lustre/target/tgt_handler.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2013, 2017, Intel Corporation.
  25  */
  26 /*
  27  * lustre/target/tgt_handler.c
  28  *
  29  * Lustre Unified Target request handler code
  30  *
  31  * Author: Brian Behlendorf <behlendorf1@llnl.gov>
  32  * Author: Mikhail Pershin <mike.pershin@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_CLASS
  36
  37 #include <linux/user_namespace.h>
  38 #include <linux/delay.h>
  39 #include <linux/uidgid.h>
  40
  41 #include <libcfs/linux/linux-mem.h>
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47 #include <lustre_acl.h>
  48
  49 #include "tgt_internal.h"
  50
  51 char *tgt_name(struct lu_target *tgt)
  52 {
  53         LASSERT(tgt->lut_obd != NULL);
  54         return tgt->lut_obd->obd_name;
  55 }
  56 EXPORT_SYMBOL(tgt_name);
  57
  58 /*
  59  * Generic code handling requests that have struct mdt_body passed in:
  60  *
  61  *  - extract mdt_body from request and save it in @tsi, if present;
  62  *
  63  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  64  *  @tsi;
  65  *
  66  *  - if HAS_BODY flag is set for this request type check whether object
  67  *  actually exists on storage (lu_object_exists()).
  68  *
  69  */
  70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  71 {
  72         const struct mdt_body   *body;
  73         struct lu_object        *obj;
  74         struct req_capsule      *pill = tsi->tsi_pill;
  75         int                      rc;
  76
  77         ENTRY;
  78
  79         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EFAULT);
  82
  83         tsi->tsi_mdt_body = body;
  84
  85         if (!(body->mbo_valid & OBD_MD_FLID))
  86                 RETURN(0);
  87
  88         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  89          * in any case in pre-2.5 clients. Fix that here if needed */
  90         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  91                 RETURN(0);
  92
  93         if (!fid_is_sane(&body->mbo_fid1)) {
  94                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  95                        PFID(&body->mbo_fid1));
  96                 RETURN(-EINVAL);
  97         }
  98
  99         obj = lu_object_find(tsi->tsi_env,
 100                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 101                              &body->mbo_fid1, NULL);
 102         if (!IS_ERR(obj)) {
 103                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 104                         lu_object_put(tsi->tsi_env, obj);
 105                         rc = -ENOENT;
 106                 } else {
 107                         tsi->tsi_corpus = obj;
 108                         rc = 0;
 109                 }
 110         } else {
 111                 rc = PTR_ERR(obj);
 112         }
 113
 114         tsi->tsi_fid = body->mbo_fid1;
 115
 116         RETURN(rc);
 117 }
 118
 119 /**
 120  * Validate oa from client.
 121  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 122  * req are valid.
 123  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 124  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 125  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 126  *       whether this is oi_fid or real ostid. So it will check
 127  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 128  *    c. Old FID-disable osc will send IDIF.
 129  *    d. new FID-enable osc/osp will send normal FID.
 130  *
 131  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 132  * be used for LAST_ID file, and only being accessed inside OST now.
 133  */
 134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 135 {
 136         struct ost_id   *oi     = &oa->o_oi;
 137         u64              seq    = ostid_seq(oi);
 138         u64              id     = ostid_id(oi);
 139         int              rc;
 140         ENTRY;
 141
 142         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 143                      fid_seq_is_echo(seq))) {
 144                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 145                  * during create, and we will reset this to 1, since this
 146                  * oi_id is basically useless in the following create process,
 147                  * but oi_id == 0 will make it difficult to tell whether it is
 148                  * real FID or ost_id. */
 149                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 150                 oi->oi_fid.f_oid = id ?: 1;
 151                 oi->oi_fid.f_ver = 0;
 152         } else {
 153                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 154
 155                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 156                         GOTO(out, rc = -EPROTO);
 157
 158                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 159                  * all of the requests are required to setup FLGROUP */
 160                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 161                         ostid_set_seq_mdt0(oi);
 162                         oa->o_valid |= OBD_MD_FLGROUP;
 163                         seq = ostid_seq(oi);
 164                 }
 165
 166                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 167                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 168                         GOTO(out, rc = -EPROTO);
 169
 170                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 171                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 172                 if (unlikely(rc != 0))
 173                         GOTO(out, rc);
 174
 175                 oi->oi_fid = tti->tti_fid1;
 176         }
 177
 178         RETURN(0);
 179
 180 out:
 181         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 182                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 183                seq, id, rc);
 184         return rc;
 185 }
 186 EXPORT_SYMBOL(tgt_validate_obdo);
 187
 188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 189 {
 190         unsigned                 max_brw;
 191         struct niobuf_remote    *rnb;
 192         struct obd_ioobj        *ioo;
 193         int                      obj_count;
 194
 195         ENTRY;
 196
 197         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 198         if (ioo == NULL)
 199                 RETURN(-EPROTO);
 200
 201         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 202         if (rnb == NULL)
 203                 RETURN(-EPROTO);
 204
 205         max_brw = ioobj_max_brw_get(ioo);
 206         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 207                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 208                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 209                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 210                        POSTID(oi), -EPROTO);
 211                 RETURN(-EPROTO);
 212         }
 213         ioo->ioo_oid = *oi;
 214
 215         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 216                                         RCL_CLIENT) / sizeof(*ioo);
 217         if (obj_count == 0) {
 218                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 219                 RETURN(-EPROTO);
 220         } else if (obj_count > 1) {
 221                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 222                        obj_count);
 223                 RETURN(-EPROTO);
 224         }
 225
 226         if (ioo->ioo_bufcnt == 0) {
 227                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 228                 RETURN(-EPROTO);
 229         }
 230
 231         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 232                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 233                           "bulk has too many pages (%d)",
 234                           ioo->ioo_bufcnt);
 235                 RETURN(-EPROTO);
 236         }
 237
 238         RETURN(0);
 239 }
 240
 241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 242 {
 243         struct ost_body         *body;
 244         struct req_capsule      *pill = tsi->tsi_pill;
 245         struct lu_nodemap       *nodemap;
 246         int                      rc;
 247
 248         ENTRY;
 249
 250         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 251         if (body == NULL)
 252                 RETURN(-EFAULT);
 253
 254         rc = tgt_validate_obdo(tsi, &body->oa);
 255         if (rc)
 256                 RETURN(rc);
 257
 258         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 259         if (IS_ERR(nodemap))
 260                 RETURN(PTR_ERR(nodemap));
 261
 262         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_uid);
 265         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 266                                         NODEMAP_CLIENT_TO_FS,
 267                                         body->oa.o_gid);
 268         body->oa.o_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
 269                                            NODEMAP_CLIENT_TO_FS,
 270                                            body->oa.o_projid);
 271         nodemap_putref(nodemap);
 272
 273         tsi->tsi_ost_body = body;
 274         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 275
 276         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 277                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 278                 if (rc < 0)
 279                         RETURN(rc);
 280         }
 281
 282         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 283                 if (flags & HAS_BODY) {
 284                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
 285                                tgt_name(tsi->tsi_tgt));
 286                         RETURN(-EPROTO);
 287                 } else {
 288                         RETURN(0);
 289                 }
 290         }
 291
 292         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 293
 294         /*
 295          * OST doesn't get object in advance for further use to prevent
 296          * situations with nested object_find which is potential deadlock.
 297          */
 298         tsi->tsi_corpus = NULL;
 299         RETURN(rc);
 300 }
 301
 302 /*
 303  * Do necessary preprocessing according to handler ->th_flags.
 304  */
 305 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 306                                   struct tgt_handler *h,
 307                                   struct ptlrpc_request *req)
 308 {
 309         struct req_capsule      *pill = tsi->tsi_pill;
 310         __u32                    flags = h->th_flags;
 311         int                      rc = 0;
 312
 313         ENTRY;
 314
 315         if (tsi->tsi_preprocessed)
 316                 RETURN(0);
 317
 318         LASSERT(h->th_act != NULL);
 319         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 320         LASSERT(current->journal_info == NULL);
 321
 322         LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
 323                      h->th_fmt != NULL));
 324         if (h->th_fmt != NULL) {
 325                 req_capsule_set(pill, h->th_fmt);
 326                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT) &&
 327                     req_capsule_field_present(pill, &RMF_MDT_BODY,
 328                                               RCL_CLIENT)) {
 329                         rc = tgt_mdt_body_unpack(tsi, flags);
 330                         if (rc < 0)
 331                                 RETURN(rc);
 332                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 333                                                  RCL_CLIENT) &&
 334                            req_capsule_field_present(pill, &RMF_OST_BODY,
 335                                                  RCL_CLIENT)) {
 336                         rc = tgt_ost_body_unpack(tsi, flags);
 337                         if (rc < 0)
 338                                 RETURN(rc);
 339                 }
 340         }
 341
 342         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 343                 RETURN(-EROFS);
 344
 345         if (flags & HAS_KEY) {
 346                 struct ldlm_request *dlm_req;
 347
 348                 LASSERT(h->th_fmt != NULL);
 349
 350                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 351                 if (dlm_req != NULL) {
 352                         union ldlm_wire_policy_data *policy =
 353                                         &dlm_req->lock_desc.l_policy_data;
 354
 355                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 356                                      LDLM_IBITS &&
 357                                      (policy->l_inodebits.bits |
 358                                       policy->l_inodebits.try_bits) == 0)) {
 359                                 /*
 360                                  * Lock without inodebits makes no sense and
 361                                  * will oops later in ldlm. If client miss to
 362                                  * set such bits, do not trigger ASSERTION.
 363                                  *
 364                                  * For liblustre flock case, it maybe zero.
 365                                  */
 366                                 rc = -EPROTO;
 367                         } else {
 368                                 tsi->tsi_dlm_req = dlm_req;
 369                         }
 370                 } else {
 371                         rc = -EFAULT;
 372                 }
 373         }
 374         tsi->tsi_preprocessed = 1;
 375         RETURN(rc);
 376 }
 377
 378 /*
 379  * Invoke handler for this request opc. Also do necessary preprocessing
 380  * (according to handler ->th_flags), and post-processing (setting of
 381  * ->last_{xid,committed}).
 382  */
 383 static int tgt_handle_request0(struct tgt_session_info *tsi,
 384                                struct tgt_handler *h,
 385                                struct ptlrpc_request *req)
 386 {
 387         int      serious = 0;
 388         int      rc;
 389         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 390
 391         ENTRY;
 392
 393
 394         /* When dealing with sec context requests, no export is associated yet,
 395          * because these requests are sent before *_CONNECT requests.
 396          * A NULL req->rq_export means the normal *_common_slice handlers will
 397          * not be called, because there is no reference to the target.
 398          * So deal with them by hand and jump directly to target_send_reply().
 399          */
 400         switch (opc) {
 401         case SEC_CTX_INIT:
 402         case SEC_CTX_INIT_CONT:
 403         case SEC_CTX_FINI:
 404                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 405                 GOTO(out, rc = 0);
 406         }
 407
 408         /*
 409          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 410          * to put same checks into handlers like mdt_close(), mdt_reint(),
 411          * etc., without talking to mdt authors first. Checking same thing
 412          * there again is useless and returning 0 error without packing reply
 413          * is buggy! Handlers either pack reply or return error.
 414          *
 415          * We return 0 here and do not send any reply in order to emulate
 416          * network failure. Do not send any reply in case any of NET related
 417          * fail_id has occured.
