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, 2016, Intel Corporation.
  25  */
  26 /*
  27  * lustre/target/tgt_handler.c
  28  *
  29  * Lustre Unified Target request handler code
  30  *
  31  * Author: Brian Behlendorf <behlendorf1@llnl.gov>
  32  * Author: Mikhail Pershin <mike.pershin@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_CLASS
  36
  37 #include <linux/user_namespace.h>
  38 #ifdef HAVE_UIDGID_HEADER
  39 # include <linux/uidgid.h>
  40 #endif
  41
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47
  48 #include "tgt_internal.h"
  49
  50 char *tgt_name(struct lu_target *tgt)
  51 {
  52         LASSERT(tgt->lut_obd != NULL);
  53         return tgt->lut_obd->obd_name;
  54 }
  55 EXPORT_SYMBOL(tgt_name);
  56
  57 /*
  58  * Generic code handling requests that have struct mdt_body passed in:
  59  *
  60  *  - extract mdt_body from request and save it in @tsi, if present;
  61  *
  62  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  63  *  @tsi;
  64  *
  65  *  - if HABEO_CORPUS flag is set for this request type check whether object
  66  *  actually exists on storage (lu_object_exists()).
  67  *
  68  */
  69 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  70 {
  71         const struct mdt_body   *body;
  72         struct lu_object        *obj;
  73         struct req_capsule      *pill = tsi->tsi_pill;
  74         int                      rc;
  75
  76         ENTRY;
  77
  78         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  79         if (body == NULL)
  80                 RETURN(-EFAULT);
  81
  82         tsi->tsi_mdt_body = body;
  83
  84         if (!(body->mbo_valid & OBD_MD_FLID))
  85                 RETURN(0);
  86
  87         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  88          * in any case in pre-2.5 clients. Fix that here if needed */
  89         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  90                 RETURN(0);
  91
  92         if (!fid_is_sane(&body->mbo_fid1)) {
  93                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  94                        PFID(&body->mbo_fid1));
  95                 RETURN(-EINVAL);
  96         }
  97
  98         obj = lu_object_find(tsi->tsi_env,
  99                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 100                              &body->mbo_fid1, NULL);
 101         if (!IS_ERR(obj)) {
 102                 if ((flags & HABEO_CORPUS) && !lu_object_exists(obj)) {
 103                         lu_object_put(tsi->tsi_env, obj);
 104                         rc = -ENOENT;
 105                 } else {
 106                         tsi->tsi_corpus = obj;
 107                         rc = 0;
 108                 }
 109         } else {
 110                 rc = PTR_ERR(obj);
 111         }
 112
 113         tsi->tsi_fid = body->mbo_fid1;
 114
 115         RETURN(rc);
 116 }
 117
 118 /**
 119  * Validate oa from client.
 120  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 121  * req are valid.
 122  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 123  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 124  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 125  *       whether this is oi_fid or real ostid. So it will check
 126  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 127  *    c. Old FID-disable osc will send IDIF.
 128  *    d. new FID-enable osc/osp will send normal FID.
 129  *
 130  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 131  * be used for LAST_ID file, and only being accessed inside OST now.
 132  */
 133 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 134 {
 135         struct ost_id   *oi     = &oa->o_oi;
 136         u64              seq    = ostid_seq(oi);
 137         u64              id     = ostid_id(oi);
 138         int              rc;
 139         ENTRY;
 140
 141         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 142                      fid_seq_is_echo(seq))) {
 143                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 144                  * during create, and we will reset this to 1, since this
 145                  * oi_id is basically useless in the following create process,
 146                  * but oi_id == 0 will make it difficult to tell whether it is
 147                  * real FID or ost_id. */
 148                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 149                 oi->oi_fid.f_oid = id ?: 1;
 150                 oi->oi_fid.f_ver = 0;
 151         } else {
 152                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 153
 154                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 155                         GOTO(out, rc = -EPROTO);
 156
 157                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 158                  * all of the requests are required to setup FLGROUP */
 159                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 160                         ostid_set_seq_mdt0(oi);
 161                         oa->o_valid |= OBD_MD_FLGROUP;
 162                         seq = ostid_seq(oi);
 163                 }
 164
 165                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 166                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 167                         GOTO(out, rc = -EPROTO);
 168
 169                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 170                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 171                 if (unlikely(rc != 0))
 172                         GOTO(out, rc);
 173
 174                 oi->oi_fid = tti->tti_fid1;
 175         }
 176
 177         RETURN(0);
 178
 179 out:
 180         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 181                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 182                seq, id, rc);
 183         return rc;
 184 }
 185 EXPORT_SYMBOL(tgt_validate_obdo);
 186
 187 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 188 {
 189         unsigned                 max_brw;
 190         struct niobuf_remote    *rnb;
 191         struct obd_ioobj        *ioo;
 192         int                      obj_count;
 193
 194         ENTRY;
 195
 196         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 197         if (ioo == NULL)
 198                 RETURN(-EPROTO);
 199
 200         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 201         if (rnb == NULL)
 202                 RETURN(-EPROTO);
 203
 204         max_brw = ioobj_max_brw_get(ioo);
 205         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 206                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 207                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 208                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 209                        POSTID(oi), -EPROTO);
 210                 RETURN(-EPROTO);
 211         }
 212         ioo->ioo_oid = *oi;
 213
 214         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 215                                         RCL_CLIENT) / sizeof(*ioo);
 216         if (obj_count == 0) {
 217                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 218                 RETURN(-EPROTO);
 219         } else if (obj_count > 1) {
 220                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 221                        obj_count);
 222                 RETURN(-EPROTO);
 223         }
 224
 225         if (ioo->ioo_bufcnt == 0) {
 226                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 227                 RETURN(-EPROTO);
 228         }
 229
 230         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 231                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 232                           "bulk has too many pages (%d)",
 233                           ioo->ioo_bufcnt);
 234                 RETURN(-EPROTO);
 235         }
 236
 237         RETURN(0);
 238 }
 239
 240 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 241 {
 242         struct ost_body         *body;
 243         struct req_capsule      *pill = tsi->tsi_pill;
 244         struct lu_nodemap       *nodemap;
 245         int                      rc;
 246
 247         ENTRY;
 248
 249         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 250         if (body == NULL)
 251                 RETURN(-EFAULT);
 252
 253         rc = tgt_validate_obdo(tsi, &body->oa);
 254         if (rc)
 255                 RETURN(rc);
 256
 257         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 258         if (IS_ERR(nodemap))
 259                 RETURN(PTR_ERR(nodemap));
 260
 261         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 262                                         NODEMAP_CLIENT_TO_FS,
 263                                         body->oa.o_uid);
 264         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 265                                         NODEMAP_CLIENT_TO_FS,
 266                                         body->oa.o_gid);
 267         nodemap_putref(nodemap);
 268
 269         tsi->tsi_ost_body = body;
 270         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 271
 272         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 273                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 274                 if (rc < 0)
 275                         RETURN(rc);
 276         }
 277
 278         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 279                 if (flags & HABEO_CORPUS) {
 280                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
 281                                "but OID/FID is mandatory with HABEO_CORPUS\n",
 282                                tgt_name(tsi->tsi_tgt));
 283                         RETURN(-EPROTO);
 284                 } else {
 285                         RETURN(0);
 286                 }
 287         }
 288
 289         ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
 290
 291         /*
 292          * OST doesn't get object in advance for further use to prevent
 293          * situations with nested object_find which is potential deadlock.
 294          */
 295         tsi->tsi_corpus = NULL;
 296         RETURN(rc);
 297 }
 298
 299 /*
 300  * Do necessary preprocessing according to handler ->th_flags.
