handle_is_used(lh));
1622         ldlm_lock_decref(lh, mode);
1623 }
1624 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1625
1626 /**
1627  * Helper function for getting server side [start, start+count] DLM lock
1628  * if asked by client.
1629  */
1630 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1631                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1632                     struct lustre_handle *lh, int mode, __u64 *flags)
1633 {
1634         union ldlm_policy_data policy;
1635         int rc;
1636
1637         ENTRY;
1638
1639         LASSERT(lh != NULL);
1640         LASSERT(ns != NULL);
1641         LASSERT(!lustre_handle_is_used(lh));
1642
1643         policy.l_extent.gid = 0;
1644         policy.l_extent.start = start & PAGE_MASK;
1645
1646         /*
1647          * If ->o_blocks is EOF it means "lock till the end of the file".
1648          * Otherwise, it's size of an extent or hole being punched (in bytes).
1649          */
1650         if (end == OBD_OBJECT_EOF || end < start)
1651                 policy.l_extent.end = OBD_OBJECT_EOF;
1652         else
1653                 policy.l_extent.end = end | ~PAGE_MASK;
1654
1655         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1656                                     mode, flags, ldlm_blocking_ast,
1657                                     ldlm_completion_ast, ldlm_glimpse_ast,
1658                                     NULL, 0, LVB_T_NONE, NULL, lh);
1659         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1660 }
1661 EXPORT_SYMBOL(tgt_extent_lock);
1662
1663 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1664 {
1665         LASSERT(lustre_handle_is_used(lh));
1666         ldlm_lock_decref(lh, mode);
1667 }
1668 EXPORT_SYMBOL(tgt_extent_unlock);
1669
1670 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1671                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1672                         struct niobuf_remote *nb, struct lustre_handle *lh,
1673                         enum ldlm_mode mode)
1674 {
1675         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1676         __u64                    flags = 0;
1677         int                      nrbufs = obj->ioo_bufcnt;
1678         int                      i;
1679         int                      rc;
1680
1681         ENTRY;
1682
1683         LASSERT(mode == LCK_PR || mode == LCK_PW);
1684         LASSERT(!lustre_handle_is_used(lh));
1685
1686         if (ns->ns_obd->obd_recovering)
1687                 RETURN(0);
1688
1689         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1690                 RETURN(0);
1691
1692         for (i = 1; i < nrbufs; i++)
1693                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1694                         RETURN(-EFAULT);
1695
1696         /* MDT IO for data-on-mdt */
1697         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1698                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1699         else
1700                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1701                                      nb[nrbufs - 1].rnb_offset +
1702                                      nb[nrbufs - 1].rnb_len - 1,
1703                                      lh, mode, &flags);
1704         RETURN(rc);
1705 }
1706
1707 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1708                            struct lustre_handle *lh, enum ldlm_mode mode)
1709 {
1710         ENTRY;
1711
1712         LASSERT(mode == LCK_PR || mode == LCK_PW);
1713         LASSERT((obj->ioo_bufcnt > 0 &&
1714                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1715                 lustre_handle_is_used(lh));
1716
1717         if (lustre_handle_is_used(lh))
1718                 tgt_extent_unlock(lh, mode);
1719         EXIT;
1720 }
1721 static int tgt_checksum_niobuf(struct lu_target *tgt,
1722                                  struct niobuf_local *local_nb, int npages,
1723                                  int opc, enum cksum_types cksum_type,
1724                                  __u32 *cksum)
1725 {
1726         struct ahash_request           *req;
1727         unsigned int                    bufsize;
1728         int                             i, err;
1729         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1730
1731         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1732         if (IS_ERR(req)) {
1733                 CERROR("%s: unable to initialize checksum hash %s\n",
1734                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1735                 return PTR_ERR(req);
1736         }
1737
1738         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1739         for (i = 0; i < npages; i++) {
1740                 /* corrupt the data before we compute the checksum, to
1741                  * simulate a client->OST data error */
1742                 if (i == 0 && opc == OST_WRITE &&
1743                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1744                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1745                         int len = local_nb[i].lnb_len;
1746                         struct page *np = tgt_page_to_corrupt;
1747
1748                         if (np) {
1749                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1750                                                         KM_USER0);
1751                                 char *ptr2 = page_address(np);
1752
1753                                 memcpy(ptr2 + off, ptr + off, len);
1754                                 memcpy(ptr2 + off, "bad3", min(4, len));
1755                                 ll_kunmap_atomic(ptr, KM_USER0);
1756
1757                                 /* LU-8376 to preserve original index for
1758                                  * display in dump_all_bulk_pages() */
1759                                 np->index = i;
1760
1761                                 cfs_crypto_hash_update_page(req, np, off,
1762                                                             len);
1763                                 continue;
1764                         } else {
1765                                 CERROR("%s: can't alloc page for corruption\n",
1766                                        tgt_name(tgt));
1767                         }
1768                 }