 418          */
 419         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 420                 RETURN(0);
 421         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 422                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 423                 RETURN(0);
 424
 425         /* drop OUT_UPDATE rpc */
 426         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == OUT_UPDATE &&
 427                      OBD_FAIL_CHECK(OBD_FAIL_OUT_UPDATE_DROP)))
 428                 RETURN(0);
 429
 430         rc = tgt_request_preprocess(tsi, h, req);
 431         /* pack reply if reply format is fixed */
 432         if (rc == 0 && h->th_flags & HAS_REPLY) {
 433                 /* Pack reply */
 434                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 435                                           RCL_SERVER))
 436                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 437                                              RCL_SERVER,
 438                                              tsi->tsi_mdt_body->mbo_eadatasize);
 439                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 440                                           RCL_SERVER))
 441                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 442                                              RCL_SERVER, 0);
 443                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 444                         req_capsule_set_size(tsi->tsi_pill,
 445                                              &RMF_ACL, RCL_SERVER,
 446                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 447
 448                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 449                                           RCL_SERVER)) {
 450                         struct niobuf_remote *remote_nb =
 451                                 req_capsule_client_get(tsi->tsi_pill,
 452                                                        &RMF_NIOBUF_REMOTE);
 453                         struct ost_body *body = tsi->tsi_ost_body;
 454
 455                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 456                                          RCL_SERVER,
 457                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 458                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 459                                          remote_nb[0].rnb_len : 0);
 460                 }
 461                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_FILE_ENCCTX,
 462                                           RCL_SERVER))
 463                         req_capsule_set_size(tsi->tsi_pill, &RMF_FILE_ENCCTX,
 464                                              RCL_SERVER, 0);
 465
 466                 rc = req_capsule_server_pack(tsi->tsi_pill);
 467         }
 468
 469         if (likely(rc == 0)) {
 470                 /*
 471                  * Process request, there can be two types of rc:
 472                  * 1) errors with msg unpack/pack, other failures outside the
 473                  * operation itself. This is counted as serious errors;
 474                  * 2) errors during fs operation, should be placed in rq_status
 475                  * only
 476                  */
 477                 rc = h->th_act(tsi);
 478                 if (!is_serious(rc) &&
 479                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 480                         DEBUG_REQ(D_ERROR, req,
 481                                   "%s: %s handler did not pack reply but returned no error",
 482                                   tgt_name(tsi->tsi_tgt), h->th_name);
 483                         LBUG();
 484                 }
 485                 serious = is_serious(rc);
 486                 rc = clear_serious(rc);
 487         } else {
 488                 serious = 1;
 489         }
 490
 491         req->rq_status = rc;
 492
 493         /*
 494          * ELDLM_* codes which > 0 should be in rq_status only as well as
 495          * all non-serious errors.
 496          */
 497         if (rc > 0 || !serious)
 498                 rc = 0;
 499
 500         LASSERT(current->journal_info == NULL);
 501
 502         if (likely(rc == 0 && req->rq_export))
 503                 target_committed_to_req(req);
 504
 505 out:
 506         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 507         RETURN(0);
 508 }
 509
 510 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 511                                        struct obd_device *obd, int *process)
 512 {
 513         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 514         case MDS_DISCONNECT:
 515         case OST_DISCONNECT:
 516         case OBD_IDX_READ:
 517                 *process = 1;
 518                 RETURN(0);
 519         case MDS_CLOSE:
 520         case MDS_SYNC: /* used in unmounting */
 521         case OBD_PING:
 522         case MDS_REINT:
 523         case OUT_UPDATE:
 524         case SEQ_QUERY:
 525         case FLD_QUERY:
 526         case FLD_READ:
 527         case LDLM_ENQUEUE:
 528         case OST_CREATE:
 529         case OST_DESTROY:
 530         case OST_PUNCH:
 531         case OST_SETATTR:
 532         case OST_SYNC:
 533         case OST_WRITE:
 534         case MDS_HSM_PROGRESS:
 535         case MDS_HSM_STATE_SET:
 536         case MDS_HSM_REQUEST:
 537         case OST_FALLOCATE:
 538                 *process = target_queue_recovery_request(req, obd);
 539                 RETURN(0);
 540
 541         default:
 542                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 543                 *process = -EAGAIN;
 544                 RETURN(0);
 545         }
 546 }
 547
 548 /*
 549  * Handle recovery. Return:
 550  *        +1: continue request processing;
 551  *       -ve: abort immediately with the given error code;
 552  *         0: send reply with error code in req->rq_status;
 553  */
 554 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 555 {
 556         ENTRY;
 557
 558         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 559         case MDS_CONNECT:
 560         case OST_CONNECT:
 561         case MGS_CONNECT:
 562         case SEC_CTX_INIT:
 563         case SEC_CTX_INIT_CONT:
 564         case SEC_CTX_FINI:
 565                 RETURN(+1);
 566         }
 567
 568         if (!req->rq_export->exp_obd->obd_replayable)
 569                 RETURN(+1);
 570
 571         /* sanity check: if the xid matches, the request must be marked as a
 572          * resent or replayed */
 573         if (req_can_reconstruct(req, NULL) == 1) {
 574                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 575                       (MSG_RESENT | MSG_REPLAY))) {
 576                         DEBUG_REQ(D_WARNING, req,
 577                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 578                                   req->rq_xid,
 579                                   lustre_msg_get_flags(req->rq_reqmsg));
 580                         req->rq_status = -ENOTCONN;
 581                         RETURN(-ENOTCONN);
 582                 }
 583         }
 584         /* else: note the opposite is not always true; a RESENT req after a
 585          * failover will usually not match the last_xid, since it was likely
 586          * never committed. A REPLAYed request will almost never match the
 587          * last xid, however it could for a committed, but still retained,
 588          * open. */
 589
 590         /* Check for aborted recovery... */
 591         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 592                 int rc;
 593                 int should_process;
 594
 595                 DEBUG_REQ(D_INFO, req, "Got new replay");
 596                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 597                                                  &should_process);
 598                 if (rc != 0 || !should_process)
 599                         RETURN(rc);
 600                 else if (should_process < 0) {
 601                         req->rq_status = should_process;
 602                         rc = ptlrpc_error(req);
 603                         RETURN(rc);
 604                 }
 605         }
 606         RETURN(+1);
 607 }
 608
 609 /* Initial check for request, it is validation mostly */
 610 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 611 {
 612         struct tgt_handler      *h;
 613         struct tgt_opc_slice    *s;
 614         struct lu_target        *tgt;
 615         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 616
 617         ENTRY;
 618
 619         tgt = class_exp2tgt(req->rq_export);
 620         if (unlikely(tgt == NULL)) {
 621                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 622                           class_exp2obd(req->rq_export)->obd_name);
 623                 RETURN(ERR_PTR(-EINVAL));
 624         }
 625
 626         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 627                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 628                         break;
 629
 630         /* opcode was not found in slice */
 631         if (unlikely(s->tos_hs == NULL)) {
 632                 static bool printed;
 633
 634                 /* don't spew error messages for unhandled RPCs */
 635                 if (!printed) {
 636                         CERROR("%s: no handler for opcode 0x%x from %s\n",
 637                                tgt_name(tgt), opc, libcfs_id2str(req->rq_peer));
 638                         printed = true;
 639                 }
 640                 RETURN(ERR_PTR(-ENOTSUPP));
 641         }
 642
 643         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 644         h = s->tos_hs + (opc - s->tos_opc_start);
 645         if (unlikely(h->th_opc == 0)) {
 646                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 647                 RETURN(ERR_PTR(-ENOTSUPP));
 648         }
 649
 650         RETURN(h);
 651 }
 652
 653 static int process_req_last_xid(struct ptlrpc_request *req)
 654 {
 655         __u64   last_xid;
 656         int rc = 0;
 657         struct obd_export *exp = req->rq_export;
 658         struct tg_export_data *ted = &exp->exp_target_data;
 659         bool need_lock = tgt_is_multimodrpcs_client(exp);
 660         ENTRY;
 661
 662         if (need_lock)
 663                 mutex_lock(&ted->ted_lcd_lock);
 664         /* check request's xid is consistent with export's last_xid */
 665         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 666         if (last_xid > exp->exp_last_xid)
 667                 exp->exp_last_xid = last_xid;
 668
 669         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 670                 /* Some request is allowed to be sent during replay,
 671                  * such as OUT update requests, FLD requests, so it
 672                  * is possible that replay requests has smaller XID
 673                  * than the exp_last_xid.
 674                  *
 675                  * Some non-replay requests may have smaller XID as
 676                  * well:
 677                  *
 678                  * - Client send a no_resend RPC, like statfs;
 679                  * - The RPC timedout (or some other error) on client,
 680                  *   then it's removed from the unreplied list;
 681                  * - Client send some other request to bump the
 682                  *   exp_last_xid on server;
 683                  * - The former RPC got chance to be processed;
 684                  */
 685                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 686                         rc = -EPROTO;
 687
 688                 DEBUG_REQ(D_WARNING, req,
 689                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 690                           req->rq_xid, exp->exp_last_xid, rc);
 691                 if (rc)
 692                         GOTO(out, rc);
 693         }
 694
 695         /* The "last_xid" is the minimum xid among unreplied requests,
 696          * if the request is from the previous connection, its xid can
 697          * still be larger than "exp_last_xid", then the above check of
 698          * xid is not enough to determine whether the request is delayed.