 301  */
 302 static int tgt_request_preprocess(struct tgt_session_info *tsi,
 303                                   struct tgt_handler *h,
 304                                   struct ptlrpc_request *req)
 305 {
 306         struct req_capsule      *pill = tsi->tsi_pill;
 307         __u32                    flags = h->th_flags;
 308         int                      rc = 0;
 309
 310         ENTRY;
 311
 312         if (tsi->tsi_preprocessed)
 313                 RETURN(0);
 314
 315         LASSERT(h->th_act != NULL);
 316         LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
 317         LASSERT(current->journal_info == NULL);
 318
 319         LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
 320                      h->th_fmt != NULL));
 321         if (h->th_fmt != NULL) {
 322                 req_capsule_set(pill, h->th_fmt);
 323                 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
 324                         rc = tgt_mdt_body_unpack(tsi, flags);
 325                         if (rc < 0)
 326                                 RETURN(rc);
 327                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 328                                                  RCL_CLIENT)) {
 329                         rc = tgt_ost_body_unpack(tsi, flags);
 330                         if (rc < 0)
 331                                 RETURN(rc);
 332                 }
 333         }
 334
 335         if (flags & MUTABOR && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 336                 RETURN(-EROFS);
 337
 338         if (flags & HABEO_CLAVIS) {
 339                 struct ldlm_request *dlm_req;
 340
 341                 LASSERT(h->th_fmt != NULL);
 342
 343                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 344                 if (dlm_req != NULL) {
 345                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 346                                      LDLM_IBITS &&
 347                                      dlm_req->lock_desc.l_policy_data.\
 348                                      l_inodebits.bits == 0)) {
 349                                 /*
 350                                  * Lock without inodebits makes no sense and
 351                                  * will oops later in ldlm. If client miss to
 352                                  * set such bits, do not trigger ASSERTION.
 353                                  *
 354                                  * For liblustre flock case, it maybe zero.
 355                                  */
 356                                 rc = -EPROTO;
 357                         } else {
 358                                 tsi->tsi_dlm_req = dlm_req;
 359                         }
 360                 } else {
 361                         rc = -EFAULT;
 362                 }
 363         }
 364         tsi->tsi_preprocessed = 1;
 365         RETURN(rc);
 366 }
 367
 368 /*
 369  * Invoke handler for this request opc. Also do necessary preprocessing
 370  * (according to handler ->th_flags), and post-processing (setting of
 371  * ->last_{xid,committed}).
 372  */
 373 static int tgt_handle_request0(struct tgt_session_info *tsi,
 374                                struct tgt_handler *h,
 375                                struct ptlrpc_request *req)
 376 {
 377         int      serious = 0;
 378         int      rc;
 379         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 380
 381         ENTRY;
 382
 383
 384         /* When dealing with sec context requests, no export is associated yet,
 385          * because these requests are sent before *_CONNECT requests.
 386          * A NULL req->rq_export means the normal *_common_slice handlers will
 387          * not be called, because there is no reference to the target.
 388          * So deal with them by hand and jump directly to target_send_reply().
 389          */
 390         switch (opc) {
 391         case SEC_CTX_INIT:
 392         case SEC_CTX_INIT_CONT:
 393         case SEC_CTX_FINI:
 394                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 395                 GOTO(out, rc = 0);
 396         }
 397
 398         /*
 399          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 400          * to put same checks into handlers like mdt_close(), mdt_reint(),
 401          * etc., without talking to mdt authors first. Checking same thing
 402          * there again is useless and returning 0 error without packing reply
 403          * is buggy! Handlers either pack reply or return error.
 404          *
 405          * We return 0 here and do not send any reply in order to emulate
 406          * network failure. Do not send any reply in case any of NET related
 407          * fail_id has occured.
 408          */
 409         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 410                 RETURN(0);
 411         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 412                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 413                 RETURN(0);
 414
 415         rc = tgt_request_preprocess(tsi, h, req);
 416         /* pack reply if reply format is fixed */
 417         if (rc == 0 && h->th_flags & HABEO_REFERO) {
 418                 /* Pack reply */
 419                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 420                                           RCL_SERVER))
 421                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 422                                              RCL_SERVER,
 423                                              tsi->tsi_mdt_body->mbo_eadatasize);
 424                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 425                                           RCL_SERVER))
 426                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 427                                              RCL_SERVER, 0);
 428
 429                 rc = req_capsule_server_pack(tsi->tsi_pill);
 430         }
 431
 432         if (likely(rc == 0)) {
 433                 /*
 434                  * Process request, there can be two types of rc:
 435                  * 1) errors with msg unpack/pack, other failures outside the
 436                  * operation itself. This is counted as serious errors;
 437                  * 2) errors during fs operation, should be placed in rq_status
 438                  * only
 439                  */
 440                 rc = h->th_act(tsi);
 441                 if (!is_serious(rc) &&
 442                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 443                         DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
 444                                   "pack reply and returned 0 error\n",
 445                                   tgt_name(tsi->tsi_tgt), h->th_name);
 446                         LBUG();
 447                 }
 448                 serious = is_serious(rc);
 449                 rc = clear_serious(rc);
 450         } else {
 451                 serious = 1;
 452         }
 453
 454         req->rq_status = rc;
 455
 456         /*
 457          * ELDLM_* codes which > 0 should be in rq_status only as well as
 458          * all non-serious errors.
 459          */
 460         if (rc > 0 || !serious)
 461                 rc = 0;
 462
 463         LASSERT(current->journal_info == NULL);
 464
 465         if (likely(rc == 0 && req->rq_export))
 466                 target_committed_to_req(req);
 467
 468 out:
 469         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 470         RETURN(0);
 471 }
 472
 473 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 474                                        struct obd_device *obd, int *process)
 475 {
 476         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 477         case MDS_DISCONNECT:
 478         case OST_DISCONNECT:
 479         case OBD_IDX_READ:
 480                 *process = 1;
 481                 RETURN(0);
 482         case MDS_CLOSE:
 483         case MDS_SYNC: /* used in unmounting */
 484         case OBD_PING:
 485         case MDS_REINT:
 486         case OUT_UPDATE:
 487         case SEQ_QUERY:
 488         case FLD_QUERY:
 489         case FLD_READ:
 490         case LDLM_ENQUEUE:
 491         case OST_CREATE:
 492         case OST_DESTROY:
 493         case OST_PUNCH:
 494         case OST_SETATTR:
 495         case OST_SYNC:
 496         case OST_WRITE:
 497         case MDS_HSM_PROGRESS:
 498         case MDS_HSM_STATE_SET:
 499         case MDS_HSM_REQUEST:
 500                 *process = target_queue_recovery_request(req, obd);
 501                 RETURN(0);
 502
 503         default:
 504                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 505                 *process = -EAGAIN;
 506                 RETURN(0);
 507         }
 508 }
 509
 510 /*
 511  * Handle recovery. Return:
 512  *        +1: continue request processing;
 513  *       -ve: abort immediately with the given error code;
 514  *         0: send reply with error code in req->rq_status;
 515  */
 516 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 517 {
 518         ENTRY;
 519
 520         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 521         case MDS_CONNECT:
 522         case OST_CONNECT:
 523         case MGS_CONNECT:
 524         case SEC_CTX_INIT:
 525         case SEC_CTX_INIT_CONT:
 526         case SEC_CTX_FINI:
 527                 RETURN(+1);
 528         }
 529
 530         if (!req->rq_export->exp_obd->obd_replayable)
 531                 RETURN(+1);
 532
 533         /* sanity check: if the xid matches, the request must be marked as a
 534          * resent or replayed */
 535         if (req_can_reconstruct(req, NULL)) {
 536                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 537                       (MSG_RESENT | MSG_REPLAY))) {
 538                         DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches "
 539                                   "saved xid, expected REPLAY or RESENT flag "
 540                                   "(%x)", req->rq_xid,
 541                                   lustre_msg_get_flags(req->rq_reqmsg));
 542                         req->rq_status = -ENOTCONN;
 543                         RETURN(-ENOTCONN);
 544                 }
 545         }
 546         /* else: note the opposite is not always true; a RESENT req after a
 547          * failover will usually not match the last_xid, since it was likely
 548          * never committed. A REPLAYed request will almost never match the
 549          * last xid, however it could for a committed, but still retained,
 550          * open. */
 551
 552         /* Check for aborted recovery... */
 553         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 554                 int rc;
 555                 int should_process;
 556
 557                 DEBUG_REQ(D_INFO, req, "Got new replay");
 558                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 559                                                  &should_process);
 560                 if (rc != 0 || !should_process)
 561                         RETURN(rc);
 562                 else if (should_process < 0) {
 563                         req->rq_status = should_process;
 564                         rc = ptlrpc_error(req);
 565                         RETURN(rc);
 566                 }
 567         }
 568         RETURN(+1);
 569 }
 570
 571 /* Initial check for request, it is validation mostly */
 572 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 573 {
 574         struct tgt_handler      *h;
 575         struct tgt_opc_slice    *s;
 576         struct lu_target        *tgt;
 577         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 578
 579         ENTRY;
 580
 581         tgt = class_exp2tgt(req->rq_export);
 582         if (unlikely(tgt == NULL)) {
 583                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
 584                           class_exp2obd(req->rq_export)->obd_name);
 585                 RETURN(ERR_PTR(-EINVAL));
 586         }
 587
 588         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 589                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 590                         break;
 591
 592         /* opcode was not found in slice */
 593         if (unlikely(s->tos_hs == NULL)) {
 594                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 595                        opc);
 596                 RETURN(ERR_PTR(-ENOTSUPP));
 597         }
 598
 599         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 600         h = s->tos_hs + (opc - s->tos_opc_start);
 601         if (unlikely(h->th_opc == 0)) {
 602                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 603                 RETURN(ERR_PTR(-ENOTSUPP));
 604         }
 605
 606         RETURN(h);
 607 }
 608
 609 static int process_req_last_xid(struct ptlrpc_request *req)
 610 {
 611         __u64   last_xid;
 612         ENTRY;
 613
 614         /* check request's xid is consistent with export's last_xid */
 615         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 616         if (last_xid > req->rq_export->exp_last_xid)
 617                 req->rq_export->exp_last_xid = last_xid;
 618
 619         if (req->rq_xid == 0 ||
 620             (req->rq_xid <= req->rq_export->exp_last_xid)) {
 621                 DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
 622                           "last_xid %llx\n", req->rq_xid,
 623                           req->rq_export->exp_last_xid);
 624                 /* Some request is allowed to be sent during replay,
 625                  * such as OUT update requests, FLD requests, so it
 626                  * is possible that replay requests has smaller XID
 627                  * than the exp_last_xid.