1769                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1770                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1771                                   local_nb[i].lnb_len);
1772
1773                  /* corrupt the data after we compute the checksum, to
1774                  * simulate an OST->client data error */
1775                 if (i == 0 && opc == OST_READ &&
1776                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1777                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1778                         int len = local_nb[i].lnb_len;
1779                         struct page *np = tgt_page_to_corrupt;
1780
1781                         if (np) {
1782                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1783                                                         KM_USER0);
1784                                 char *ptr2 = page_address(np);
1785
1786                                 memcpy(ptr2 + off, ptr + off, len);
1787                                 memcpy(ptr2 + off, "bad4", min(4, len));
1788                                 ll_kunmap_atomic(ptr, KM_USER0);
1789
1790                                 /* LU-8376 to preserve original index for
1791                                  * display in dump_all_bulk_pages() */
1792                                 np->index = i;
1793
1794                                 cfs_crypto_hash_update_page(req, np, off,
1795                                                             len);
1796                                 continue;
1797                         } else {
1798                                 CERROR("%s: can't alloc page for corruption\n",
1799                                        tgt_name(tgt));
1800                         }
1801                 }
1802         }
1803
1804         bufsize = sizeof(*cksum);
1805         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1806
1807         return 0;
1808 }
1809
1810 char dbgcksum_file_name[PATH_MAX];
1811
1812 static void dump_all_bulk_pages(struct obdo *oa, int count,
1813                                 struct niobuf_local *local_nb,
1814                                 __u32 server_cksum, __u32 client_cksum)
1815 {
1816         struct file *filp;
1817         int rc, i;
1818         unsigned int len;
1819         char *buf;
1820
1821         /* will only keep dump of pages on first error for the same range in
1822          * file/fid, not during the resends/retries. */
1823         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1824                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1825                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1826                   libcfs_debug_file_path_arr :
1827                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1828                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1829                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1830                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1831                  local_nb[0].lnb_file_offset,
1832                  local_nb[count-1].lnb_file_offset +
1833                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1834         filp = filp_open(dbgcksum_file_name,
1835                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1836         if (IS_ERR(filp)) {
1837                 rc = PTR_ERR(filp);
1838                 if (rc == -EEXIST)
1839                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1840                                "checksum error: rc = %d\n", dbgcksum_file_name,
1841                                rc);
1842                 else
1843                         CERROR("%s: can't open to dump pages with checksum "
1844                                "error: rc = %d\n", dbgcksum_file_name, rc);
1845                 return;
1846         }
1847
1848         for (i = 0; i < count; i++) {
1849                 len = local_nb[i].lnb_len;
1850                 buf = kmap(local_nb[i].lnb_page);
1851                 while (len != 0) {
1852                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1853                         if (rc < 0) {
1854                                 CERROR("%s: wanted to write %u but got %d "
1855                                        "error\n", dbgcksum_file_name, len, rc);
1856                                 break;
1857                         }
1858                         len -= rc;
1859                         buf += rc;
1860                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1861                                dbgcksum_file_name, rc);
1862                 }
1863                 kunmap(local_nb[i].lnb_page);
1864         }
1865
1866         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1867         if (rc)
1868                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1869         filp_close(filp, NULL);
1870         return;
1871 }
1872
1873 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1874                                struct obd_export *exp, struct obdo *oa,
1875                                const struct lnet_process_id *peer,
1876                                __u32 client_cksum, __u32 server_cksum,
1877                                enum cksum_types server_cksum_type)
1878 {
1879         char *msg;
1880         enum cksum_types cksum_type;
1881         loff_t start, end;
1882
1883         /* unlikely to happen and only if resend does not occur due to cksum
1884          * control failure on Client */
1885         if (unlikely(server_cksum == client_cksum)) {
1886                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1887                        client_cksum);
1888                 return 0;
1889         }
1890
1891         if (exp->exp_obd->obd_checksum_dump)
1892                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1893                                     client_cksum);
1894
1895         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1896                                            oa->o_flags : 0);
1897
1898         if (cksum_type != server_cksum_type)
1899                 msg = "the server may have not used the checksum type specified"
1900                       " in the original request - likely a protocol problem";
1901         else
1902                 msg = "should have changed on the client or in transit";