 699          *
 700          * For example, if some replay request was delayed and caused
 701          * timeout at client and the replay is restarted, the delayed
 702          * replay request will have the larger xid than "exp_last_xid"
 703          */
 704         if (req->rq_export->exp_conn_cnt >
 705             lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 706                 CDEBUG(D_RPCTRACE,
 707                        "Dropping request %llu from an old epoch %u/%u\n",
 708                        req->rq_xid,
 709                        lustre_msg_get_conn_cnt(req->rq_reqmsg),
 710                        req->rq_export->exp_conn_cnt);
 711                 req->rq_no_reply = 1;
 712                 GOTO(out, rc = -ESTALE);
 713         }
 714
 715         /* try to release in-memory reply data */
 716         if (tgt_is_multimodrpcs_client(exp)) {
 717                 tgt_handle_received_xid(exp, last_xid);
 718                 rc = tgt_handle_tag(req);
 719         }
 720
 721 out:
 722         if (need_lock)
 723                 mutex_unlock(&ted->ted_lcd_lock);
 724
 725         RETURN(rc);
 726 }
 727
 728 int tgt_request_handle(struct ptlrpc_request *req)
 729 {
 730         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 731
 732         struct lustre_msg       *msg = req->rq_reqmsg;
 733         struct tgt_handler      *h;
 734         struct lu_target        *tgt;
 735         int                      request_fail_id = 0;
 736         __u32                    opc = lustre_msg_get_opc(msg);
 737         struct obd_device       *obd;
 738         int                      rc;
 739         bool                     is_connect = false;
 740         ENTRY;
 741
 742         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 743                 if (cfs_fail_val == 0 &&
 744                     lustre_msg_get_opc(msg) != OBD_PING &&
 745                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 746                         cfs_fail_val = 1;
 747                         cfs_race_state = 0;
 748                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 749                 }
 750         }
 751
 752         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 753         tsi->tsi_pill = &req->rq_pill;
 754         tsi->tsi_env = req->rq_svc_thread->t_env;
 755
 756         /* if request has export then get handlers slice from corresponding
 757          * target, otherwise that should be connect operation */
 758         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 759             opc == MGS_CONNECT) {
 760                 is_connect = true;
 761                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 762                 rc = target_handle_connect(req);
 763                 if (rc != 0) {
 764                         rc = ptlrpc_error(req);
 765                         GOTO(out, rc);
 766                 }
 767                 /* recovery-small test 18c asks to drop connect reply */
 768                 if (unlikely(opc == OST_CONNECT &&
 769                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 770                         GOTO(out, rc = 0);
 771         }
 772
 773         if (unlikely(!class_connected_export(req->rq_export))) {
 774                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 775                     opc == SEC_CTX_FINI) {
 776                         /* sec context initialization has to be handled
 777                          * by hand in tgt_handle_request0() */
 778                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 779                         h = NULL;
 780                         GOTO(handle_recov, rc = 0);
 781                 }
 782                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 783                        opc, libcfs_id2str(req->rq_peer));
 784                 req->rq_status = -ENOTCONN;
 785                 rc = ptlrpc_error(req);
 786                 GOTO(out, rc);
 787         }
 788
 789         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 790         tsi->tsi_exp = req->rq_export;
 791         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 792                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 793         else
 794                 tsi->tsi_jobid = NULL;
 795
 796         if (tgt == NULL) {
 797                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 798                           class_exp2obd(req->rq_export)->obd_name);
 799                 req->rq_status = -EINVAL;
 800                 rc = ptlrpc_error(req);
 801                 GOTO(out, rc);
 802         }
 803
 804         /* Skip last_xid processing for the recovery thread, otherwise, the
 805          * last_xid on same request could be processed twice: first time when
 806          * processing the incoming request, second time when the request is
 807          * being processed by recovery thread. */
 808         obd = class_exp2obd(req->rq_export);
 809         if (is_connect) {
 810                 /* reset the exp_last_xid on each connection. */
 811                 req->rq_export->exp_last_xid = 0;
 812         } else if (obd->obd_recovery_data.trd_processing_task !=
 813                    current->pid) {
 814                 rc = process_req_last_xid(req);
 815                 if (rc) {
 816                         req->rq_status = rc;
 817                         rc = ptlrpc_error(req);
 818                         GOTO(out, rc);
 819                 }
 820         }
 821
 822         request_fail_id = tgt->lut_request_fail_id;
 823         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 824
 825         h = tgt_handler_find_check(req);
 826         if (IS_ERR(h)) {
 827                 req->rq_status = PTR_ERR(h);
 828                 rc = ptlrpc_error(req);
 829                 GOTO(out, rc);
 830         }
 831
 832         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 833                  h->th_opc, opc);
 834
 835         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 836              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 837                 GOTO(out, rc = 0);
 838
 839         rc = lustre_msg_check_version(msg, h->th_version);
 840         if (unlikely(rc)) {
 841                 DEBUG_REQ(D_ERROR, req,
 842                           "%s: drop malformed request version=%08x expect=%08x",
 843                           tgt_name(tgt), lustre_msg_get_version(msg),
 844                           h->th_version);
 845                 req->rq_status = -EINVAL;
 846                 rc = ptlrpc_error(req);
 847                 GOTO(out, rc);
 848         }
 849
 850 handle_recov:
 851         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 852         if (likely(rc == 1)) {
 853                 rc = tgt_handle_request0(tsi, h, req);
 854                 if (rc)
 855                         GOTO(out, rc);
 856         }
 857         EXIT;
 858 out:
 859         req_capsule_fini(tsi->tsi_pill);
 860         if (tsi->tsi_corpus != NULL) {
 861                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 862                 tsi->tsi_corpus = NULL;
 863         }
 864         return rc;
 865 }
 866 EXPORT_SYMBOL(tgt_request_handle);
 867
 868 /** Assign high priority operations to the request if needed. */
 869 int tgt_hpreq_handler(struct ptlrpc_request *req)
 870 {
 871         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 872         struct tgt_handler      *h;
 873         int                      rc;
 874
 875         ENTRY;
 876
 877         if (req->rq_export == NULL)
 878                 RETURN(0);
 879
 880         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 881         tsi->tsi_pill = &req->rq_pill;
 882         tsi->tsi_env = req->rq_svc_thread->t_env;
 883         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 884         tsi->tsi_exp = req->rq_export;
 885
 886         h = tgt_handler_find_check(req);
 887         if (IS_ERR(h)) {
 888                 rc = PTR_ERR(h);
 889                 RETURN(rc);
 890         }
 891
 892         rc = tgt_request_preprocess(tsi, h, req);
 893         if (unlikely(rc != 0))
 894                 RETURN(rc);
 895
 896         if (h->th_hp != NULL)
 897                 h->th_hp(tsi);
 898         RETURN(0);
 899 }
 900 EXPORT_SYMBOL(tgt_hpreq_handler);
 901
 902 void tgt_counter_incr(struct obd_export *exp, int opcode)
 903 {
 904         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 905         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 906                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 907 }
 908 EXPORT_SYMBOL(tgt_counter_incr);
 909
 910 /*
 911  * Unified target generic handlers.
 912  */
 913
 914 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 915 {
 916         struct lu_target *tgt = class_exp2tgt(exp);
 917         struct sptlrpc_flavor flvr;
 918         int rc = 0;
 919
 920         LASSERT(tgt);
 921         LASSERT(tgt->lut_obd);
 922         LASSERT(tgt->lut_slice);
 923
 924         /* always allow ECHO client */
 925         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 926                             LUSTRE_ECHO_NAME) == 0)) {
 927                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 928                 return 0;
 929         }
 930
 931         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 932                 read_lock(&tgt->lut_sptlrpc_lock);
 933                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 934                                              req->rq_sp_from,
 935                                              req->rq_peer.nid,
 936                                              &flvr);
 937                 read_unlock(&tgt->lut_sptlrpc_lock);
 938
 939                 spin_lock(&exp->exp_lock);
 940                 exp->exp_sp_peer = req->rq_sp_from;
 941                 exp->exp_flvr = flvr;
 942
 943                 /* when on mgs, if no restriction is set, or if the client
 944                  * NID is on the local node, allow any flavor
 945                  */
 946                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 947                             LUSTRE_MGS_NAME) == 0) &&
 948                     (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 949                      LNetIsPeerLocal(&exp->exp_connection->c_peer.nid)))
 950                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 951
 952                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 953                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 954                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 955                                "expect %x\n", tgt_name(tgt),
 956                                req->rq_flvr.sf_rpc,
 957                                libcfs_nid2str(req->rq_peer.nid),
 958                                exp->exp_flvr.sf_rpc);
 959                         rc = -EACCES;
 960                 }
 961                 spin_unlock(&exp->exp_lock);
 962         } else {
 963                 if (exp->exp_sp_peer != req->rq_sp_from) {
 964                         CERROR("%s: RPC source %s doesn't match %s\n",
 965                                tgt_name(tgt),
 966                                sptlrpc_part2name(req->rq_sp_from),
 967                                sptlrpc_part2name(exp->exp_sp_peer));
 968                         rc = -EACCES;
 969                 } else {
 970                         rc = sptlrpc_target_export_check(exp, req);
 971                 }
 972         }
 973
 974         return rc;
 975 }
 976
 977 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 978 {
 979         struct sptlrpc_rule_set  tmp_rset;
 980         int                      rc;
 981
 982         if (unlikely(tgt == NULL)) {
 983                 CERROR("No target passed\n");
 984                 return -EINVAL;
 985         }
 986
 987         sptlrpc_rule_set_init(&tmp_rset);
 988         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 989         if (rc) {
 990                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 991                        tgt_name(tgt), rc);
 992                 return rc;
 993         }
 994
 995         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 996
 997         write_lock(&tgt->lut_sptlrpc_lock);
 998         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 999         tgt->lut_sptlrpc_rset = tmp_rset;
1000         write_unlock(&tgt->lut_sptlrpc_lock);
1001
1002         return 0;
1003 }
1004 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
1005
1006 int tgt_connect(struct tgt_session_info *tsi)
1007 {
1008         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1009         struct obd_connect_data *reply;
1010         int                      rc;
1011
1012         ENTRY;
1013
1014         /* XXX: better to call this check right after getting new export but
1015          * before last_rcvd slot allocation to avoid server load upon insecure
1016          * connects. This is to be fixed after unifiyng all targets.
1017          */
1018         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1019         if (rc)
1020                 GOTO(out, rc);
1021
1022         /* To avoid exposing partially initialized connection flags, changes up
1023          * to this point have been staged in reply->ocd_connect_flags. Now that
1024          * connection handling has completed successfully, atomically update
1025          * the connect flags in the shared export data structure. LU-1623 */
1026         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1027         spin_lock(&tsi->tsi_exp->exp_lock);
1028         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1029         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1030                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1031                         reply->ocd_connect_flags2;
1032         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1033         spin_unlock(&tsi->tsi_exp->exp_lock);
1034
1035         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1036                    LUSTRE_MDT_NAME) == 0) {
1037                 rc = req_check_sepol(tsi->tsi_pill);
1038                 if (rc)
1039                         GOTO(out, rc);
1040
1041                 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1042                     reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT &&
1043                     tsi->tsi_pill->rc_req->rq_export) {
1044                         bool forbid_encrypt = true;
1045                         struct lu_nodemap *nm =
1046                          nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1047
1048                         if (!nm) {
1049                                 /* nodemap_get_from_exp returns NULL in case
1050                                  * nodemap is not active, so we do not forbid
1051                                  */
1052                                 forbid_encrypt = false;
1053                         } else if (!IS_ERR(nm)) {
1054                                 forbid_encrypt = nm->nmf_forbid_encryption;
1055                                 nodemap_putref(nm);
1056                         }
1057
1058                         if (forbid_encrypt)
1059                                 GOTO(out, rc = -EACCES);
1060                 }
1061         }
1062
1063         RETURN(0);
1064 out:
1065         obd_disconnect(class_export_get(tsi->tsi_exp));
1066         return rc;
1067 }
1068 EXPORT_SYMBOL(tgt_connect);
1069
1070 int tgt_disconnect(struct tgt_session_info *tsi)
1071 {
1072         int rc;
1073
1074         ENTRY;
1075
1076         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1077
1078         rc = target_handle_disconnect(tgt_ses_req(tsi));
1079         if (rc)
1080                 RETURN(err_serious(rc));
1081
1082         RETURN(rc);
1083 }
1084 EXPORT_SYMBOL(tgt_disconnect);
1085
1086 /*
1087  * Unified target OBD handlers
1088  */
1089 int tgt_obd_ping(struct tgt_session_info *tsi)
1090 {
1091         int rc;
1092
1093         ENTRY;
1094
1095         /* The target-specific part of OBD_PING request handling.
1096          * It controls Filter Modification Data (FMD) expiration each time
1097          * PING is received.