 628                  *
 629                  * Some non-replay requests may have smaller XID as
 630                  * well:
 631                  *
 632                  * - Client send a no_resend RPC, like statfs;
 633                  * - The RPC timedout (or some other error) on client,
 634                  *   then it's removed from the unreplied list;
 635                  * - Client send some other request to bump the
 636                  *   exp_last_xid on server;
 637                  * - The former RPC got chance to be processed;
 638                  */
 639                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 640                         RETURN(-EPROTO);
 641         }
 642
 643         /* try to release in-memory reply data */
 644         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 645                 tgt_handle_received_xid(req->rq_export,
 646                                 lustre_msg_get_last_xid(req->rq_reqmsg));
 647                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 648                       (MSG_RESENT | MSG_REPLAY)))
 649                         tgt_handle_tag(req->rq_export,
 650                                        lustre_msg_get_tag(req->rq_reqmsg));
 651         }
 652         RETURN(0);
 653 }
 654
 655 int tgt_request_handle(struct ptlrpc_request *req)
 656 {
 657         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 658
 659         struct lustre_msg       *msg = req->rq_reqmsg;
 660         struct tgt_handler      *h;
 661         struct lu_target        *tgt;
 662         int                      request_fail_id = 0;
 663         __u32                    opc = lustre_msg_get_opc(msg);
 664         struct obd_device       *obd;
 665         int                      rc;
 666         bool                     is_connect = false;
 667         ENTRY;
 668
 669         /* Refill the context, to make sure all thread keys are allocated */
 670         lu_env_refill(req->rq_svc_thread->t_env);
 671
 672         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 673         tsi->tsi_pill = &req->rq_pill;
 674         tsi->tsi_env = req->rq_svc_thread->t_env;
 675
 676         /* if request has export then get handlers slice from corresponding
 677          * target, otherwise that should be connect operation */
 678         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 679             opc == MGS_CONNECT) {
 680                 is_connect = true;
 681                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 682                 rc = target_handle_connect(req);
 683                 if (rc != 0) {
 684                         rc = ptlrpc_error(req);
 685                         GOTO(out, rc);
 686                 }
 687                 /* recovery-small test 18c asks to drop connect reply */
 688                 if (unlikely(opc == OST_CONNECT &&
 689                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 690                         GOTO(out, rc = 0);
 691         }
 692
 693         if (unlikely(!class_connected_export(req->rq_export))) {
 694                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 695                     opc == SEC_CTX_FINI) {
 696                         /* sec context initialization has to be handled
 697                          * by hand in tgt_handle_request0() */
 698                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 699                         h = NULL;
 700                         GOTO(handle_recov, rc = 0);
 701                 }
 702                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 703                        opc, libcfs_id2str(req->rq_peer));
 704                 req->rq_status = -ENOTCONN;
 705                 rc = ptlrpc_error(req);
 706                 GOTO(out, rc);
 707         }
 708
 709         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 710         tsi->tsi_exp = req->rq_export;
 711         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 712                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 713         else
 714                 tsi->tsi_jobid = NULL;
 715
 716         if (tgt == NULL) {
 717                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
 718                           class_exp2obd(req->rq_export)->obd_name);
 719                 req->rq_status = -EINVAL;
 720                 rc = ptlrpc_error(req);
 721                 GOTO(out, rc);
 722         }
 723
 724         /* Skip last_xid processing for the recovery thread, otherwise, the
 725          * last_xid on same request could be processed twice: first time when
 726          * processing the incoming request, second time when the request is
 727          * being processed by recovery thread. */
 728         obd = class_exp2obd(req->rq_export);
 729         if (is_connect) {
 730                 /* reset the exp_last_xid on each connection. */
 731                 req->rq_export->exp_last_xid = 0;
 732         } else if (obd->obd_recovery_data.trd_processing_task !=
 733                    current_pid()) {
 734                 rc = process_req_last_xid(req);
 735                 if (rc) {
 736                         req->rq_status = rc;
 737                         rc = ptlrpc_error(req);
 738                         GOTO(out, rc);
 739                 }
 740         }
 741
 742         request_fail_id = tgt->lut_request_fail_id;
 743         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 744
 745         h = tgt_handler_find_check(req);
 746         if (IS_ERR(h)) {
 747                 req->rq_status = PTR_ERR(h);
 748                 rc = ptlrpc_error(req);
 749                 GOTO(out, rc);
 750         }
 751
 752         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 753                  h->th_opc, opc);
 754
 755         if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 756                 GOTO(out, rc = 0);
 757
 758         rc = lustre_msg_check_version(msg, h->th_version);
 759         if (unlikely(rc)) {
 760                 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
 761                           " %08x, expecting %08x\n", tgt_name(tgt),
 762                           lustre_msg_get_version(msg), h->th_version);
 763                 req->rq_status = -EINVAL;
 764                 rc = ptlrpc_error(req);
 765                 GOTO(out, rc);
 766         }
 767
 768 handle_recov:
 769         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 770         if (likely(rc == 1)) {
 771                 rc = tgt_handle_request0(tsi, h, req);
 772                 if (rc)
 773                         GOTO(out, rc);
 774         }
 775         EXIT;
 776 out:
 777         req_capsule_fini(tsi->tsi_pill);
 778         if (tsi->tsi_corpus != NULL) {
 779                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 780                 tsi->tsi_corpus = NULL;
 781         }
 782         return rc;
 783 }
 784 EXPORT_SYMBOL(tgt_request_handle);
 785
 786 /** Assign high priority operations to the request if needed. */
 787 int tgt_hpreq_handler(struct ptlrpc_request *req)
 788 {
 789         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 790         struct tgt_handler      *h;
 791         int                      rc;
 792
 793         ENTRY;
 794
 795         if (req->rq_export == NULL)
 796                 RETURN(0);
 797
 798         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 799         tsi->tsi_pill = &req->rq_pill;
 800         tsi->tsi_env = req->rq_svc_thread->t_env;
 801         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 802         tsi->tsi_exp = req->rq_export;
 803
 804         h = tgt_handler_find_check(req);
 805         if (IS_ERR(h)) {
 806                 rc = PTR_ERR(h);
 807                 RETURN(rc);
 808         }
 809
 810         rc = tgt_request_preprocess(tsi, h, req);
 811         if (unlikely(rc != 0))
 812                 RETURN(rc);
 813
 814         if (h->th_hp != NULL)
 815                 h->th_hp(tsi);
 816         RETURN(0);
 817 }
 818 EXPORT_SYMBOL(tgt_hpreq_handler);
 819
 820 void tgt_counter_incr(struct obd_export *exp, int opcode)
 821 {
 822         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 823         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 824                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 825 }
 826 EXPORT_SYMBOL(tgt_counter_incr);
 827
 828 /*
 829  * Unified target generic handlers.