1903
1904         start = local_nb[0].lnb_file_offset;
1905         end = local_nb[npages-1].lnb_file_offset +
1906                                         local_nb[npages-1].lnb_len - 1;
1907
1908         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1909                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1910                 " %x (type %x), server csum %x (type %x)\n",
1911                 exp->exp_obd->obd_name,
1912                 msg, libcfs_nid2str(peer->nid),
1913                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1914                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1915                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1916                 POSTID(&oa->o_oi),
1917                 start, end, client_cksum, cksum_type, server_cksum,
1918                 server_cksum_type);
1919
1920         return 1;
1921 }
1922
1923 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1924                              unsigned char *buf, int size)
1925 {
1926         int     i, off, len, copied = size;
1927         char    *ptr;
1928
1929         for (i = 0; i < npages; i++) {
1930                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1931                 len = local[i].lnb_len;
1932
1933                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1934                        i, off, len, size);
1935                 if (len > size)
1936                         return -EINVAL;
1937
1938                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1939                 memcpy(buf + off, ptr, len);
1940                 ll_kunmap_atomic(ptr, KM_USER0);
1941                 buf += len;
1942                 size -= len;
1943         }
1944         return copied - size;
1945 }
1946
1947 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1948                                      struct niobuf_local *local_nb,
1949                                      int npages, int opc,
1950                                      obd_dif_csum_fn *fn,
1951                                      int sector_size,
1952                                      u32 *check_sum)
1953 {
1954         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1955         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1956         const char *obd_name = tgt->lut_obd->obd_name;
1957         struct ahash_request *req;
1958         unsigned int bufsize;
1959         unsigned char *buffer;
1960         struct page *__page;
1961         __u16 *guard_start;
1962         int guard_number;
1963         int used_number = 0;
1964         __u32 cksum;
1965         int rc = 0;
1966         int used;
1967         int i;
1968
1969         __page = alloc_page(GFP_KERNEL);
1970         if (__page == NULL)
1971                 return -ENOMEM;
1972
1973         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1974         if (IS_ERR(req)) {
1975                 CERROR("%s: unable to initialize checksum hash %s\n",
1976                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1977                 return PTR_ERR(req);
1978         }
1979
1980         buffer = kmap(__page);
1981         guard_start = (__u16 *)buffer;
1982         guard_number = PAGE_SIZE / sizeof(*guard_start);
1983         for (i = 0; i < npages; i++) {
1984                 /* corrupt the data before we compute the checksum, to
1985                  * simulate a client->OST data error */
1986                 if (i == 0 && opc == OST_WRITE &&
1987                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1988                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1989                         int len = local_nb[i].lnb_len;
1990                         struct page *np = tgt_page_to_corrupt;
1991
1992                         if (np) {
1993                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1994                                                         KM_USER0);
1995                                 char *ptr2 = page_address(np);
1996
1997                                 memcpy(ptr2 + off, ptr + off, len);
1998                                 memcpy(ptr2 + off, "bad3", min(4, len));
1999                                 ll_kunmap_atomic(ptr, KM_USER0);
2000
2001                                 /* LU-8376 to preserve original index for
2002                                  * display in dump_all_bulk_pages() */
2003                                 np->index = i;
2004
2005                                 cfs_crypto_hash_update_page(req, np, off,
2006                                                             len);
2007                                 continue;
2008                         } else {
2009                                 CERROR("%s: can't alloc page for corruption\n",
2010                                        tgt_name(tgt));
2011                         }
2012                 }
2013
2014                 /*
2015                  * The left guard number should be able to hold checksums of a
2016                  * whole page
2017                  */
2018                 if (t10_cksum_type && opc == OST_READ &&
2019                     local_nb[i].lnb_guard_disk) {
2020                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2021                         if (used > (guard_number - used_number)) {
2022                                 rc = -E2BIG;
2023                                 break;
2024                         }
2025                         memcpy(guard_start + used_number,
2026                                local_nb[i].lnb_guards,
2027                                used * sizeof(*local_nb[i].lnb_guards));
2028                 } else {
2029                         rc = obd_page_dif_generate_buffer(obd_name,
2030                                 local_nb[i].lnb_page,
2031                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2032                                 local_nb[i].lnb_len, guard_start + used_number,
2033                                 guard_number - used_number, &used, sector_size,
2034                                 fn);
2035                         if (rc)
2036                                 break;
2037                 }
2038
2039                 LASSERT(used <= MAX_GUARD_NUMBER);
2040                 /*
2041                  * If disk support T10PI checksum, copy guards to local_nb.