1098          *
1099          * Valid only for replayable targets, e.g. MDT and OFD
1100          */
1101         if (tsi->tsi_exp->exp_obd->obd_replayable)
1102                 tgt_fmd_expire(tsi->tsi_exp);
1103
1104         rc = req_capsule_server_pack(tsi->tsi_pill);
1105         if (rc)
1106                 RETURN(err_serious(rc));
1107
1108         RETURN(rc);
1109 }
1110 EXPORT_SYMBOL(tgt_obd_ping);
1111
1112 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1113 {
1114         return err_serious(-EOPNOTSUPP);
1115 }
1116
1117 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1118 {
1119         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1120         struct obd_export       *exp = req->rq_export;
1121         struct ptlrpc_bulk_desc *desc;
1122         int                      i;
1123         int                      rc;
1124         int                      pages = 0;
1125
1126         ENTRY;
1127
1128         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1129                 unsigned int offset;
1130
1131                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1132                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1133                                       PAGE_SIZE);
1134         }
1135
1136         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1137                                   PTLRPC_BULK_PUT_SOURCE,
1138                                     MDS_BULK_PORTAL,
1139                                     &ptlrpc_bulk_kiov_nopin_ops);
1140         if (desc == NULL)
1141                 RETURN(-ENOMEM);
1142
1143         for (i = 0; i < rdbuf->rb_nbufs; i++)
1144                 desc->bd_frag_ops->add_iov_frag(desc,
1145                                         rdbuf->rb_bufs[i].lb_buf,
1146                                         rdbuf->rb_bufs[i].lb_len);
1147
1148         rc = target_bulk_io(exp, desc);
1149         ptlrpc_free_bulk(desc);
1150         RETURN(rc);
1151 }
1152 EXPORT_SYMBOL(tgt_send_buffer);
1153
1154 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1155 {
1156         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1157         struct obd_export       *exp = req->rq_export;
1158         struct ptlrpc_bulk_desc *desc;
1159         int                      tmpcount;
1160         int                      tmpsize;
1161         int                      i;
1162         int                      rc;
1163
1164         ENTRY;
1165
1166         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1167                                     PTLRPC_BULK_PUT_SOURCE,
1168                                     MDS_BULK_PORTAL,
1169                                     &ptlrpc_bulk_kiov_pin_ops);
1170         if (desc == NULL)
1171                 RETURN(-ENOMEM);
1172
1173         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1174                 /* old client requires reply size in it's PAGE_SIZE,
1175                  * which is rdpg->rp_count */
1176                 nob = rdpg->rp_count;
1177
1178         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1179              i++, tmpcount -= tmpsize) {
1180                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1181                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1182                                                  tmpsize);
1183         }
1184
1185         LASSERT(desc->bd_nob == nob);
1186         rc = target_bulk_io(exp, desc);
1187         ptlrpc_free_bulk(desc);
1188         RETURN(rc);
1189 }
1190 EXPORT_SYMBOL(tgt_sendpage);
1191
1192 /*
1193  * OBD_IDX_READ handler
1194  */
1195 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1196 {
1197         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1198         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1199         struct idx_info         *req_ii, *rep_ii;
1200         int                      rc, i;
1201
1202         ENTRY;
1203
1204         memset(rdpg, 0, sizeof(*rdpg));
1205         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1206
1207         /* extract idx_info buffer from request & reply */
1208         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1209         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1210                 RETURN(err_serious(-EPROTO));
1211
1212         rc = req_capsule_server_pack(tsi->tsi_pill);
1213         if (rc)
1214                 RETURN(err_serious(rc));
1215
1216         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1217         if (rep_ii == NULL)
1218                 RETURN(err_serious(-EFAULT));
1219         rep_ii->ii_magic = IDX_INFO_MAGIC;
1220
1221         /* extract hash to start with */
1222         rdpg->rp_hash = req_ii->ii_hash_start;
1223
1224         /* extract requested attributes */
1225         rdpg->rp_attrs = req_ii->ii_attrs;
1226
1227         /* check that fid packed in request is valid and supported */
1228         if (!fid_is_sane(&req_ii->ii_fid))
1229                 RETURN(-EINVAL);
1230         rep_ii->ii_fid = req_ii->ii_fid;
1231
1232         /* copy flags */
1233         rep_ii->ii_flags = req_ii->ii_flags;
1234
1235         /* compute number of pages to allocate, ii_count is the number of 4KB
1236          * containers */
1237         if (req_ii->ii_count <= 0)
1238                 GOTO(out, rc = -EFAULT);
1239         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1240                                exp_max_brw_size(tsi->tsi_exp));
1241         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1242
1243         /* allocate pages to store the containers */
1244         OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1245         if (rdpg->rp_pages == NULL)
1246                 GOTO(out, rc = -ENOMEM);
1247         for (i = 0; i < rdpg->rp_npages; i++) {
1248                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1249                 if (rdpg->rp_pages[i] == NULL)
1250                         GOTO(out, rc = -ENOMEM);
1251         }
1252
1253         /* populate pages with key/record pairs */
1254         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1255         if (rc < 0)
1256                 GOTO(out, rc);
1257
1258         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1259                  "asked %d > %d\n", rc, rdpg->rp_count);
1260
1261         /* send pages to client */
1262         rc = tgt_sendpage(tsi, rdpg, rc);
1263         if (rc)
1264                 GOTO(out, rc);
1265         EXIT;
1266 out:
1267         if (rdpg->rp_pages) {
1268                 for (i = 0; i < rdpg->rp_npages; i++)
1269                         if (rdpg->rp_pages[i])
1270                                 __free_page(rdpg->rp_pages[i]);
1271                 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1272         }
1273         return rc;
1274 }
1275
1276 struct tgt_handler tgt_obd_handlers[] = {
1277 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1278 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1279 };
1280 EXPORT_SYMBOL(tgt_obd_handlers);
1281
1282 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1283              struct dt_object *obj, __u64 start, __u64 end)
1284 {
1285         int rc = 0;
1286
1287         ENTRY;
1288
1289         /* if no objid is specified, it means "sync whole filesystem" */
1290         if (obj == NULL) {
1291                 rc = dt_sync(env, tgt->lut_bottom);
1292         } else if (dt_version_get(env, obj) >
1293                    tgt->lut_obd->obd_last_committed) {
1294                 rc = dt_object_sync(env, obj, start, end);
1295         }
1296         atomic_inc(&tgt->lut_sync_count);
1297
1298         RETURN(rc);
1299 }
1300 EXPORT_SYMBOL(tgt_sync);
1301 /*
1302  * Unified target DLM handlers.
1303  */
1304
1305 /**
1306  * Unified target BAST
1307  *
1308  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1309  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1310  * but if it's cross-MDT lock, because remote object version is not set, a
1311  * filesystem sync is needed.
1312  *
1313  * \param lock server side lock
1314  * \param desc lock desc
1315  * \param data ldlm_cb_set_arg
1316  * \param flag  indicates whether this cancelling or blocking callback
1317  * \retval      0 on success
1318  * \retval      negative number on error
1319  */
1320 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1321                             void *data, int flag)
1322 {
1323         struct lu_env            env;
1324         struct lu_target        *tgt;
1325         struct dt_object        *obj = NULL;
1326         struct lu_fid            fid;
1327         int                      rc = 0;
1328
1329         ENTRY;
1330
1331         tgt = class_exp2tgt(lock->l_export);
1332
1333         if (unlikely(tgt == NULL)) {
1334                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1335                        class_exp2obd(lock->l_export)->obd_name);
1336                 RETURN(-EINVAL);
1337         }
1338
1339         if (flag == LDLM_CB_CANCELING &&
1340             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1341             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1342              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1343               ldlm_is_cbpending(lock))) &&
1344             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1345              lock->l_resource->lr_type == LDLM_EXTENT)) {
1346                 __u64 start = 0;
1347                 __u64 end = OBD_OBJECT_EOF;
1348
1349                 rc = lu_env_init(&env, LCT_DT_THREAD);
1350                 if (unlikely(rc != 0))
1351                         GOTO(err, rc);
1352
1353                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1354                                    tgt->lut_lsd.lsd_osd_index);
1355
1356                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1357                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1358                         if (IS_ERR(obj))
1359                                 GOTO(err_env, rc = PTR_ERR(obj));
1360
1361                         if (!dt_object_exists(obj))
1362                                 GOTO(err_put, rc = -ENOENT);
1363
1364                         start = lock->l_policy_data.l_extent.start;
1365                         end = lock->l_policy_data.l_extent.end;
1366                 }
1367
1368                 rc = tgt_sync(&env, tgt, obj, start, end);
1369                 if (rc < 0) {
1370                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1371                                "cancel: rc = %d\n",
1372                                tgt_name(tgt), PFID(&fid),
1373                                lock->l_policy_data.l_extent.start,
1374                                lock->l_policy_data.l_extent.end, rc);
1375                 }
1376 err_put:
1377                 if (obj != NULL)
1378                         dt_object_put(&env, obj);
1379 err_env:
1380                 lu_env_fini(&env);
1381         }
1382 err:
1383         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1384         RETURN(rc);
1385 }
1386
1387 static struct ldlm_callback_suite tgt_dlm_cbs = {
1388         .lcs_completion = ldlm_server_completion_ast,
1389         .lcs_blocking   = tgt_blocking_ast,
1390         .lcs_glimpse    = ldlm_server_glimpse_ast
1391 };
1392
1393 int tgt_enqueue(struct tgt_session_info *tsi)
1394 {
1395         struct ptlrpc_request *req = tgt_ses_req(tsi);
1396         int rc;
1397
1398         ENTRY;
1399         /*
1400          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1401          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1402          */
1403         LASSERT(tsi->tsi_dlm_req != NULL);
1404         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1405                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1406         if (rc)
1407                 RETURN(err_serious(rc));
1408
1409         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1410         case LUT_FAIL_MDT:
1411                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1412                 break;
1413         case LUT_FAIL_OST:
1414                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1415                 break;
1416         case LUT_FAIL_MGT:
1417                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1418                 break;
1419         default:
1420                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1421                 break;
1422         }
1423         RETURN(req->rq_status);
1424 }
1425 EXPORT_SYMBOL(tgt_enqueue);
1426
1427 int tgt_convert(struct tgt_session_info *tsi)
1428 {
1429         struct ptlrpc_request *req = tgt_ses_req(tsi);
1430         int rc;
1431
1432         ENTRY;
1433         LASSERT(tsi->tsi_dlm_req);
1434         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1435         if (rc)
1436                 RETURN(err_serious(rc));
1437
1438         RETURN(req->rq_status);
1439 }
1440
1441 int tgt_bl_callback(struct tgt_session_info *tsi)
1442 {
1443         return err_serious(-EOPNOTSUPP);
1444 }
1445
1446 int tgt_cp_callback(struct tgt_session_info *tsi)
1447 {
1448         return err_serious(-EOPNOTSUPP);
1449 }
1450
1451 /* generic LDLM target handler */
1452 struct tgt_handler tgt_dlm_handlers[] = {
1453 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1454 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1455 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1456 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1457 };
1458 EXPORT_SYMBOL(tgt_dlm_handlers);
1459
1460 /*
1461  * Unified target LLOG handlers.
1462  */
1463 int tgt_llog_open(struct tgt_session_info *tsi)
1464 {
1465         int rc;
1466
1467         ENTRY;
1468
1469         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1470
1471         RETURN(rc);
1472 }
1473 EXPORT_SYMBOL(tgt_llog_open);
1474
1475 int tgt_llog_read_header(struct tgt_session_info *tsi)
1476 {
1477         int rc;
1478
1479         ENTRY;
1480
1481         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1482
1483         RETURN(rc);
1484 }
1485 EXPORT_SYMBOL(tgt_llog_read_header);
1486
1487 int tgt_llog_next_block(struct tgt_session_info *tsi)
1488 {
1489         int rc;
1490
1491         ENTRY;
1492
1493         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1494
1495         RETURN(rc);
1496 }
1497 EXPORT_SYMBOL(tgt_llog_next_block);
1498
1499 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1500 {
1501         int rc;
1502
1503         ENTRY;
1504
1505         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1506
1507         RETURN(rc);
1508 }
1509 EXPORT_SYMBOL(tgt_llog_prev_block);
1510
1511 /* generic llog target handler */
1512 struct tgt_handler tgt_llog_handlers[] = {
1513 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1514 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1515 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1516 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1517 };
1518 EXPORT_SYMBOL(tgt_llog_handlers);
1519
1520 /*
1521  * sec context handlers
1522  */
1523 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1524 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1525 {
1526         return 0;
1527 }
1528
1529 struct tgt_handler tgt_sec_ctx_handlers[] = {
1530 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1531 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1532 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1533 };
1534 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1535
1536 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1537                                  struct dt_device *key,
1538                                  struct lfsck_req_local *lrl,
1539                                  struct thandle *th) = NULL;
1540
1541 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1542                                                       struct dt_device *,
1543                                                       struct lfsck_req_local *,
1544                                                       struct thandle *))
1545 {
1546         tgt_lfsck_in_notify_local = notify;
1547 }
1548 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1549
1550 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1551                            struct dt_device *key,
1552                            struct lfsck_request *lr) = NULL;
1553
1554 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1555                                                 struct dt_device *,
1556                                                 struct lfsck_request *))
1557 {
1558         tgt_lfsck_in_notify = notify;
1559 }
1560 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1561
1562 static int (*tgt_lfsck_query)(const struct lu_env *env,
1563                               struct dt_device *key,
1564                               struct lfsck_request *req,
1565                               struct lfsck_reply *rep,
1566                               struct lfsck_query *que) = NULL;
1567
1568 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1569                                            struct dt_device *,
1570                                            struct lfsck_request *,
1571                                            struct lfsck_reply *,
1572                                            struct lfsck_query *))
1573 {
1574         tgt_lfsck_query = query;
1575 }
1576 EXPORT_SYMBOL(tgt_register_lfsck_query);
1577
1578 /* LFSCK request handlers */
1579 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1580 {
1581         const struct lu_env     *env = tsi->tsi_env;
1582         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1583         struct lfsck_request    *lr;
1584         int                      rc;
1585         ENTRY;
1586
1587         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1588         if (lr == NULL)
1589                 RETURN(-EPROTO);
1590
1591         rc = tgt_lfsck_in_notify(env, key, lr);
1592
1593         RETURN(rc);
1594 }
1595
1596 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1597 {
1598         struct lfsck_request    *request;
1599         struct lfsck_reply      *reply;
1600         int                      rc;
1601         ENTRY;
1602
1603         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1604         if (request == NULL)
1605                 RETURN(-EPROTO);
1606
1607         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1608         if (reply == NULL)
1609                 RETURN(-ENOMEM);
1610
1611         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1612                              request, reply, NULL);
1613
1614         RETURN(rc < 0 ? rc : 0);
1615 }
1616
1617 struct tgt_handler tgt_lfsck_handlers[] = {
1618 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1619 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1620 };
1621 EXPORT_SYMBOL(tgt_lfsck_handlers);
1622
1623 /*
1624  * initialize per-thread page pool (bug 5137).