 830  */
 831
 832 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 833 {
 834         struct lu_target        *tgt = class_exp2tgt(exp);
 835         struct sptlrpc_flavor    flvr;
 836         int                      rc = 0;
 837
 838         LASSERT(tgt);
 839         LASSERT(tgt->lut_obd);
 840         LASSERT(tgt->lut_slice);
 841
 842         /* always allow ECHO client */
 843         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 844                             LUSTRE_ECHO_NAME) == 0)) {
 845                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 846                 return 0;
 847         }
 848
 849         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 850                 read_lock(&tgt->lut_sptlrpc_lock);
 851                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 852                                              req->rq_sp_from,
 853                                              req->rq_peer.nid,
 854                                              &flvr);
 855                 read_unlock(&tgt->lut_sptlrpc_lock);
 856
 857                 spin_lock(&exp->exp_lock);
 858                 exp->exp_sp_peer = req->rq_sp_from;
 859                 exp->exp_flvr = flvr;
 860
 861                 /* when on mgs, if no restriction is set, or if client
 862                  * is loopback, allow any flavor */
 863                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 864                            LUSTRE_MGS_NAME) == 0) &&
 865                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 866                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 867                       == LOLND))
 868                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 869
 870                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 871                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 872                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 873                                "expect %x\n", tgt_name(tgt),
 874                                req->rq_flvr.sf_rpc,
 875                                libcfs_nid2str(req->rq_peer.nid),
 876                                exp->exp_flvr.sf_rpc);
 877                         rc = -EACCES;
 878                 }
 879                 spin_unlock(&exp->exp_lock);
 880         } else {
 881                 if (exp->exp_sp_peer != req->rq_sp_from) {
 882                         CERROR("%s: RPC source %s doesn't match %s\n",
 883                                tgt_name(tgt),
 884                                sptlrpc_part2name(req->rq_sp_from),
 885                                sptlrpc_part2name(exp->exp_sp_peer));
 886                         rc = -EACCES;
 887                 } else {
 888                         rc = sptlrpc_target_export_check(exp, req);
 889                 }
 890         }
 891
 892         return rc;
 893 }
 894
 895 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 896 {
 897         struct sptlrpc_rule_set  tmp_rset;
 898         int                      rc;
 899
 900         if (unlikely(tgt == NULL)) {
 901                 CERROR("No target passed");
 902                 return -EINVAL;
 903         }
 904
 905         sptlrpc_rule_set_init(&tmp_rset);
 906         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 907         if (rc) {
 908                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 909                        tgt_name(tgt), rc);
 910                 return rc;
 911         }
 912
 913         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 914
 915         write_lock(&tgt->lut_sptlrpc_lock);
 916         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 917         tgt->lut_sptlrpc_rset = tmp_rset;
 918         write_unlock(&tgt->lut_sptlrpc_lock);
 919
 920         return 0;
 921 }
 922 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 923
 924 int tgt_connect(struct tgt_session_info *tsi)
 925 {
 926         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 927         struct obd_connect_data *reply;
 928         int                      rc;
 929
 930         ENTRY;
 931
 932         /* XXX: better to call this check right after getting new export but
 933          * before last_rcvd slot allocation to avoid server load upon insecure
 934          * connects. This is to be fixed after unifiyng all targets.
 935          */
 936         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 937         if (rc)
 938                 GOTO(out, rc);
 939
 940         /* To avoid exposing partially initialized connection flags, changes up
 941          * to this point have been staged in reply->ocd_connect_flags. Now that
 942          * connection handling has completed successfully, atomically update
 943          * the connect flags in the shared export data structure. LU-1623 */
 944         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
 945         spin_lock(&tsi->tsi_exp->exp_lock);
 946         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
 947         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
 948         spin_unlock(&tsi->tsi_exp->exp_lock);
 949
 950         RETURN(0);
 951 out:
 952         obd_disconnect(class_export_get(tsi->tsi_exp));
 953         return rc;
 954 }
 955 EXPORT_SYMBOL(tgt_connect);
 956
 957 int tgt_disconnect(struct tgt_session_info *tsi)
 958 {
 959         int rc;
 960
 961         ENTRY;
 962
 963         rc = target_handle_disconnect(tgt_ses_req(tsi));
 964         if (rc)
 965                 RETURN(err_serious(rc));
 966
 967         RETURN(rc);
 968 }
 969 EXPORT_SYMBOL(tgt_disconnect);
 970
 971 /*
 972  * Unified target OBD handlers
 973  */
 974 int tgt_obd_ping(struct tgt_session_info *tsi)
 975 {
 976         int rc;
 977
 978         ENTRY;
 979
 980         rc = target_handle_ping(tgt_ses_req(tsi));
 981         if (rc)
 982                 RETURN(err_serious(rc));
 983
 984         RETURN(rc);
 985 }
 986 EXPORT_SYMBOL(tgt_obd_ping);
 987
 988 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
 989 {
 990         return err_serious(-EOPNOTSUPP);
 991 }
 992
 993 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
 994 {
 995         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
 996         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 997         struct obd_export       *exp = req->rq_export;
 998         struct ptlrpc_bulk_desc *desc;
 999         struct l_wait_info      *lwi = &tti->tti_u.update.tti_wait_info;
1000         int                      i;
1001         int                      rc;
1002
1003         ENTRY;
1004
1005         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1006                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1007                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1008         if (desc == NULL)
1009                 RETURN(-ENOMEM);
1010
1011         for (i = 0; i < rdbuf->rb_nbufs; i++)
1012                 desc->bd_frag_ops->add_iov_frag(desc,
1013                                         rdbuf->rb_bufs[i].lb_buf,
1014                                         rdbuf->rb_bufs[i].lb_len);
1015
1016         rc = target_bulk_io(exp, desc, lwi);
1017         ptlrpc_free_bulk(desc);
1018         RETURN(rc);
1019 }
1020 EXPORT_SYMBOL(tgt_send_buffer);
1021
1022 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1023 {
1024         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1025         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1026         struct obd_export       *exp = req->rq_export;
1027         struct ptlrpc_bulk_desc *desc;
1028         struct l_wait_info      *lwi = &tti->tti_u.rdpg.tti_wait_info;
1029         int                      tmpcount;
1030         int                      tmpsize;
1031         int                      i;
1032         int                      rc;
1033
1034         ENTRY;
1035
1036         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1037                                     PTLRPC_BULK_PUT_SOURCE |
1038                                         PTLRPC_BULK_BUF_KIOV,
1039                                     MDS_BULK_PORTAL,
1040                                     &ptlrpc_bulk_kiov_pin_ops);
1041         if (desc == NULL)
1042                 RETURN(-ENOMEM);
1043
1044         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1045                 /* old client requires reply size in it's PAGE_SIZE,
1046                  * which is rdpg->rp_count */
1047                 nob = rdpg->rp_count;
1048
1049         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1050              i++, tmpcount -= tmpsize) {
1051                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1052                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1053                                                  tmpsize);
1054         }
1055
1056         LASSERT(desc->bd_nob == nob);
1057         rc = target_bulk_io(exp, desc, lwi);
1058         ptlrpc_free_bulk(desc);
1059         RETURN(rc);
1060 }
1061 EXPORT_SYMBOL(tgt_sendpage);
1062
1063 /*
1064  * OBD_IDX_READ handler
1065  */
1066 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1067 {
1068         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1069         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1070         struct idx_info         *req_ii, *rep_ii;
1071         int                      rc, i;
1072
1073         ENTRY;
1074
1075         memset(rdpg, 0, sizeof(*rdpg));
1076         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1077
1078         /* extract idx_info buffer from request & reply */