2042                  * If the write is partial page, do not use the guards for bio
2043                  * submission since the data might not be full-sector. The bio
2044                  * guards will be generated later based on the full sectors. If
2045                  * the sector size is 512B rather than 4 KB, or the page size
2046                  * is larger than 4KB, this might drop some useful guards for
2047                  * partial page write, but it will only add minimal extra time
2048                  * of checksum calculation.
2049                  */
2050                 if (t10_cksum_type && opc == OST_WRITE &&
2051                     local_nb[i].lnb_len == PAGE_SIZE) {
2052                         local_nb[i].lnb_guard_rpc = 1;
2053                         memcpy(local_nb[i].lnb_guards,
2054                                guard_start + used_number,
2055                                used * sizeof(*local_nb[i].lnb_guards));
2056                 }
2057
2058                 used_number += used;
2059                 if (used_number == guard_number) {
2060                         cfs_crypto_hash_update_page(req, __page, 0,
2061                                 used_number * sizeof(*guard_start));
2062                         used_number = 0;
2063                 }
2064
2065                  /* corrupt the data after we compute the checksum, to
2066                  * simulate an OST->client data error */
2067                 if (unlikely(i == 0 && opc == OST_READ &&
2068                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2069                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2070                         int len = local_nb[i].lnb_len;
2071                         struct page *np = tgt_page_to_corrupt;
2072
2073                         if (np) {
2074                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2075                                                         KM_USER0);
2076                                 char *ptr2 = page_address(np);
2077
2078                                 memcpy(ptr2 + off, ptr + off, len);
2079                                 memcpy(ptr2 + off, "bad4", min(4, len));
2080                                 ll_kunmap_atomic(ptr, KM_USER0);
2081
2082                                 /* LU-8376 to preserve original index for
2083                                  * display in dump_all_bulk_pages() */
2084                                 np->index = i;
2085
2086                                 cfs_crypto_hash_update_page(req, np, off,
2087                                                             len);
2088                                 continue;
2089                         } else {
2090                                 CERROR("%s: can't alloc page for corruption\n",
2091                                        tgt_name(tgt));
2092                         }
2093                 }
2094         }
2095         kunmap(__page);
2096         if (rc)
2097                 GOTO(out, rc);
2098
2099         if (used_number != 0)
2100                 cfs_crypto_hash_update_page(req, __page, 0,
2101                         used_number * sizeof(*guard_start));
2102
2103         bufsize = sizeof(cksum);
2104         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2105
2106         if (rc == 0)
2107                 *check_sum = cksum;
2108 out:
2109         __free_page(__page);
2110         return rc;
2111 }
2112
2113 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2114                                   enum cksum_types cksum_type,
2115                                   struct niobuf_local *local_nb,
2116                                   int npages, int opc, u32 *check_sum)
2117 {
2118         obd_dif_csum_fn *fn = NULL;
2119         int sector_size = 0;
2120         int rc;
2121
2122         ENTRY;
2123         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2124
2125         if (fn)
2126                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2127                                                opc, fn, sector_size,
2128                                                check_sum);
2129         else
2130                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2131                                          cksum_type, check_sum);
2132         RETURN(rc);
2133 }
2134
2135 int tgt_brw_read(struct tgt_session_info *tsi)
2136 {
2137         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2138         struct ptlrpc_bulk_desc *desc = NULL;
2139         struct obd_export       *exp = tsi->tsi_exp;
2140         struct niobuf_remote    *remote_nb;
2141         struct niobuf_local     *local_nb;
2142         struct obd_ioobj        *ioo;
2143         struct ost_body         *body, *repbody;
2144         struct l_wait_info       lwi;
2145         struct lustre_handle     lockh = { 0 };
2146         int                      npages, nob = 0, rc, i, no_reply = 0,
2147                                  npages_read;
2148         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2149         const char *obd_name = exp->exp_obd->obd_name;
2150
2151         ENTRY;
2152
2153         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2154             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2155                 CERROR("%s: deny read request from %s to portal %u\n",
2156                        tgt_name(tsi->tsi_tgt),
2157                        obd_export_nid2str(req->rq_export),
2158                        ptlrpc_req2svc(req)->srv_req_portal);
2159                 RETURN(-EPROTO);
2160         }
2161
2162         req->rq_bulk_read = 1;
2163
2164         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2165                 RETURN(-EIO);
2166
2167         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2168                          cfs_fail_val : (obd_timeout + 1) / 4);
2169
2170         /* Check if there is eviction in progress, and if so, wait for it to
2171          * finish */
2172         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2173                 /* We do not care how long it takes */