1625  */
1626 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1627 {
1628         struct tgt_thread_big_cache *tbc;
1629
1630         ENTRY;
1631
1632         LASSERT(thread != NULL);
1633         LASSERT(thread->t_data == NULL);
1634
1635         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1636         if (tbc == NULL)
1637                 RETURN(-ENOMEM);
1638         thread->t_data = tbc;
1639         RETURN(0);
1640 }
1641 EXPORT_SYMBOL(tgt_io_thread_init);
1642
1643 /*
1644  * free per-thread pool created by tgt_thread_init().
1645  */
1646 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1647 {
1648         struct tgt_thread_big_cache *tbc;
1649
1650         ENTRY;
1651
1652         LASSERT(thread != NULL);
1653
1654         /*
1655          * be prepared to handle partially-initialized pools (because this is
1656          * called from ost_io_thread_init() for cleanup.
1657          */
1658         tbc = thread->t_data;
1659         if (tbc != NULL) {
1660                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1661                 thread->t_data = NULL;
1662         }
1663         EXIT;
1664 }
1665 EXPORT_SYMBOL(tgt_io_thread_done);
1666
1667 /**
1668  * Helper function for getting Data-on-MDT file server DLM lock
1669  * if asked by client.
1670  */
1671 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1672                       struct lustre_handle *lh, int mode, __u64 *flags)
1673 {
1674         union ldlm_policy_data policy = {
1675                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1676         };
1677         int rc;
1678
1679         ENTRY;
1680
1681         LASSERT(lh != NULL);
1682         LASSERT(ns != NULL);
1683         LASSERT(!lustre_handle_is_used(lh));
1684
1685         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1686                                     flags, ldlm_blocking_ast,
1687                                     ldlm_completion_ast, ldlm_glimpse_ast,
1688                                     NULL, 0, LVB_T_NONE, NULL, lh);
1689
1690         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1691 }
1692 EXPORT_SYMBOL(tgt_mdt_data_lock);
1693
1694 /**
1695  * Helper function for getting server side [start, start+count] DLM lock
1696  * if asked by client.
1697  */
1698 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1699                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1700                     struct lustre_handle *lh, int mode, __u64 *flags)
1701 {
1702         union ldlm_policy_data policy;
1703         int rc;
1704
1705         ENTRY;
1706
1707         LASSERT(lh != NULL);
1708         LASSERT(ns != NULL);
1709         LASSERT(!lustre_handle_is_used(lh));
1710
1711         policy.l_extent.gid = 0;
1712         policy.l_extent.start = start & PAGE_MASK;
1713
1714         /*
1715          * If ->o_blocks is EOF it means "lock till the end of the file".
1716          * Otherwise, it's size of an extent or hole being punched (in bytes).
1717          */
1718         if (end == OBD_OBJECT_EOF || end < start)
1719                 policy.l_extent.end = OBD_OBJECT_EOF;
1720         else
1721                 policy.l_extent.end = end | ~PAGE_MASK;
1722
1723         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1724                                     mode, flags, ldlm_blocking_ast,
1725                                     ldlm_completion_ast, ldlm_glimpse_ast,
1726                                     NULL, 0, LVB_T_NONE, NULL, lh);
1727         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1728 }
1729 EXPORT_SYMBOL(tgt_extent_lock);
1730
1731 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1732                          struct ldlm_res_id *res_id, __u64 start, __u64 end,
1733                          struct lustre_handle *lh, enum ldlm_mode mode)
1734 {
1735         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1736         __u64 flags = 0;
1737
1738         /* MDT IO for data-on-mdt */
1739         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1740                 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1741
1742         return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1743 }
1744
1745 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1746 {
1747         LASSERT(lustre_handle_is_used(lh));
1748         ldlm_lock_decref(lh, mode);
1749 }
1750 EXPORT_SYMBOL(tgt_data_unlock);
1751
1752 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1753                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1754                         struct niobuf_remote *nb, struct lustre_handle *lh,
1755                         enum ldlm_mode mode)
1756 {
1757         int nrbufs = obj->ioo_bufcnt;
1758         int i;
1759
1760         ENTRY;
1761
1762         LASSERT(mode == LCK_PR || mode == LCK_PW);
1763         LASSERT(!lustre_handle_is_used(lh));
1764
1765         if (exp->exp_obd->obd_recovering)
1766                 RETURN(0);
1767
1768         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1769                 RETURN(0);
1770
1771         for (i = 1; i < nrbufs; i++)
1772                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1773                         RETURN(-EFAULT);
1774
1775         return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1776                              nb[nrbufs - 1].rnb_offset +
1777                              nb[nrbufs - 1].rnb_len - 1, lh, mode);
1778 }
1779
1780 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1781                            struct niobuf_remote *niob,
1782                            struct lustre_handle *lh, enum ldlm_mode mode)
1783 {
1784         ENTRY;
1785
1786         LASSERT(mode == LCK_PR || mode == LCK_PW);
1787         LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1788                  niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1789                 lustre_handle_is_used(lh));
1790
1791         if (lustre_handle_is_used(lh))
1792                 tgt_data_unlock(lh, mode);
1793         EXIT;
1794 }
1795
1796 static int tgt_checksum_niobuf(struct lu_target *tgt,
1797                                  struct niobuf_local *local_nb, int npages,
1798                                  int opc, enum cksum_types cksum_type,
1799                                  __u32 *cksum)
1800 {
1801         struct ahash_request           *req;
1802         unsigned int                    bufsize;
1803         int                             i, err;
1804         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1805
1806         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1807         if (IS_ERR(req)) {
1808                 CERROR("%s: unable to initialize checksum hash %s\n",
1809                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1810                 return PTR_ERR(req);
1811         }
1812
1813         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1814         for (i = 0; i < npages; i++) {
1815                 /* corrupt the data before we compute the checksum, to
1816                  * simulate a client->OST data error */
1817                 if (i == 0 && opc == OST_WRITE &&
1818                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1819                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1820                         int len = local_nb[i].lnb_len;
1821                         struct page *np = tgt_page_to_corrupt;
1822
1823                         if (np) {
1824                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1825                                 char *ptr2 = page_address(np);
1826
1827                                 memcpy(ptr2 + off, ptr + off, len);
1828                                 memcpy(ptr2 + off, "bad3", min(4, len));
1829                                 kunmap_atomic(ptr);
1830
1831                                 /* LU-8376 to preserve original index for
1832                                  * display in dump_all_bulk_pages() */
1833                                 np->index = i;
1834
1835                                 cfs_crypto_hash_update_page(req, np, off,
1836                                                             len);
1837                                 continue;
1838                         } else {
1839                                 CERROR("%s: can't alloc page for corruption\n",
1840                                        tgt_name(tgt));
1841                         }
1842                 }
1843                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1844                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1845                                   local_nb[i].lnb_len);
1846
1847                  /* corrupt the data after we compute the checksum, to
1848                  * simulate an OST->client data error */
1849                 if (i == 0 && opc == OST_READ &&
1850                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1851                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1852                         int len = local_nb[i].lnb_len;
1853                         struct page *np = tgt_page_to_corrupt;
1854
1855                         if (np) {
1856                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1857                                 char *ptr2 = page_address(np);
1858
1859                                 memcpy(ptr2 + off, ptr + off, len);
1860                                 memcpy(ptr2 + off, "bad4", min(4, len));
1861                                 kunmap_atomic(ptr);
1862
1863                                 /* LU-8376 to preserve original index for
1864                                  * display in dump_all_bulk_pages() */
1865                                 np->index = i;
1866
1867                                 cfs_crypto_hash_update_page(req, np, off,
1868                                                             len);
1869                                 continue;
1870                         } else {
1871                                 CERROR("%s: can't alloc page for corruption\n",
1872                                        tgt_name(tgt));
1873                         }
1874                 }
1875         }
1876
1877         bufsize = sizeof(*cksum);
1878         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1879
1880         return 0;
1881 }
1882
1883 char dbgcksum_file_name[PATH_MAX];
1884
1885 static void dump_all_bulk_pages(struct obdo *oa, int count,
1886                                 struct niobuf_local *local_nb,
1887                                 __u32 server_cksum, __u32 client_cksum)
1888 {
1889         struct file *filp;
1890         int rc, i;
1891         unsigned int len;
1892         char *buf;
1893
1894         /* will only keep dump of pages on first error for the same range in
1895          * file/fid, not during the resends/retries. */
1896         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1897                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1898                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1899                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1900                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1901                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1902                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1903                  local_nb[0].lnb_file_offset,
1904                  local_nb[count-1].lnb_file_offset +
1905                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1906         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1907         filp = filp_open(dbgcksum_file_name,
1908                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1909         if (IS_ERR(filp)) {
1910                 rc = PTR_ERR(filp);
1911                 if (rc == -EEXIST)
1912                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1913                                "checksum error: rc = %d\n", dbgcksum_file_name,
1914                                rc);
1915                 else
1916                         CERROR("%s: can't open to dump pages with checksum "
1917                                "error: rc = %d\n", dbgcksum_file_name, rc);
1918                 return;
1919         }
1920
1921         for (i = 0; i < count; i++) {
1922                 len = local_nb[i].lnb_len;
1923                 buf = kmap(local_nb[i].lnb_page);
1924                 while (len != 0) {
1925                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1926                         if (rc < 0) {
1927                                 CERROR("%s: wanted to write %u but got %d "
1928                                        "error\n", dbgcksum_file_name, len, rc);
1929                                 break;
1930                         }
1931                         len -= rc;
1932                         buf += rc;
1933                 }
1934                 kunmap(local_nb[i].lnb_page);
1935         }
1936
1937         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1938         if (rc)
1939                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1940         filp_close(filp, NULL);
1941
1942         libcfs_debug_dumplog();
1943 }
1944
1945 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1946                                struct obd_export *exp, struct obdo *oa,
1947                                const struct lnet_process_id *peer,
1948                                __u32 client_cksum, __u32 server_cksum,
1949                                enum cksum_types server_cksum_type)
1950 {
1951         char *msg;
1952         enum cksum_types cksum_type;
1953         loff_t start, end;
1954
1955         if (unlikely(npages <= 0))
1956                 return 0;
1957
1958         /* unlikely to happen and only if resend does not occur due to cksum
1959          * control failure on Client */
1960         if (unlikely(server_cksum == client_cksum)) {
1961                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1962                        client_cksum);
1963                 return 0;
1964         }
1965
1966         if (exp->exp_obd->obd_checksum_dump)
1967                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1968                                     client_cksum);
1969
1970         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1971                                            oa->o_flags : 0);
1972
1973         if (cksum_type != server_cksum_type)
1974                 msg = "the server may have not used the checksum type specified"
1975                       " in the original request - likely a protocol problem";