1079         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1080         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1081                 RETURN(err_serious(-EPROTO));
1082
1083         rc = req_capsule_server_pack(tsi->tsi_pill);
1084         if (rc)
1085                 RETURN(err_serious(rc));
1086
1087         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1088         if (rep_ii == NULL)
1089                 RETURN(err_serious(-EFAULT));
1090         rep_ii->ii_magic = IDX_INFO_MAGIC;
1091
1092         /* extract hash to start with */
1093         rdpg->rp_hash = req_ii->ii_hash_start;
1094
1095         /* extract requested attributes */
1096         rdpg->rp_attrs = req_ii->ii_attrs;
1097
1098         /* check that fid packed in request is valid and supported */
1099         if (!fid_is_sane(&req_ii->ii_fid))
1100                 RETURN(-EINVAL);
1101         rep_ii->ii_fid = req_ii->ii_fid;
1102
1103         /* copy flags */
1104         rep_ii->ii_flags = req_ii->ii_flags;
1105
1106         /* compute number of pages to allocate, ii_count is the number of 4KB
1107          * containers */
1108         if (req_ii->ii_count <= 0)
1109                 GOTO(out, rc = -EFAULT);
1110         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1111                                exp_max_brw_size(tsi->tsi_exp));
1112         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1113
1114         /* allocate pages to store the containers */
1115         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1116         if (rdpg->rp_pages == NULL)
1117                 GOTO(out, rc = -ENOMEM);
1118         for (i = 0; i < rdpg->rp_npages; i++) {
1119                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1120                 if (rdpg->rp_pages[i] == NULL)
1121                         GOTO(out, rc = -ENOMEM);
1122         }
1123
1124         /* populate pages with key/record pairs */
1125         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1126         if (rc < 0)
1127                 GOTO(out, rc);
1128
1129         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1130                  "asked %d > %d\n", rc, rdpg->rp_count);
1131
1132         /* send pages to client */
1133         rc = tgt_sendpage(tsi, rdpg, rc);
1134         if (rc)
1135                 GOTO(out, rc);
1136         EXIT;
1137 out:
1138         if (rdpg->rp_pages) {
1139                 for (i = 0; i < rdpg->rp_npages; i++)
1140                         if (rdpg->rp_pages[i])
1141                                 __free_page(rdpg->rp_pages[i]);
1142                 OBD_FREE(rdpg->rp_pages,
1143                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1144         }
1145         return rc;
1146 }
1147
1148 struct tgt_handler tgt_obd_handlers[] = {
1149 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1150 TGT_OBD_HDL_VAR(0,      OBD_LOG_CANCEL,         tgt_obd_log_cancel),
1151 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1152 };
1153 EXPORT_SYMBOL(tgt_obd_handlers);
1154
1155 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1156              struct dt_object *obj, __u64 start, __u64 end)
1157 {
1158         int rc = 0;
1159
1160         ENTRY;
1161
1162         /* if no objid is specified, it means "sync whole filesystem" */
1163         if (obj == NULL) {
1164                 rc = dt_sync(env, tgt->lut_bottom);
1165         } else if (dt_version_get(env, obj) >
1166                    tgt->lut_obd->obd_last_committed) {
1167                 rc = dt_object_sync(env, obj, start, end);
1168         }
1169         atomic_inc(&tgt->lut_sync_count);
1170
1171         RETURN(rc);
1172 }
1173 EXPORT_SYMBOL(tgt_sync);
1174 /*
1175  * Unified target DLM handlers.
1176  */
1177
1178 /**
1179  * Unified target BAST
1180  *
1181  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1182  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1183  * but if it's cross-MDT lock, because remote object version is not set, a
1184  * filesystem sync is needed.
1185  *
1186  * \param lock server side lock
1187  * \param desc lock desc
1188  * \param data ldlm_cb_set_arg
1189  * \param flag  indicates whether this cancelling or blocking callback
1190  * \retval      0 on success
1191  * \retval      negative number on error
1192  */
1193 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1194                             void *data, int flag)
1195 {
1196         struct lu_env            env;
1197         struct lu_target        *tgt;
1198         struct dt_object        *obj = NULL;
1199         struct lu_fid            fid;
1200         int                      rc = 0;
1201
1202         ENTRY;
1203
1204         tgt = class_exp2tgt(lock->l_export);
1205
1206         if (unlikely(tgt == NULL)) {
1207                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1208                        class_exp2obd(lock->l_export)->obd_name);
1209                 RETURN(-EINVAL);
1210         }
1211
1212         if (flag == LDLM_CB_CANCELING &&
1213             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1214             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1215              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1216               ldlm_is_cbpending(lock))) &&
1217             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1218              lock->l_resource->lr_type == LDLM_EXTENT)) {
1219                 __u64 start = 0;
1220                 __u64 end = OBD_OBJECT_EOF;
1221
1222                 rc = lu_env_init(&env, LCT_DT_THREAD);
1223                 if (unlikely(rc != 0))
1224                         RETURN(rc);
1225
1226                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1227                                    tgt->lut_lsd.lsd_osd_index);
1228
1229                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1230                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1231                         if (IS_ERR(obj))
1232                                 GOTO(err_env, rc = PTR_ERR(obj));
1233
1234                         if (!dt_object_exists(obj))
1235                                 GOTO(err_put, rc = -ENOENT);
1236
1237                         start = lock->l_policy_data.l_extent.start;
1238                         end = lock->l_policy_data.l_extent.end;
1239                 }
1240
1241                 rc = tgt_sync(&env, tgt, obj, start, end);
1242                 if (rc < 0) {
1243                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1244                                "cancel: rc = %d\n",
1245                                tgt_name(tgt), PFID(&fid),
1246                                lock->l_policy_data.l_extent.start,
1247                                lock->l_policy_data.l_extent.end, rc);
1248                 }
1249 err_put:
1250                 if (obj != NULL)
1251                         dt_object_put(&env, obj);
1252 err_env:
1253                 lu_env_fini(&env);
1254         }
1255
1256         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1257         RETURN(rc);
1258 }
1259
1260 static struct ldlm_callback_suite tgt_dlm_cbs = {
1261         .lcs_completion = ldlm_server_completion_ast,
1262         .lcs_blocking   = tgt_blocking_ast,
1263         .lcs_glimpse    = ldlm_server_glimpse_ast
1264 };
1265
1266 int tgt_enqueue(struct tgt_session_info *tsi)
1267 {
1268         struct ptlrpc_request *req = tgt_ses_req(tsi);
1269         int rc;
1270
1271         ENTRY;
1272         /*
1273          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1274          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1275          */
1276         LASSERT(tsi->tsi_dlm_req != NULL);
1277         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1278                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1279         if (rc)
1280                 RETURN(err_serious(rc));
1281
1282         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1283         case LUT_FAIL_MDT:
1284                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1285                 break;
1286         case LUT_FAIL_OST:
1287                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1288                 break;
1289         case LUT_FAIL_MGT:
1290                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1291                 break;
1292         default:
1293                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1294                 break;
1295         }
1296         RETURN(req->rq_status);
1297 }
1298 EXPORT_SYMBOL(tgt_enqueue);
1299
1300 int tgt_convert(struct tgt_session_info *tsi)
1301 {
1302         struct ptlrpc_request *req = tgt_ses_req(tsi);
1303         int rc;
1304
1305         ENTRY;
1306         LASSERT(tsi->tsi_dlm_req);
1307         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1308         if (rc)
1309                 RETURN(err_serious(rc));
1310
1311         RETURN(req->rq_status);
1312 }
1313
1314 int tgt_bl_callback(struct tgt_session_info *tsi)
1315 {
1316         return err_serious(-EOPNOTSUPP);
1317 }
1318
1319 int tgt_cp_callback(struct tgt_session_info *tsi)
1320 {
1321         return err_serious(-EOPNOTSUPP);
1322 }
1323
1324 /* generic LDLM target handler */
1325 struct tgt_handler tgt_dlm_handlers[] = {
1326 TGT_DLM_HDL    (HABEO_CLAVIS,   LDLM_ENQUEUE,           tgt_enqueue),
1327 TGT_DLM_HDL_VAR(HABEO_CLAVIS,   LDLM_CONVERT,           tgt_convert),
1328 TGT_DLM_HDL_VAR(0,              LDLM_BL_CALLBACK,       tgt_bl_callback),
1329 TGT_DLM_HDL_VAR(0,              LDLM_CP_CALLBACK,       tgt_cp_callback)
1330 };
1331 EXPORT_SYMBOL(tgt_dlm_handlers);
1332
1333 /*
1334  * Unified target LLOG handlers.