2174                 lwi = LWI_INTR(NULL, NULL);
2175                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2176                          !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2177                          &lwi);
2178         }
2179
2180         /* There must be big cache in current thread to process this request
2181          * if it is NULL then something went wrong and it wasn't allocated,
2182          * report -ENOMEM in that case */
2183         if (tbc == NULL)
2184                 RETURN(-ENOMEM);
2185
2186         body = tsi->tsi_ost_body;
2187         LASSERT(body != NULL);
2188
2189         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2190         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2191
2192         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2193         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2194
2195         local_nb = tbc->local;
2196
2197         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2198                           &lockh, LCK_PR);
2199         if (rc != 0)
2200                 RETURN(rc);
2201
2202         /*
2203          * If getting the lock took more time than
2204          * client was willing to wait, drop it. b=11330
2205          */
2206         if (ktime_get_real_seconds() > req->rq_deadline ||
2207             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2208                 no_reply = 1;
2209                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2210                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2211                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2212                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2213                 GOTO(out_lock, rc = -ETIMEDOUT);
2214         }
2215
2216         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2217         repbody->oa = body->oa;
2218
2219         npages = PTLRPC_MAX_BRW_PAGES;
2220         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2221                         ioo, remote_nb, &npages, local_nb);
2222         if (rc != 0)
2223                 GOTO(out_lock, rc);
2224
2225         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2226                 desc = NULL;
2227         } else {
2228                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2229                                             PTLRPC_BULK_PUT_SOURCE |
2230                                                 PTLRPC_BULK_BUF_KIOV,
2231                                             OST_BULK_PORTAL,
2232                                             &ptlrpc_bulk_kiov_nopin_ops);
2233                 if (desc == NULL)
2234                         GOTO(out_commitrw, rc = -ENOMEM);
2235         }
2236
2237         nob = 0;
2238         npages_read = npages;
2239         for (i = 0; i < npages; i++) {
2240                 int page_rc = local_nb[i].lnb_rc;
2241
2242                 if (page_rc < 0) {
2243                         rc = page_rc;
2244                         npages_read = i;
2245                         break;
2246                 }
2247
2248                 nob += page_rc;
2249                 if (page_rc != 0 && desc != NULL) { /* some data! */
2250                         LASSERT(local_nb[i].lnb_page != NULL);
2251                         desc->bd_frag_ops->add_kiov_frag
2252                           (desc, local_nb[i].lnb_page,
2253                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2254                            page_rc);
2255                 }
2256
2257                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2258                         local_nb[i].lnb_len = page_rc;
2259                         npages_read = i + (page_rc != 0 ? 1 : 0);
2260                         /* All subsequent pages should be 0 */
2261                         while (++i < npages)
2262                                 LASSERT(local_nb[i].lnb_rc == 0);
2263                         break;
2264                 }
2265         }
2266         if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2267             nob != cfs_fail_val)
2268                 rc = -E2BIG;
2269
2270         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2271                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2272                            body->oa.o_flags : 0;
2273                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2274
2275                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2276                                                           cksum_type);
2277                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2278
2279                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2280                                             local_nb, npages_read, OST_READ,
2281                                             &repbody->oa.o_cksum);
2282                 if (rc < 0)
2283                         GOTO(out_commitrw, rc);
2284                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2285                        repbody->oa.o_cksum);
2286
2287                 /* if a resend it could be for a cksum error, so check Server
2288                  * cksum with returned Client cksum (this should even cover
2289                  * zero-cksum case) */
2290                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2291                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2292                         check_read_checksum(local_nb, npages_read, exp,
2293                                             &body->oa, &req->rq_peer,
2294                                             body->oa.o_cksum,
2295                                             repbody->oa.o_cksum, cksum_type);
2296         } else {
2297                 repbody->oa.o_valid = 0;
2298         }
2299         /* We're finishing using body->oa as an input variable */
2300
2301         /* Check if client was evicted while we were doing i/o before touching
2302          * network */
2303         if (rc == 0) {