1976         else
1977                 msg = "should have changed on the client or in transit";
1978
1979         start = local_nb[0].lnb_file_offset;
1980         end = local_nb[npages-1].lnb_file_offset +
1981                                         local_nb[npages-1].lnb_len - 1;
1982
1983         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1984                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1985                 " %x (type %x), server csum %x (type %x)\n",
1986                 exp->exp_obd->obd_name,
1987                 msg, libcfs_nid2str(peer->nid),
1988                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1989                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1990                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1991                 POSTID(&oa->o_oi),
1992                 start, end, client_cksum, cksum_type, server_cksum,
1993                 server_cksum_type);
1994
1995         return 1;
1996 }
1997
1998 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1999                              unsigned char *buf, int size)
2000 {
2001         int     i, off, len, copied = size;
2002         char    *ptr;
2003
2004         for (i = 0; i < npages; i++) {
2005                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2006                 len = local[i].lnb_len;
2007
2008                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2009                        i, off, len, size);
2010                 if (len > size)
2011                         return -EINVAL;
2012
2013                 ptr = kmap_atomic(local[i].lnb_page);
2014                 memcpy(buf, ptr + off, len);
2015                 kunmap_atomic(ptr);
2016                 buf += len;
2017                 size -= len;
2018         }
2019         return copied - size;
2020 }
2021
2022 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2023                                      struct niobuf_local *local_nb, int npages,
2024                                      int opc, obd_dif_csum_fn *fn,
2025                                      int sector_size, u32 *check_sum,
2026                                      bool resend)
2027 {
2028         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2029         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2030         const char *obd_name = tgt->lut_obd->obd_name;
2031         struct ahash_request *req;
2032         unsigned int bufsize;
2033         unsigned char *buffer;
2034         struct page *__page;
2035         __u16 *guard_start;
2036         int guard_number;
2037         int used_number = 0;
2038         __u32 cksum;
2039         int rc = 0;
2040         int used;
2041         int i;
2042
2043         __page = alloc_page(GFP_KERNEL);
2044         if (__page == NULL)
2045                 return -ENOMEM;
2046
2047         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2048         if (IS_ERR(req)) {
2049                 rc = PTR_ERR(req);
2050                 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
2051                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg), rc);
2052                 goto out;
2053         }
2054
2055         buffer = kmap(__page);
2056         guard_start = (__u16 *)buffer;
2057         guard_number = PAGE_SIZE / sizeof(*guard_start);
2058         if (unlikely(resend))
2059                 CDEBUG(D_PAGE | D_HA, "GRD tags per page = %u\n", guard_number);
2060         for (i = 0; i < npages; i++) {
2061                 bool use_t10_grd;
2062
2063                 /* corrupt the data before we compute the checksum, to
2064                  * simulate a client->OST data error */
2065                 if (i == 0 && opc == OST_WRITE &&
2066                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
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 = kmap_atomic(local_nb[i].lnb_page);
2073                                 char *ptr2 = page_address(np);
2074
2075                                 memcpy(ptr2 + off, ptr + off, len);
2076                                 memcpy(ptr2 + off, "bad3", min(4, len));
2077                                 kunmap_atomic(ptr);
2078
2079                                 /* LU-8376 to preserve original index for
2080                                  * display in dump_all_bulk_pages() */
2081                                 np->index = i;
2082
2083                                 cfs_crypto_hash_update_page(req, np, off,
2084                                                             len);
2085                                 continue;
2086                         } else {
2087                                 CERROR("%s: can't alloc page for corruption\n",
2088                                        tgt_name(tgt));
2089                         }
2090                 }
2091
2092                 /*
2093                  * The left guard number should be able to hold checksums of a
2094                  * whole page
2095                  */
2096                 use_t10_grd = t10_cksum_type && opc == OST_READ &&
2097                               local_nb[i].lnb_len == PAGE_SIZE &&
2098                               local_nb[i].lnb_guard_disk;
2099                 if (use_t10_grd) {
2100                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2101                         if (used > (guard_number - used_number)) {
2102                                 rc = -E2BIG;
2103                                 break;
2104                         }
2105                         memcpy(guard_start + used_number,
2106                                local_nb[i].lnb_guards,
2107                                used * sizeof(*guard_start));
2108                         if (unlikely(resend))
2109                                 CDEBUG(D_PAGE | D_HA,
2110                                        "lnb[%u]: used %u off %u+%u lnb checksum: %*phN\n",
2111                                        i, used,
2112                                        local_nb[i].lnb_page_offset,
2113                                        local_nb[i].lnb_len,
2114                                        (int)(used * sizeof(*guard_start)),
2115                                        guard_start + used_number);
2116                 }
2117                 if (!use_t10_grd || unlikely(resend)) {
2118                         __u16 guard_tmp[MAX_GUARD_NUMBER];
2119                         __u16 *guards = guard_start + used_number;
2120                         int used_tmp = -1, *usedp = &used;
2121
2122                         if (unlikely(use_t10_grd)) {
2123                                 guards = guard_tmp;
2124                                 usedp = &used_tmp;
2125                         }
2126                         rc = obd_page_dif_generate_buffer(obd_name,
2127                                 local_nb[i].lnb_page,
2128                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2129                                 local_nb[i].lnb_len, guards,
2130                                 guard_number - used_number, usedp, sector_size,
2131                                 fn);
2132                         if (unlikely(resend)) {
2133                                 bool bad = use_t10_grd &&
2134                                         memcmp(guard_tmp,
2135                                                local_nb[i].lnb_guards,
2136                                                used_tmp * sizeof(*guard_tmp));
2137
2138                                 if (bad)
2139                                         CERROR("lnb[%u]: used %u/%u off %u+%u tmp checksum: %*phN\n",
2140                                                i, used, used_tmp,
2141                                                local_nb[i].lnb_page_offset,
2142                                                local_nb[i].lnb_len,
2143                                                (int)(used_tmp * sizeof(*guard_start)),
2144                                                guard_tmp);
2145                                 CDEBUG_LIMIT(D_PAGE | D_HA | (bad ? D_ERROR : 0),
2146                                        "lnb[%u]: used %u/%u off %u+%u gen checksum: %*phN\n",
2147                                        i, used, used_tmp,
2148                                        local_nb[i].lnb_page_offset,
2149                                        local_nb[i].lnb_len,
2150                                        (int)(used * sizeof(*guard_start)),
2151                                        guard_start + used_number);
2152                         }
2153                         if (rc)
2154                                 break;
2155                 }
2156
2157                 LASSERT(used <= MAX_GUARD_NUMBER);
2158                 /*
2159                  * If disk support T10PI checksum, copy guards to local_nb.
2160                  * If the write is partial page, do not use the guards for bio
2161                  * submission since the data might not be full-sector. The bio
2162                  * guards will be generated later based on the full sectors. If
2163                  * the sector size is 512B rather than 4 KB, or the page size
2164                  * is larger than 4KB, this might drop some useful guards for
2165                  * partial page write, but it will only add minimal extra time
2166                  * of checksum calculation.
2167                  */
2168                 if (t10_cksum_type && opc == OST_WRITE &&
2169                     local_nb[i].lnb_len == PAGE_SIZE) {
2170                         local_nb[i].lnb_guard_rpc = 1;
2171                         memcpy(local_nb[i].lnb_guards,
2172                                guard_start + used_number,
2173                                used * sizeof(*local_nb[i].lnb_guards));
2174                 }
2175
2176                 used_number += used;
2177                 if (used_number == guard_number) {
2178                         cfs_crypto_hash_update_page(req, __page, 0,
2179                                 used_number * sizeof(*guard_start));
2180                         used_number = 0;
2181                 }
2182
2183                  /* corrupt the data after we compute the checksum, to
2184                  * simulate an OST->client data error */
2185                 if (unlikely(i == 0 && opc == OST_READ &&
2186                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2187                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2188                         int len = local_nb[i].lnb_len;
2189                         struct page *np = tgt_page_to_corrupt;
2190
2191                         if (np) {
2192                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2193                                 char *ptr2 = page_address(np);
2194
2195                                 memcpy(ptr2 + off, ptr + off, len);
2196                                 memcpy(ptr2 + off, "bad4", min(4, len));
2197                                 kunmap_atomic(ptr);
2198
2199                                 /* LU-8376 to preserve original index for
2200                                  * display in dump_all_bulk_pages() */
2201                                 np->index = i;
2202
2203                                 cfs_crypto_hash_update_page(req, np, off,
2204                                                             len);
2205                                 continue;
2206                         } else {
2207                                 CERROR("%s: can't alloc page for corruption\n",
2208                                        tgt_name(tgt));
2209                         }
2210                 }
2211         }
2212         kunmap(__page);
2213         if (rc)
2214                 GOTO(out, rc);
2215
2216         if (used_number != 0)
2217                 cfs_crypto_hash_update_page(req, __page, 0,
2218                         used_number * sizeof(*guard_start));
2219
2220         bufsize = sizeof(cksum);
2221         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2222
2223         if (rc == 0)
2224                 *check_sum = cksum;
2225 out:
2226         __free_page(__page);
2227         return rc;
2228 }
2229
2230 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2231                                   enum cksum_types cksum_type,
2232                                   struct niobuf_local *local_nb,
2233                                   int npages, int opc, u32 *check_sum,
2234                                   bool resend)
2235 {
2236         obd_dif_csum_fn *fn = NULL;
2237         int sector_size = 0;
2238         int rc;
2239
2240         ENTRY;
2241         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2242
2243         if (fn)
2244                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2245                                                opc, fn, sector_size,
2246                                                check_sum, resend);
2247         else
2248                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2249                                          cksum_type, check_sum);
2250
2251         RETURN(rc);
2252 }
2253
2254 int tgt_brw_read(struct tgt_session_info *tsi)
2255 {
2256         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2257         struct ptlrpc_bulk_desc *desc = NULL;
2258         struct obd_export       *exp = tsi->tsi_exp;
2259         struct niobuf_remote    *remote_nb;
2260         struct niobuf_local     *local_nb;
2261         struct obd_ioobj        *ioo;
2262         struct ost_body         *body, *repbody;
2263         struct lustre_handle     lockh = { 0 };
2264         int                      npages, nob = 0, rc, i, no_reply = 0,
2265                                  npages_read;
2266         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2267         const char *obd_name = exp->exp_obd->obd_name;
2268         ktime_t kstart;
2269
2270         ENTRY;
2271
2272         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2273             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2274                 CERROR("%s: deny read request from %s to portal %u\n",
2275                        tgt_name(tsi->tsi_tgt),
2276                        obd_export_nid2str(req->rq_export),
2277                        ptlrpc_req2svc(req)->srv_req_portal);
2278                 RETURN(-EPROTO);
2279         }
2280
2281         req->rq_bulk_read = 1;
2282
2283         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) {
2284                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2285                  * this server to fail requests.