1335  */
1336 int tgt_llog_open(struct tgt_session_info *tsi)
1337 {
1338         int rc;
1339
1340         ENTRY;
1341
1342         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1343
1344         RETURN(rc);
1345 }
1346 EXPORT_SYMBOL(tgt_llog_open);
1347
1348 int tgt_llog_close(struct tgt_session_info *tsi)
1349 {
1350         int rc;
1351
1352         ENTRY;
1353
1354         rc = llog_origin_handle_close(tgt_ses_req(tsi));
1355
1356         RETURN(rc);
1357 }
1358 EXPORT_SYMBOL(tgt_llog_close);
1359
1360
1361 int tgt_llog_destroy(struct tgt_session_info *tsi)
1362 {
1363         int rc;
1364
1365         ENTRY;
1366
1367         rc = llog_origin_handle_destroy(tgt_ses_req(tsi));
1368
1369         RETURN(rc);
1370 }
1371
1372 int tgt_llog_read_header(struct tgt_session_info *tsi)
1373 {
1374         int rc;
1375
1376         ENTRY;
1377
1378         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1379
1380         RETURN(rc);
1381 }
1382 EXPORT_SYMBOL(tgt_llog_read_header);
1383
1384 int tgt_llog_next_block(struct tgt_session_info *tsi)
1385 {
1386         int rc;
1387
1388         ENTRY;
1389
1390         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1391
1392         RETURN(rc);
1393 }
1394 EXPORT_SYMBOL(tgt_llog_next_block);
1395
1396 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1397 {
1398         int rc;
1399
1400         ENTRY;
1401
1402         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1403
1404         RETURN(rc);
1405 }
1406 EXPORT_SYMBOL(tgt_llog_prev_block);
1407
1408 /* generic llog target handler */
1409 struct tgt_handler tgt_llog_handlers[] = {
1410 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1411 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1412 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1413 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1414 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_DESTROY,     tgt_llog_destroy),
1415 TGT_LLOG_HDL_VAR(0,     LLOG_ORIGIN_HANDLE_CLOSE,       tgt_llog_close),
1416 };
1417 EXPORT_SYMBOL(tgt_llog_handlers);
1418
1419 /*
1420  * sec context handlers
1421  */
1422 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1423 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1424 {
1425         return 0;
1426 }
1427
1428 struct tgt_handler tgt_sec_ctx_handlers[] = {
1429 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1430 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1431 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1432 };
1433 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1434
1435 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1436                                  struct dt_device *key,
1437                                  struct lfsck_req_local *lrl,
1438                                  struct thandle *th) = NULL;
1439
1440 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1441                                                       struct dt_device *,
1442                                                       struct lfsck_req_local *,
1443                                                       struct thandle *))
1444 {
1445         tgt_lfsck_in_notify_local = notify;
1446 }
1447 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1448
1449 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1450                            struct dt_device *key,
1451                            struct lfsck_request *lr) = NULL;
1452
1453 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1454                                                 struct dt_device *,
1455                                                 struct lfsck_request *))
1456 {
1457         tgt_lfsck_in_notify = notify;
1458 }
1459 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1460
1461 static int (*tgt_lfsck_query)(const struct lu_env *env,
1462                               struct dt_device *key,
1463                               struct lfsck_request *req,
1464                               struct lfsck_reply *rep,
1465                               struct lfsck_query *que) = NULL;
1466
1467 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1468                                            struct dt_device *,
1469                                            struct lfsck_request *,
1470                                            struct lfsck_reply *,
1471                                            struct lfsck_query *))
1472 {
1473         tgt_lfsck_query = query;
1474 }
1475 EXPORT_SYMBOL(tgt_register_lfsck_query);
1476
1477 /* LFSCK request handlers */
1478 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1479 {
1480         const struct lu_env     *env = tsi->tsi_env;
1481         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1482         struct lfsck_request    *lr;
1483         int                      rc;
1484         ENTRY;
1485
1486         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1487         if (lr == NULL)
1488                 RETURN(-EPROTO);
1489
1490         rc = tgt_lfsck_in_notify(env, key, lr);
1491
1492         RETURN(rc);
1493 }
1494
1495 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1496 {
1497         struct lfsck_request    *request;
1498         struct lfsck_reply      *reply;
1499         int                      rc;
1500         ENTRY;
1501
1502         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1503         if (request == NULL)
1504                 RETURN(-EPROTO);
1505
1506         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1507         if (reply == NULL)
1508                 RETURN(-ENOMEM);
1509
1510         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1511                              request, reply, NULL);
1512
1513         RETURN(rc < 0 ? rc : 0);
1514 }
1515
1516 struct tgt_handler tgt_lfsck_handlers[] = {
1517 TGT_LFSCK_HDL(HABEO_REFERO,     LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1518 TGT_LFSCK_HDL(HABEO_REFERO,     LFSCK_QUERY,    tgt_handle_lfsck_query),
1519 };
1520 EXPORT_SYMBOL(tgt_lfsck_handlers);
1521
1522 /*
1523  * initialize per-thread page pool (bug 5137).
1524  */
1525 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1526 {
1527         struct tgt_thread_big_cache *tbc;
1528
1529         ENTRY;
1530
1531         LASSERT(thread != NULL);
1532         LASSERT(thread->t_data == NULL);
1533
1534         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1535         if (tbc == NULL)
1536                 RETURN(-ENOMEM);
1537         thread->t_data = tbc;
1538         RETURN(0);
1539 }
1540 EXPORT_SYMBOL(tgt_io_thread_init);
1541
1542 /*
1543  * free per-thread pool created by tgt_thread_init().
1544  */
1545 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1546 {
1547         struct tgt_thread_big_cache *tbc;
1548
1549         ENTRY;
1550
1551         LASSERT(thread != NULL);
1552
1553         /*
1554          * be prepared to handle partially-initialized pools (because this is
1555          * called from ost_io_thread_init() for cleanup.
1556          */
1557         tbc = thread->t_data;
1558         if (tbc != NULL) {
1559                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1560                 thread->t_data = NULL;
1561         }
1562         EXIT;
1563 }
1564 EXPORT_SYMBOL(tgt_io_thread_done);
1565 /**
1566  * Helper function for getting server side [start, start+count] DLM lock
1567  * if asked by client.
1568  */
1569 int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1570                     __u64 start, __u64 end, struct lustre_handle *lh,
1571                     int mode, __u64 *flags)
1572 {
1573         union ldlm_policy_data policy;
1574         int rc;
1575
1576         ENTRY;
1577
1578         LASSERT(lh != NULL);
1579         LASSERT(ns != NULL);
1580         LASSERT(!lustre_handle_is_used(lh));
1581
1582         policy.l_extent.gid = 0;
1583         policy.l_extent.start = start & PAGE_MASK;
1584
1585         /*
1586          * If ->o_blocks is EOF it means "lock till the end of the file".
1587          * Otherwise, it's size of an extent or hole being punched (in bytes).