2304                 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2305                         unsigned char *short_io_buf;
2306                         int short_io_size;
2307
2308                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2309                                                               &RMF_SHORT_IO);
2310                         short_io_size = req_capsule_get_size(&req->rq_pill,
2311                                                              &RMF_SHORT_IO,
2312                                                              RCL_SERVER);
2313                         rc = tgt_pages2shortio(local_nb, npages_read,
2314                                                short_io_buf, short_io_size);
2315                         if (rc >= 0)
2316                                 req_capsule_shrink(&req->rq_pill,
2317                                                    &RMF_SHORT_IO, rc,
2318                                                    RCL_SERVER);
2319                         rc = rc > 0 ? 0 : rc;
2320                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2321                         rc = target_bulk_io(exp, desc, &lwi);
2322                 }
2323                 no_reply = rc != 0;
2324         } else {
2325                 if (body->oa.o_flags & OBD_FL_SHORT_IO)
2326                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2327                                            RCL_SERVER);
2328         }
2329
2330 out_commitrw:
2331         /* Must commit after prep above in all cases */
2332         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2333                           remote_nb, npages, local_nb, rc);
2334 out_lock:
2335         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2336
2337         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2338                 ptlrpc_free_bulk(desc);
2339
2340         LASSERT(rc <= 0);
2341         if (rc == 0) {
2342                 rc = nob;
2343                 ptlrpc_lprocfs_brw(req, nob);
2344         } else if (no_reply) {
2345                 req->rq_no_reply = 1;
2346                 /* reply out callback would free */
2347                 ptlrpc_req_drop_rs(req);
2348                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2349                               "client will retry: rc %d\n",
2350                               obd_name,
2351                               obd_uuid2str(&exp->exp_client_uuid),
2352                               obd_export_nid2str(exp), rc);
2353         }
2354         /* send a bulk after reply to simulate a network delay or reordering
2355          * by a router - Note that !desc implies short io, so there is no bulk
2356          * to reorder. */
2357         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2358             desc) {
2359                 wait_queue_head_t        waitq;
2360                 struct l_wait_info       lwi1;
2361
2362                 CDEBUG(D_INFO, "reorder BULK\n");
2363                 init_waitqueue_head(&waitq);
2364
2365                 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2366                 l_wait_event(waitq, 0, &lwi1);
2367                 target_bulk_io(exp, desc, &lwi);
2368                 ptlrpc_free_bulk(desc);
2369         }
2370
2371         RETURN(rc);
2372 }
2373 EXPORT_SYMBOL(tgt_brw_read);
2374
2375 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2376                              unsigned char *buf, int size)
2377 {
2378         int     i, off, len;
2379         char    *ptr;
2380
2381         for (i = 0; i < npages; i++) {
2382                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2383                 len = local[i].lnb_len;
2384
2385                 if (len == 0)
2386                         continue;
2387
2388                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2389                        i, off, len, size);
2390                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2391                 if (ptr == NULL)
2392                         return -EINVAL;
2393                 memcpy(ptr + off, buf, len < size ? len : size);
2394                 ll_kunmap_atomic(ptr, KM_USER0);
2395                 buf += len;
2396                 size -= len;
2397         }
2398         return 0;
2399 }
2400
2401 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2402                               struct ptlrpc_bulk_desc *desc,
2403                               struct niobuf_local *local_nb, int npages,
2404                               u32 client_cksum, u32 server_cksum,
2405                               bool mmap)
2406 {
2407         struct obd_export *exp = req->rq_export;
2408         struct ost_body *body;
2409         char *router = "";
2410         char *via = "";
2411
2412         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2413         LASSERT(body != NULL);
2414
2415         if (desc && req->rq_peer.nid != desc->bd_sender) {
2416                 via = " via ";
2417                 router = libcfs_nid2str(desc->bd_sender);
2418         }
2419
2420         if (exp->exp_obd->obd_checksum_dump)
2421                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2422                                     client_cksum);
2423
2424         if (mmap) {
2425                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2426                              client_cksum, server_cksum);
2427                 return;
2428         }
2429
2430         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2431                            DFID" object "DOSTID" extent [%llu-%llu"
2432                            "]: client csum %x, server csum %x\n",
2433                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2434                            via, router,
2435                            body->oa.o_valid & OBD_MD_FLFID ?
2436                            body->oa.o_parent_seq : (__u64)0,
2437                            body->oa.o_valid & OBD_MD_FLFID ?
2438                            body->oa.o_parent_oid : 0,
2439                            body->oa.o_valid & OBD_MD_FLFID ?
