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