2286                  */
2287                 char fail_ost_name[MAX_OBD_NAME];
2288
2289                 if (cfs_fail_val > 0) {
2290                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2291                                  cfs_fail_val - 1);
2292
2293                         if (strstr(obd_name, fail_ost_name))
2294                                 RETURN(err_serious(-EIO));
2295                 } else {
2296                         RETURN(err_serious(-EIO));
2297                 }
2298         }
2299
2300         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2301                          cfs_fail_val : (obd_timeout + 1) / 4);
2302
2303         /* Check if there is eviction in progress, and if so, wait for it to
2304          * finish */
2305         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2306                 /* We do not care how long it takes */
2307                 wait_event_idle(
2308                         exp->exp_obd->obd_evict_inprogress_waitq,
2309                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2310         }
2311
2312         /* There must be big cache in current thread to process this request
2313          * if it is NULL then something went wrong and it wasn't allocated,
2314          * report -ENOMEM in that case */
2315         if (tbc == NULL)
2316                 RETURN(-ENOMEM);
2317
2318         body = tsi->tsi_ost_body;
2319         LASSERT(body != NULL);
2320
2321         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2322             body->oa.o_flags & OBD_FL_NORPC)
2323                 RETURN(0);
2324
2325         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2326         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2327
2328         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2329         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2330
2331         local_nb = tbc->local;
2332
2333         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2334                           &lockh, LCK_PR);
2335         if (rc != 0)
2336                 RETURN(rc);
2337
2338         /*
2339          * If getting the lock took more time than
2340          * client was willing to wait, drop it. b=11330
2341          */
2342         if (ktime_get_real_seconds() > req->rq_deadline ||
2343             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2344                 no_reply = 1;
2345                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2346                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2347                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2348                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2349                 GOTO(out_lock, rc = -ETIMEDOUT);
2350         }
2351
2352         /*
2353          * Because we already sync grant info with client when
2354          * reconnect, grant info will be cleared for resent req,
2355          * otherwise, outdated grant count in the rpc would de-sync
2356          * grant counters in case of shrink
2357          */
2358         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2359                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2360                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2361         }
2362
2363         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2364         repbody->oa = body->oa;
2365
2366         npages = PTLRPC_MAX_BRW_PAGES;
2367         kstart = ktime_get();
2368         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2369                         ioo, remote_nb, &npages, local_nb);
2370         if (rc != 0)
2371                 GOTO(out_lock, rc);
2372
2373         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2374             body->oa.o_flags & OBD_FL_SHORT_IO) {
2375                 desc = NULL;
2376         } else {
2377                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2378                                             PTLRPC_BULK_PUT_SOURCE,
2379                                             OST_BULK_PORTAL,
2380                                             &ptlrpc_bulk_kiov_nopin_ops);
2381                 if (desc == NULL)
2382                         GOTO(out_commitrw, rc = -ENOMEM);
2383         }
2384
2385         npages_read = npages;
2386         for (i = 0; i < npages; i++) {
2387                 int page_rc = local_nb[i].lnb_rc;
2388
2389                 if (page_rc < 0) {
2390                         rc = page_rc;
2391                         npages_read = i;
2392                         break;
2393                 }
2394
2395                 nob += page_rc;
2396                 if (page_rc != 0 && desc != NULL) { /* some data! */
2397                         LASSERT(local_nb[i].lnb_page != NULL);
2398                         desc->bd_frag_ops->add_kiov_frag
2399                           (desc, local_nb[i].lnb_page,
2400                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2401                            page_rc);
2402                 }
2403
2404                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2405                         local_nb[i].lnb_len = page_rc;
2406                         npages_read = i + (page_rc != 0 ? 1 : 0);
2407                         /* All subsequent pages should be 0 */
2408                         while (++i < npages)
2409                                 LASSERT(local_nb[i].lnb_rc == 0);
2410                         break;
2411                 }
2412         }
2413
2414         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2415                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2416                            body->oa.o_flags : 0;
2417                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2418                 bool resend = (body->oa.o_valid & OBD_MD_FLFLAGS) &&
2419                         (body->oa.o_flags & OBD_FL_RECOV_RESEND);
2420
2421                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2422                                                           cksum_type);
2423                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2424
2425                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2426                                             local_nb, npages_read, OST_READ,
2427                                             &repbody->oa.o_cksum, resend);
2428                 if (rc < 0)
2429                         GOTO(out_commitrw, rc);
2430                 CDEBUG(D_PAGE | (resend ? D_HA : 0),
2431                        "checksum at read origin: %x (%x)\n",
2432                        repbody->oa.o_cksum, cksum_type);
2433
2434                 /* if a resend it could be for a cksum error, so check Server
2435                  * cksum with returned Client cksum (this should even cover
2436                  * zero-cksum case) */
2437                 if (resend)
2438                         check_read_checksum(local_nb, npages_read, exp,
2439                                             &body->oa, &req->rq_peer,
2440                                             body->oa.o_cksum,
2441                                             repbody->oa.o_cksum, cksum_type);
2442         } else {
2443                 repbody->oa.o_valid = 0;
2444         }
2445         if (body->oa.o_valid & OBD_MD_FLGRANT)
2446                 repbody->oa.o_valid |= OBD_MD_FLGRANT;
2447         /* We're finishing using body->oa as an input variable */
2448
2449         /* Check if client was evicted while we were doing i/o before touching
2450          * network */
2451         if (rc == 0) {
2452                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2453                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2454                         unsigned char *short_io_buf;
2455                         int short_io_size;
2456
2457                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2458                                                               &RMF_SHORT_IO);
2459                         short_io_size = req_capsule_get_size(&req->rq_pill,
2460                                                              &RMF_SHORT_IO,
2461                                                              RCL_SERVER);
2462                         rc = tgt_pages2shortio(local_nb, npages_read,
2463                                                short_io_buf, short_io_size);
2464                         if (rc >= 0)
2465                                 req_capsule_shrink(&req->rq_pill,
2466                                                    &RMF_SHORT_IO, rc,
2467                                                    RCL_SERVER);
2468                         rc = rc > 0 ? 0 : rc;
2469                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2470                         rc = target_bulk_io(exp, desc);
2471                 }
2472                 no_reply = rc != 0;
2473         } else {
2474                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2475                     body->oa.o_flags & OBD_FL_SHORT_IO)
2476                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2477                                            RCL_SERVER);
2478         }
2479
2480 out_commitrw:
2481         /* Must commit after prep above in all cases */
2482         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2483                           remote_nb, npages, local_nb, rc, nob, kstart);
2484 out_lock:
2485         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2486
2487         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2488                 ptlrpc_free_bulk(desc);
2489
2490         LASSERT(rc <= 0);
2491         if (rc == 0) {
2492                 rc = nob;
2493                 ptlrpc_lprocfs_brw(req, nob);
2494         } else if (no_reply) {
2495                 req->rq_no_reply = 1;
2496                 /* reply out callback would free */
2497                 ptlrpc_req_drop_rs(req);
2498                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2499                               "client will retry: rc %d\n",
2500                               obd_name,
2501                               obd_uuid2str(&exp->exp_client_uuid),
2502                               obd_export_nid2str(exp), rc);
2503         }
2504         /* send a bulk after reply to simulate a network delay or reordering
2505          * by a router - Note that !desc implies short io, so there is no bulk
2506          * to reorder. */
2507         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2508             desc) {
2509                 /* Calculate checksum before request transfer, original
2510                  * it is done by target_bulk_io() */
2511                 rc = sptlrpc_svc_wrap_bulk(req, desc);
2512                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
2513                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2514                 else /* old version, bulk matchbits is rq_xid */
2515                         req->rq_mbits = req->rq_xid;
2516
2517                 req->rq_status = rc;
2518                 target_committed_to_req(req);
2519                 target_send_reply(req, 0, 0);
2520
2521                 CDEBUG(D_INFO, "reorder BULK\n");
2522                 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
2523                                  cfs_fail_val ? : 3);
2524
2525                 target_bulk_io(exp, desc);
2526                 ptlrpc_free_bulk(desc);
2527         }
2528
2529         RETURN(rc);
2530 }
2531 EXPORT_SYMBOL(tgt_brw_read);
2532
2533 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2534                              unsigned char *buf, unsigned int size)
2535 {
2536         int     i, off, len;
2537         char    *ptr;
2538
2539         for (i = 0; i < npages; i++) {
2540                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2541                 len = local[i].lnb_len;
2542
2543                 if (len == 0)
2544                         continue;
2545
2546                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2547                        i, off, len, size);
2548                 ptr = kmap_atomic(local[i].lnb_page);
2549                 if (ptr == NULL)
2550                         return -EINVAL;
2551                 memcpy(ptr + off, buf, len < size ? len : size);
2552                 kunmap_atomic(ptr);
2553                 buf += len;
2554                 size -= len;
2555         }
2556         return 0;
2557 }
2558
2559 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2560                               struct ptlrpc_bulk_desc *desc,
2561                               struct niobuf_local *local_nb, int npages,
2562                               u32 client_cksum, u32 server_cksum,
2563                               bool mmap)
2564 {
2565         struct obd_export *exp = req->rq_export;
2566         struct ost_body *body;
2567         char *router = "";
2568         char *via = "";
2569
2570         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2571         LASSERT(body != NULL);
2572
2573         if (desc && req->rq_peer.nid != desc->bd_sender) {
2574                 via = " via ";
2575                 router = libcfs_nid2str(desc->bd_sender);
2576         }
2577
2578         if (exp->exp_obd->obd_checksum_dump)
2579                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2580                                     client_cksum);
2581
2582         if (mmap) {
2583                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2584                              client_cksum, server_cksum);
2585                 return;
2586         }
2587
2588         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2589                            DFID" object "DOSTID" extent [%llu-%llu"
2590                            "]: client csum %x, server csum %x\n",
2591                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2592                            via, router,
2593                            body->oa.o_valid & OBD_MD_FLFID ?
2594                            body->oa.o_parent_seq : (__u64)0,
2595                            body->oa.o_valid & OBD_MD_FLFID ?
2596                            body->oa.o_parent_oid : 0,
2597                            body->oa.o_valid & OBD_MD_FLFID ?
2598                            body->oa.o_parent_ver : 0,
2599                            POSTID(&body->oa.o_oi),
2600                            local_nb[0].lnb_file_offset,
2601                            local_nb[npages-1].lnb_file_offset +
2602                            local_nb[npages - 1].lnb_len - 1,
2603                            client_cksum, server_cksum);
2604 }
2605
2606 int tgt_brw_write(struct tgt_session_info *tsi)
2607 {
2608         struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
2609         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2610         struct ptlrpc_bulk_desc *desc = NULL;
2611         struct obd_export       *exp = req->rq_export;
2612         struct niobuf_remote    *remote_nb;
2613         struct niobuf_local     *local_nb;
2614         struct obd_ioobj        *ioo;
2615         struct ost_body         *body, *repbody;
2616         struct lustre_handle     lockh = {0};
2617         __u32                   *rcs;
2618         int                      objcount, niocount, npages;
2619         int                      rc = 0;
2620         int                      i, j;
2621         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2622         bool                     no_reply = false, mmap;
2623         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2624         bool wait_sync = false;
2625         const char *obd_name = exp->exp_obd->obd_name;
2626         /* '1' for consistency with code that checks !mpflag to restore */
2627         unsigned int mpflags = 1;
2628         ktime_t kstart;
2629         int nob = 0;
2630
2631         ENTRY;
2632
2633         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2634             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2635                 CERROR("%s: deny write request from %s to portal %u\n",
2636                        tgt_name(tsi->tsi_tgt),
2637                        obd_export_nid2str(req->rq_export),
2638                        ptlrpc_req2svc(req)->srv_req_portal);
2639                 RETURN(err_serious(-EPROTO));
2640         }
2641
2642         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2643                 RETURN(err_serious(-ENOSPC));
2644         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2645                 RETURN(err_serious(-EROFS));
2646
2647         req->rq_bulk_write = 1;
2648
2649         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2650                 rc = -EIO;
2651         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2652                 rc = -EFAULT;
2653         if (rc < 0) {
2654                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2655                  * this server to fail requests.