2440                            body->oa.o_parent_ver : 0,
2441                            POSTID(&body->oa.o_oi),
2442                            local_nb[0].lnb_file_offset,
2443                            local_nb[npages-1].lnb_file_offset +
2444                            local_nb[npages - 1].lnb_len - 1,
2445                            client_cksum, server_cksum);
2446 }
2447
2448 int tgt_brw_write(struct tgt_session_info *tsi)
2449 {
2450         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2451         struct ptlrpc_bulk_desc *desc = NULL;
2452         struct obd_export       *exp = req->rq_export;
2453         struct niobuf_remote    *remote_nb;
2454         struct niobuf_local     *local_nb;
2455         struct obd_ioobj        *ioo;
2456         struct ost_body         *body, *repbody;
2457         struct l_wait_info       lwi;
2458         struct lustre_handle     lockh = {0};
2459         __u32                   *rcs;
2460         int                      objcount, niocount, npages;
2461         int                      rc, i, j;
2462         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2463         bool                     no_reply = false, mmap;
2464         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2465         bool wait_sync = false;
2466         const char *obd_name = exp->exp_obd->obd_name;
2467
2468         ENTRY;
2469
2470         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2471             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2472                 CERROR("%s: deny write request from %s to portal %u\n",
2473                        tgt_name(tsi->tsi_tgt),
2474                        obd_export_nid2str(req->rq_export),
2475                        ptlrpc_req2svc(req)->srv_req_portal);
2476                 RETURN(err_serious(-EPROTO));
2477         }
2478
2479         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2480                 RETURN(err_serious(-ENOSPC));
2481         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2482                 RETURN(err_serious(-EROFS));
2483
2484         req->rq_bulk_write = 1;
2485
2486         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2487                 RETURN(err_serious(-EIO));
2488         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2489                 RETURN(err_serious(-EFAULT));
2490
2491         /* pause before transaction has been started */
2492         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2493                          cfs_fail_val : (obd_timeout + 1) / 4);
2494
2495         /* There must be big cache in current thread to process this request
2496          * if it is NULL then something went wrong and it wasn't allocated,
2497          * report -ENOMEM in that case */
2498         if (tbc == NULL)
2499                 RETURN(-ENOMEM);
2500
2501         body = tsi->tsi_ost_body;
2502         LASSERT(body != NULL);
2503
2504         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2505         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2506
2507         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2508                                         RCL_CLIENT) / sizeof(*ioo);
2509
2510         for (niocount = i = 0; i < objcount; i++)
2511                 niocount += ioo[i].ioo_bufcnt;
2512
2513         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2514         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2515         if (niocount != req_capsule_get_size(&req->rq_pill,
2516                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2517                         sizeof(*remote_nb))
2518                 RETURN(err_serious(-EPROTO));
2519
2520         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2521             ptlrpc_connection_is_local(exp->exp_connection))
2522                 memory_pressure_set();
2523
2524         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2525                              niocount * sizeof(*rcs));
2526         rc = req_capsule_server_pack(&req->rq_pill);
2527         if (rc != 0)
2528                 GOTO(out, rc = err_serious(rc));
2529
2530         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2531         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2532
2533         local_nb = tbc->local;
2534
2535         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2536                           &lockh, LCK_PW);
2537         if (rc != 0)
2538                 GOTO(out, rc);
2539
2540         /*
2541          * If getting the lock took more time than
2542          * client was willing to wait, drop it. b=11330
2543          */
2544         if (ktime_get_real_seconds() > req->rq_deadline ||
2545             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2546                 no_reply = true;
2547                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2548                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2549                        POSTID(&ioo->ioo_oid),
2550                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2551                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2552                 GOTO(out_lock, rc = -ETIMEDOUT);
2553         }
2554
2555         /* Because we already sync grant info with client when reconnect,
2556          * grant info will be cleared for resent req, then fed_grant and
2557          * total_grant will not be modified in following preprw_write */
2558         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2559                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2560                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2561         }
2562
2563         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2564         if (repbody == NULL)
2565                 GOTO(out_lock, rc = -ENOMEM);
2566         repbody->oa = body->oa;
2567
2568         npages = PTLRPC_MAX_BRW_PAGES;
2569         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2570                         objcount, ioo, remote_nb, &npages, local_nb);
2571         if (rc < 0)
2572                 GOTO(out_lock, rc);
2573         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2574                 int short_io_size;
2575                 unsigned char *short_io_buf;
2576
2577                 short_io_size = req_capsule_get_size(&req->rq_pill,
2578                                                      &RMF_SHORT_IO,
2579                                                      RCL_CLIENT);
2580                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2581                                                       &RMF_SHORT_IO);
2582                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2583                                " size = %d\n", short_io_size);
2584
2585                 /* Copy short io buf to pages */
2586                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2587                                        short_io_size);
2588                 desc = NULL;
2589         } else {
2590                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2591                                             PTLRPC_BULK_GET_SINK |
2592                                             PTLRPC_BULK_BUF_KIOV,
2593                                             OST_BULK_PORTAL,
2594                                             &ptlrpc_bulk_kiov_nopin_ops);