1588          */
1589         if (end == OBD_OBJECT_EOF || end < start)
1590                 policy.l_extent.end = OBD_OBJECT_EOF;
1591         else
1592                 policy.l_extent.end = end | ~PAGE_MASK;
1593
1594         rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_EXTENT, &policy, mode,
1595                                     flags, ldlm_blocking_ast,
1596                                     ldlm_completion_ast, ldlm_glimpse_ast,
1597                                     NULL, 0, LVB_T_NONE, NULL, lh);
1598         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1599 }
1600 EXPORT_SYMBOL(tgt_extent_lock);
1601
1602 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1603 {
1604         LASSERT(lustre_handle_is_used(lh));
1605         ldlm_lock_decref(lh, mode);
1606 }
1607 EXPORT_SYMBOL(tgt_extent_unlock);
1608
1609 int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1610                  struct obd_ioobj *obj, struct niobuf_remote *nb,
1611                  struct lustre_handle *lh, enum ldlm_mode mode)
1612 {
1613         __u64                    flags = 0;
1614         int                      nrbufs = obj->ioo_bufcnt;
1615         int                      i;
1616
1617         ENTRY;
1618
1619         LASSERT(mode == LCK_PR || mode == LCK_PW);
1620         LASSERT(!lustre_handle_is_used(lh));
1621
1622         if (ns->ns_obd->obd_recovering)
1623                 RETURN(0);
1624
1625         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1626                 RETURN(0);
1627
1628         for (i = 1; i < nrbufs; i++)
1629                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1630                         RETURN(-EFAULT);
1631
1632         RETURN(tgt_extent_lock(ns, res_id, nb[0].rnb_offset,
1633                                nb[nrbufs - 1].rnb_offset +
1634                                nb[nrbufs - 1].rnb_len - 1,
1635                                lh, mode, &flags));
1636 }
1637
1638 void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1639                     struct lustre_handle *lh, enum ldlm_mode mode)
1640 {
1641         ENTRY;
1642
1643         LASSERT(mode == LCK_PR || mode == LCK_PW);
1644         LASSERT((obj->ioo_bufcnt > 0 &&
1645                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1646                 lustre_handle_is_used(lh));
1647
1648         if (lustre_handle_is_used(lh))
1649                 tgt_extent_unlock(lh, mode);
1650         EXIT;
1651 }
1652
1653 static __u32 tgt_checksum_bulk(struct lu_target *tgt,
1654                                struct ptlrpc_bulk_desc *desc, int opc,
1655                                cksum_type_t cksum_type)
1656 {
1657         struct cfs_crypto_hash_desc     *hdesc;
1658         unsigned int                    bufsize;
1659         int                             i, err;
1660         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1661         __u32                           cksum;
1662
1663         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
1664
1665         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1666         if (IS_ERR(hdesc)) {
1667                 CERROR("%s: unable to initialize checksum hash %s\n",
1668                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1669                 return PTR_ERR(hdesc);
1670         }
1671
1672         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1673         for (i = 0; i < desc->bd_iov_count; i++) {
1674                 /* corrupt the data before we compute the checksum, to
1675                  * simulate a client->OST data error */
1676                 if (i == 0 && opc == OST_WRITE &&
1677                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1678                         int off = BD_GET_KIOV(desc, i).kiov_offset &
1679                                 ~PAGE_MASK;
1680                         int len = BD_GET_KIOV(desc, i).kiov_len;
1681                         struct page *np = tgt_page_to_corrupt;
1682                         char *ptr = kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
1683
1684                         if (np) {
1685                                 char *ptr2 = kmap(np) + off;
1686
1687                                 memcpy(ptr2, ptr, len);
1688                                 memcpy(ptr2, "bad3", min(4, len));
1689                                 kunmap(np);
1690
1691                                 /* LU-8376 to preserve original index for
1692                                  * display in dump_all_bulk_pages() */
1693                                 np->index = BD_GET_KIOV(desc,
1694                                                         i).kiov_page->index;
1695
1696                                 BD_GET_KIOV(desc, i).kiov_page = np;
1697                         } else {
1698                                 CERROR("%s: can't alloc page for corruption\n",
1699                                        tgt_name(tgt));
1700                         }
1701                 }
1702                 cfs_crypto_hash_update_page(hdesc,
1703                                   BD_GET_KIOV(desc, i).kiov_page,
1704                                   BD_GET_KIOV(desc, i).kiov_offset &
1705                                         ~PAGE_MASK,
1706                                   BD_GET_KIOV(desc, i).kiov_len);
1707
1708                  /* corrupt the data after we compute the checksum, to
1709                  * simulate an OST->client data error */
1710                 if (i == 0 && opc == OST_READ &&
1711                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1712                         int off = BD_GET_KIOV(desc, i).kiov_offset
1713                           & ~PAGE_MASK;
1714                         int len = BD_GET_KIOV(desc, i).kiov_len;
1715                         struct page *np = tgt_page_to_corrupt;
1716                         char *ptr =
1717                           kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
1718
1719                         if (np) {
1720                                 char *ptr2 = kmap(np) + off;
1721
1722                                 memcpy(ptr2, ptr, len);
1723                                 memcpy(ptr2, "bad4", min(4, len));
1724                                 kunmap(np);
1725
1726                                 /* LU-8376 to preserve original index for
1727                                  * display in dump_all_bulk_pages() */
1728                                 np->index = BD_GET_KIOV(desc,
1729                                                         i).kiov_page->index;
1730
1731                                 BD_GET_KIOV(desc, i).kiov_page = np;
1732                         } else {
1733                                 CERROR("%s: can't alloc page for corruption\n",
1734                                        tgt_name(tgt));
1735                         }
1736                 }
1737         }
1738
1739         bufsize = sizeof(cksum);
1740         err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1741
1742         return cksum;
1743 }
1744
1745 char dbgcksum_file_name[PATH_MAX];
1746
1747 static void dump_all_bulk_pages(struct obdo *oa, int count,
1748                                     lnet_kiov_t *iov, __u32 server_cksum,
1749                                     __u32 client_cksum)
1750 {
1751         struct file *filp;
1752         int rc, i;
1753         unsigned int len;
1754         char *buf;
1755         mm_segment_t oldfs;
1756
1757         /* will only keep dump of pages on first error for the same range in
1758          * file/fid, not during the resends/retries. */
1759         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1760                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1761                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1762                   libcfs_debug_file_path_arr :
1763                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1764                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1765                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1766                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1767                  (__u64)iov[0].kiov_page->index << PAGE_SHIFT,
1768                  ((__u64)iov[count - 1].kiov_page->index << PAGE_SHIFT) +
1769                  iov[count - 1].kiov_len - 1, client_cksum, server_cksum);
1770         filp = filp_open(dbgcksum_file_name,
1771                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1772         if (IS_ERR(filp)) {
1773                 rc = PTR_ERR(filp);
1774                 if (rc == -EEXIST)
1775                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1776                                "checksum error: rc = %d\n", dbgcksum_file_name,
1777                                rc);
1778                 else
1779                         CERROR("%s: can't open to dump pages with checksum "
1780                                "error: rc = %d\n", dbgcksum_file_name, rc);
1781                 return;
1782         }
1783
1784         oldfs = get_fs();
1785         set_fs(KERNEL_DS);
1786         for (i = 0; i < count; i++) {
1787                 len = iov[i].kiov_len;
1788                 buf = kmap(iov[i].kiov_page);
1789                 while (len != 0) {
1790                         rc = vfs_write(filp, (__force const char __user *)buf,
1791                                        len, &filp->f_pos);
1792                         if (rc < 0) {
1793                                 CERROR("%s: wanted to write %u but got %d "
1794                                        "error\n", dbgcksum_file_name, len, rc);
1795                                 break;
1796                         }
1797                         len -= rc;
1798                         buf += rc;
1799                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1800                                dbgcksum_file_name, rc);
1801                 }
1802                 kunmap(iov[i].kiov_page);
1803         }
1804         set_fs(oldfs);
1805
1806         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1807         if (rc)
1808                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1809         filp_close(filp, NULL);
1810         return;
1811 }
1812
1813 static int check_read_checksum(struct ptlrpc_bulk_desc *desc, struct obdo *oa,
1814                                const lnet_process_id_t *peer,
1815                                __u32 client_cksum, __u32 server_cksum,
1816                                cksum_type_t server_cksum_type)
1817 {
1818         char *msg;
1819         cksum_type_t cksum_type;
1820
1821         /* unlikely to happen and only if resend does not occur due to cksum
1822          * control failure on Client */
1823         if (unlikely(server_cksum == client_cksum)) {