2656                  */
2657                 char fail_ost_name[MAX_OBD_NAME];
2658
2659                 if (cfs_fail_val > 0) {
2660                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2661                                  cfs_fail_val - 1);
2662
2663                         if (strstr(obd_name, fail_ost_name))
2664                                 RETURN(err_serious(rc));
2665                 } else {
2666                         RETURN(err_serious(rc));
2667                 }
2668         }
2669
2670         /* pause before transaction has been started */
2671         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2672                          cfs_fail_val : (obd_timeout + 1) / 4);
2673
2674         /* Delay write commit to show stale size information */
2675         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2676
2677         /* There must be big cache in current thread to process this request
2678          * if it is NULL then something went wrong and it wasn't allocated,
2679          * report -ENOMEM in that case */
2680         if (tbc == NULL)
2681                 RETURN(-ENOMEM);
2682
2683         body = tsi->tsi_ost_body;
2684         LASSERT(body != NULL);
2685
2686         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2687             body->oa.o_flags & OBD_FL_NORPC)
2688                 RETURN(0);
2689
2690
2691         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2692         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2693
2694         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2695                                         RCL_CLIENT) / sizeof(*ioo);
2696
2697         for (niocount = i = 0; i < objcount; i++)
2698                 niocount += ioo[i].ioo_bufcnt;
2699
2700         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2701         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2702         if (niocount != req_capsule_get_size(&req->rq_pill,
2703                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2704                         sizeof(*remote_nb))
2705                 RETURN(err_serious(-EPROTO));
2706
2707         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2708             ptlrpc_connection_is_local(exp->exp_connection))
2709                 mpflags = memalloc_noreclaim_save();
2710
2711         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2712                              niocount * sizeof(*rcs));
2713         rc = req_capsule_server_pack(&req->rq_pill);
2714         if (rc != 0)
2715                 GOTO(out, rc = err_serious(rc));
2716
2717         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2718         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2719
2720         local_nb = tbc->local;
2721
2722         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2723                           &lockh, LCK_PW);
2724         if (rc != 0)
2725                 GOTO(out, rc);
2726
2727         /*
2728          * If getting the lock took more time than
2729          * client was willing to wait, drop it. b=11330
2730          */
2731         if (ktime_get_real_seconds() > req->rq_deadline ||
2732             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2733                 no_reply = true;
2734                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2735                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2736                        POSTID(&ioo->ioo_oid),
2737                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2738                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2739                 GOTO(out_lock, rc = -ETIMEDOUT);
2740         }
2741
2742         /* Because we already sync grant info with client when reconnect,
2743          * grant info will be cleared for resent req, then fed_grant and
2744          * total_grant will not be modified in following preprw_write */
2745         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2746                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2747                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2748         }
2749
2750         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2751         if (repbody == NULL)
2752                 GOTO(out_lock, rc = -ENOMEM);
2753         repbody->oa = body->oa;
2754
2755         npages = PTLRPC_MAX_BRW_PAGES;
2756         kstart = ktime_get();
2757         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2758                         objcount, ioo, remote_nb, &npages, local_nb);
2759         if (rc < 0)
2760                 GOTO(out_lock, rc);
2761         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2762             body->oa.o_flags & OBD_FL_SHORT_IO) {
2763                 unsigned int short_io_size;
2764                 unsigned char *short_io_buf;
2765
2766                 short_io_size = req_capsule_get_size(&req->rq_pill,
2767                                                      &RMF_SHORT_IO,
2768                                                      RCL_CLIENT);
2769                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2770                                                       &RMF_SHORT_IO);
2771                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2772                                " size = %d\n", short_io_size);
2773
2774                 /* Copy short io buf to pages */
2775                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2776                                        short_io_size);
2777                 desc = NULL;
2778         } else {
2779                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2780                                             PTLRPC_BULK_GET_SINK,
2781                                             OST_BULK_PORTAL,
2782                                             &ptlrpc_bulk_kiov_nopin_ops);
2783                 if (desc == NULL)
2784                         GOTO(skip_transfer, rc = -ENOMEM);
2785
2786                 /* NB Having prepped, we must commit... */
2787                 for (i = 0; i < npages; i++)
2788                         desc->bd_frag_ops->add_kiov_frag(desc,
2789                                         local_nb[i].lnb_page,
2790                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2791                                         local_nb[i].lnb_len);
2792
2793                 rc = sptlrpc_svc_prep_bulk(req, desc);
2794                 if (rc != 0)
2795                         GOTO(skip_transfer, rc);
2796
2797                 rc = target_bulk_io(exp, desc);
2798         }
2799
2800         no_reply = rc != 0;
2801
2802 skip_transfer:
2803         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2804                 static int cksum_counter;
2805
2806                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2807                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2808
2809                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2810                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2811                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2812                                                            cksum_type);
2813
2814                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2815                                             local_nb, npages, OST_WRITE,
2816                                             &repbody->oa.o_cksum, false);
2817                 if (rc < 0)
2818                         GOTO(out_commitrw, rc);
2819
2820                 cksum_counter++;
2821
2822                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2823                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2824                                 body->oa.o_flags & OBD_FL_MMAP);
2825
2826                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2827                                           body->oa.o_cksum,
2828                                           repbody->oa.o_cksum, mmap);
2829                         cksum_counter = 0;
2830                 } else if ((cksum_counter & (-cksum_counter)) ==
2831                            cksum_counter) {
2832                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2833                                cksum_counter, libcfs_id2str(req->rq_peer),
2834                                repbody->oa.o_cksum);
2835                 }
2836         }
2837
2838         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2839
2840 out_commitrw:
2841         /* calculate the expected actual write bytes (nob) for OFD stats.
2842          * Technically, if commit fails this would be wrong, but that should be
2843          * very rare
2844          */
2845         for (i = 0; i < niocount; i++) {
2846                 int len = remote_nb[i].rnb_len;
2847
2848                 nob += len;
2849         }
2850
2851         /* multiple transactions can be assigned during write commit */
2852         tti->tti_mult_trans = 1;
2853
2854         /* Must commit after prep above in all cases */
2855         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2856                           objcount, ioo, remote_nb, npages, local_nb, rc, nob,
2857                           kstart);
2858         if (rc == -ENOTCONN)
2859                 /* quota acquire process has been given up because
2860                  * either the client has been evicted or the client
2861                  * has timed out the request already
2862                  */
2863                 no_reply = true;
2864
2865         for (i = 0; i < niocount; i++) {
2866                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2867                         wait_sync = true;
2868                         break;
2869                 }
2870         }
2871         /*
2872          * Disable sending mtime back to the client. If the client locked the
2873          * whole object, then it has already updated the mtime on its side,
2874          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2875          */
2876         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2877
2878         if (rc == 0) {
2879                 /* set per-requested niobuf return codes */
2880                 for (i = j = 0; i < niocount; i++) {
2881                         int len = remote_nb[i].rnb_len;
2882
2883                         rcs[i] = 0;
2884                         do {
2885                                 LASSERT(j < npages);
2886                                 if (local_nb[j].lnb_rc < 0)
2887                                         rcs[i] = local_nb[j].lnb_rc;
2888                                 len -= local_nb[j].lnb_len;
2889                                 j++;
2890                         } while (len > 0);
2891                         LASSERT(len == 0);
2892                 }
2893                 LASSERT(j == npages);
2894                 ptlrpc_lprocfs_brw(req, nob);
2895         }
2896 out_lock:
2897         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2898         if (desc)
2899                 ptlrpc_free_bulk(desc);
2900 out:
2901         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2902                 req->rq_no_reply = 1;
2903                 /* reply out callback would free */
2904                 ptlrpc_req_drop_rs(req);
2905                 if (!exp->exp_obd->obd_no_transno)
2906                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2907                                       " client will retry: rc = %d\n",
2908                                       obd_name,
2909                                       obd_uuid2str(&exp->exp_client_uuid),
2910                                       obd_export_nid2str(exp), rc);
2911         }
2912
2913         if (mpflags)
2914                 memalloc_noreclaim_restore(mpflags);
2915
2916         RETURN(rc);
2917 }
2918 EXPORT_SYMBOL(tgt_brw_write);
2919
2920 /**
2921  * Common request handler for OST_SEEK RPC.
2922  *
2923  * Unified request handling for OST_SEEK RPC.
2924  * It takes object by its FID, does needed lseek and packs result
2925  * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2926  *
2927  * \param[in] tsi       target session environment for this request
2928  *
2929  * \retval              0 if successful
2930  * \retval              negative value on error
2931  */
2932 int tgt_lseek(struct tgt_session_info *tsi)
2933 {
2934         struct lustre_handle lh = { 0 };
2935         struct dt_object *dob;
2936         struct ost_body *repbody;
2937         loff_t offset = tsi->tsi_ost_body->oa.o_size;
2938         int whence = tsi->tsi_ost_body->oa.o_mode;
2939         bool srvlock;
2940         int rc = 0;
2941
2942         ENTRY;
2943
2944         if (whence != SEEK_HOLE && whence != SEEK_DATA)
2945                 RETURN(-EPROTO);
2946
2947         /* Negative offset is prohibited on wire and must be handled on client
2948          * prior sending RPC.
2949          */
2950         if (offset < 0)
2951                 RETURN(-EPROTO);
2952
2953         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2954         if (repbody == NULL)
2955                 RETURN(-ENOMEM);
2956         repbody->oa = tsi->tsi_ost_body->oa;
2957
2958         srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
2959                   tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
2960         if (srvlock) {
2961                 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
2962                                    offset, OBD_OBJECT_EOF, &lh, LCK_PR);
2963                 if (rc)
2964                         RETURN(rc);
2965         }
2966
2967         dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
2968         if (IS_ERR(dob))
2969                 GOTO(out, rc = PTR_ERR(dob));
2970
2971         if (!dt_object_exists(dob))
2972                 GOTO(obj_put, rc = -ENOENT);
2973
2974         repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
2975         rc = 0;
2976 obj_put:
2977         dt_object_put(tsi->tsi_env, dob);
2978 out:
2979         if (srvlock)
2980                 tgt_data_unlock(&lh, LCK_PR);
2981
2982         RETURN(rc);
2983 }
2984 EXPORT_SYMBOL(tgt_lseek);
2985
2986 /* Check if request can be reconstructed from saved reply data
2987  * A copy of the reply data is returned in @trd if the pointer is not NULL
2988  */
2989 int req_can_reconstruct(struct ptlrpc_request *req,
2990                          struct tg_reply_data *trd)
2991 {
2992         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2993         struct lsd_client_data *lcd = ted->ted_lcd;
2994         int found;
2995
2996         if (tgt_is_multimodrpcs_client(req->rq_export))
2997                 return tgt_lookup_reply(req, trd);
2998
2999         mutex_lock(&ted->ted_lcd_lock);
3000         found = req->rq_xid == lcd->lcd_last_xid ||
3001                 req->rq_xid == lcd->lcd_last_close_xid;
3002
3003         if (found && trd != NULL) {
3004                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
3005                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
3006                         trd->trd_reply.lrd_transno =
3007                                                 lcd->lcd_last_close_transno;
3008                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
3009                 } else {
3010                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
3011                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
3012                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
3013                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
3014                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
3015                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
3016                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
3017                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
3018                 }
3019         }
3020         mutex_unlock(&ted->ted_lcd_lock);
3021
3022         return found;
3023 }
3024 EXPORT_SYMBOL(req_can_reconstruct);
3025