2595                 if (desc == NULL)
2596                         GOTO(skip_transfer, rc = -ENOMEM);
2597
2598                 /* NB Having prepped, we must commit... */
2599                 for (i = 0; i < npages; i++)
2600                         desc->bd_frag_ops->add_kiov_frag(desc,
2601                                         local_nb[i].lnb_page,
2602                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2603                                         local_nb[i].lnb_len);
2604
2605                 rc = sptlrpc_svc_prep_bulk(req, desc);
2606                 if (rc != 0)
2607                         GOTO(skip_transfer, rc);
2608
2609                 rc = target_bulk_io(exp, desc, &lwi);
2610         }
2611
2612         no_reply = rc != 0;
2613
2614 skip_transfer:
2615         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2616                 static int cksum_counter;
2617
2618                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2619                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2620
2621                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2622                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2623                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2624                                                            cksum_type);
2625
2626                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2627                                             local_nb, npages, OST_WRITE,
2628                                             &repbody->oa.o_cksum);
2629                 if (rc < 0)
2630                         GOTO(out_commitrw, rc);
2631
2632                 cksum_counter++;
2633
2634                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2635                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2636                                 body->oa.o_flags & OBD_FL_MMAP);
2637
2638                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2639                                           body->oa.o_cksum,
2640                                           repbody->oa.o_cksum, mmap);
2641                         cksum_counter = 0;
2642                 } else if ((cksum_counter & (-cksum_counter)) ==
2643                            cksum_counter) {
2644                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2645                                cksum_counter, libcfs_id2str(req->rq_peer),
2646                                repbody->oa.o_cksum);
2647                 }
2648         }
2649
2650 out_commitrw:
2651         /* Must commit after prep above in all cases */
2652         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2653                           objcount, ioo, remote_nb, npages, local_nb, rc);
2654         if (rc == -ENOTCONN)
2655                 /* quota acquire process has been given up because
2656                  * either the client has been evicted or the client
2657                  * has timed out the request already */
2658                 no_reply = true;
2659
2660         for (i = 0; i < niocount; i++) {
2661                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2662                         wait_sync = true;
2663                         break;
2664                 }
2665         }
2666         /*
2667          * Disable sending mtime back to the client. If the client locked the
2668          * whole object, then it has already updated the mtime on its side,
2669          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2670          */
2671         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2672
2673         if (rc == 0) {
2674                 int nob = 0;
2675
2676                 /* set per-requested niobuf return codes */
2677                 for (i = j = 0; i < niocount; i++) {
2678                         int len = remote_nb[i].rnb_len;
2679
2680                         nob += len;
2681                         rcs[i] = 0;
2682                         do {
2683                                 LASSERT(j < npages);
2684                                 if (local_nb[j].lnb_rc < 0)
2685                                         rcs[i] = local_nb[j].lnb_rc;
2686                                 len -= local_nb[j].lnb_len;
2687                                 j++;
2688                         } while (len > 0);
2689                         LASSERT(len == 0);
2690                 }
2691                 LASSERT(j == npages);
2692                 ptlrpc_lprocfs_brw(req, nob);
2693         }
2694 out_lock:
2695         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2696         if (desc)
2697                 ptlrpc_free_bulk(desc);
2698 out:
2699         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2700                 req->rq_no_reply = 1;
2701                 /* reply out callback would free */
2702                 ptlrpc_req_drop_rs(req);
2703                 if (!exp->exp_obd->obd_no_transno)
2704                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2705                                       " client will retry: rc = %d\n",
2706                                       obd_name,
2707                                       obd_uuid2str(&exp->exp_client_uuid),
2708                                       obd_export_nid2str(exp), rc);
2709         }
2710         memory_pressure_clr();
2711         RETURN(rc);
2712 }
2713 EXPORT_SYMBOL(tgt_brw_write);
2714
2715 /* Check if request can be reconstructed from saved reply data
2716  * A copy of the reply data is returned in @trd if the pointer is not NULL
2717  */
2718 bool req_can_reconstruct(struct ptlrpc_request *req,
2719                          struct tg_reply_data *trd)
2720 {
2721         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2722         struct lsd_client_data *lcd = ted->ted_lcd;
2723         bool found;
2724
2725         if (tgt_is_multimodrpcs_client(req->rq_export))
2726                 return tgt_lookup_reply(req, trd);
2727
2728         mutex_lock(&ted->ted_lcd_lock);
2729         found = req->rq_xid == lcd->lcd_last_xid ||
2730                 req->rq_xid == lcd->lcd_last_close_xid;
2731
2732         if (found && trd != NULL) {
2733                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2734                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2735                         trd->trd_reply.lrd_transno =
2736                                                 lcd->lcd_last_close_transno;
2737                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2738                 } else {
2739                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2740                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2741                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2742                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2743                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2744                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2745                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2746                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2747                 }
2748         }
2749         mutex_unlock(&ted->ted_lcd_lock);
2750
2751         return found;
2752 }
2753 EXPORT_SYMBOL(req_can_reconstruct);
2754