1824                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1825                        client_cksum);
1826                 return 0;
1827         }
1828
1829         if (desc->bd_export->exp_obd->obd_checksum_dump)
1830                 dump_all_bulk_pages(oa, desc->bd_iov_count,
1831                                     &BD_GET_KIOV(desc, 0), server_cksum,
1832                                     client_cksum);
1833
1834         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1835                                        oa->o_flags : 0);
1836
1837         if (cksum_type != server_cksum_type)
1838                 msg = "the server may have not used the checksum type specified"
1839                       " in the original request - likely a protocol problem";
1840         else
1841                 msg = "should have changed on the client or in transit";
1842
1843         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1844                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1845                 " %x (type %x), server csum %x (type %x)\n",
1846                 desc->bd_export->exp_obd->obd_name,
1847                 msg, libcfs_nid2str(peer->nid),
1848                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1849                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1850                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1851                 POSTID(&oa->o_oi),
1852                 (__u64)BD_GET_KIOV(desc, 0).kiov_page->index << PAGE_SHIFT,
1853                 ((__u64)BD_GET_KIOV(desc,
1854                                     desc->bd_iov_count - 1).kiov_page->index
1855                         << PAGE_SHIFT) +
1856                         BD_GET_KIOV(desc, desc->bd_iov_count - 1).kiov_len - 1,
1857                 client_cksum, cksum_type, server_cksum, server_cksum_type);
1858         return 1;
1859 }
1860
1861 int tgt_brw_read(struct tgt_session_info *tsi)
1862 {
1863         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1864         struct ptlrpc_bulk_desc *desc = NULL;
1865         struct obd_export       *exp = tsi->tsi_exp;
1866         struct niobuf_remote    *remote_nb;
1867         struct niobuf_local     *local_nb;
1868         struct obd_ioobj        *ioo;
1869         struct ost_body         *body, *repbody;
1870         struct l_wait_info       lwi;
1871         struct lustre_handle     lockh = { 0 };
1872         int                      npages, nob = 0, rc, i, no_reply = 0;
1873         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
1874
1875         ENTRY;
1876
1877         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
1878                 CERROR("%s: deny read request from %s to portal %u\n",
1879                        tgt_name(tsi->tsi_tgt),
1880                        obd_export_nid2str(req->rq_export),
1881                        ptlrpc_req2svc(req)->srv_req_portal);
1882                 RETURN(-EPROTO);
1883         }
1884
1885         req->rq_bulk_read = 1;
1886
1887         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
1888                 RETURN(-EIO);
1889
1890         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
1891                          cfs_fail_val : (obd_timeout + 1) / 4);
1892
1893         /* Check if there is eviction in progress, and if so, wait for it to
1894          * finish */
1895         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1896                 /* We do not care how long it takes */
1897                 lwi = LWI_INTR(NULL, NULL);
1898                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1899                          !atomic_read(&exp->exp_obd->obd_evict_inprogress),
1900                          &lwi);
1901         }
1902
1903         /* There must be big cache in current thread to process this request
1904          * if it is NULL then something went wrong and it wasn't allocated,
1905          * report -ENOMEM in that case */
1906         if (tbc == NULL)
1907                 RETURN(-ENOMEM);
1908
1909         body = tsi->tsi_ost_body;
1910         LASSERT(body != NULL);
1911
1912         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
1913         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
1914
1915         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1916         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
1917
1918         local_nb = tbc->local;
1919
1920         rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
1921                           remote_nb, &lockh, LCK_PR);
1922         if (rc != 0)
1923                 RETURN(rc);
1924
1925         /*
1926          * If getting the lock took more time than
1927          * client was willing to wait, drop it. b=11330
1928          */
1929         if (ktime_get_real_seconds() > req->rq_deadline ||
1930             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1931                 no_reply = 1;
1932                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
1933                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
1934                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
1935                        req->rq_deadline - req->rq_arrival_time.tv_sec);
1936                 GOTO(out_lock, rc = -ETIMEDOUT);
1937         }
1938
1939         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1940         repbody->oa = body->oa;
1941
1942         npages = PTLRPC_MAX_BRW_PAGES;
1943         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
1944                         ioo, remote_nb, &npages, local_nb);
1945         if (rc != 0)
1946                 GOTO(out_lock, rc);
1947
1948         desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
1949                                     PTLRPC_BULK_PUT_SOURCE |
1950                                         PTLRPC_BULK_BUF_KIOV,
1951                                     OST_BULK_PORTAL,
1952                                     &ptlrpc_bulk_kiov_nopin_ops);
1953         if (desc == NULL)
1954                 GOTO(out_commitrw, rc = -ENOMEM);
1955
1956         nob = 0;
1957         for (i = 0; i < npages; i++) {
1958                 int page_rc = local_nb[i].lnb_rc;
1959
1960                 if (page_rc < 0) {
1961                         rc = page_rc;
1962                         break;
1963                 }
1964
1965                 nob += page_rc;
1966                 if (page_rc != 0) { /* some data! */
1967                         LASSERT(local_nb[i].lnb_page != NULL);
1968                         desc->bd_frag_ops->add_kiov_frag
1969                           (desc, local_nb[i].lnb_page,
1970                            local_nb[i].lnb_page_offset,
1971                            page_rc);
1972                 }
1973
1974                 if (page_rc != local_nb[i].lnb_len) { /* short read */
1975                         /* All subsequent pages should be 0 */
1976                         while (++i < npages)
1977                                 LASSERT(local_nb[i].lnb_rc == 0);
1978                         break;
1979                 }
1980         }
1981         if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
1982             nob != cfs_fail_val)
1983                 rc = -E2BIG;
1984
1985         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1986                 cksum_type_t cksum_type =
1987                         cksum_type_unpack(body->oa.o_valid & OBD_MD_FLFLAGS ?
1988                                           body->oa.o_flags : 0);
1989
1990                 repbody->oa.o_flags = cksum_type_pack(cksum_type);
1991                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1992                 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
1993                                                         OST_READ, cksum_type);
1994                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
1995                        repbody->oa.o_cksum);
1996
1997                 /* if a resend it could be for a cksum error, so check Server
1998                  * cksum with returned Client cksum (this should even cover
1999                  * zero-cksum case) */
2000                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2001                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2002                         check_read_checksum(desc, &body->oa, &req->rq_peer,
2003                                             body->oa.o_cksum,
2004                                             repbody->oa.o_cksum, cksum_type);
2005         } else {
2006                 repbody->oa.o_valid = 0;
2007         }
2008         /* We're finishing using body->oa as an input variable */
2009
2010         /* Check if client was evicted while we were doing i/o before touching
2011          * network */
2012         if (likely(rc == 0 &&
2013                    !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2) &&
2014                    !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_BULK))) {
2015                 rc = target_bulk_io(exp, desc, &lwi);
2016                 no_reply = rc != 0;
2017         }
2018
2019 out_commitrw:
2020         /* Must commit after prep above in all cases */
2021         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2022                           remote_nb, npages, local_nb, rc);
2023 out_lock:
2024         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2025
2026         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2027                 ptlrpc_free_bulk(desc);
2028
2029         LASSERT(rc <= 0);
2030         if (rc == 0) {
2031                 rc = nob;
2032                 ptlrpc_lprocfs_brw(req, nob);
2033         } else if (no_reply) {
2034                 req->rq_no_reply = 1;
2035                 /* reply out callback would free */
2036                 ptlrpc_req_drop_rs(req);
2037                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2038                               "client will retry: rc %d\n",
2039                               exp->exp_obd->obd_name,
2040                               obd_uuid2str(&exp->exp_client_uuid),
2041                               obd_export_nid2str(exp), rc);
2042         }
2043         /* send a bulk after reply to simulate a network delay or reordering
2044          * by a router */
2045         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
2046                 wait_queue_head_t        waitq;
2047                 struct l_wait_info       lwi1;
2048
2049                 CDEBUG(D_INFO, "reorder BULK\n");
2050                 init_waitqueue_head(&waitq);
2051
2052                 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2053                 l_wait_event(waitq, 0, &lwi1);
2054                 target_bulk_io(exp, desc, &lwi);
2055                 ptlrpc_free_bulk(desc);
2056         }
2057
2058         RETURN(rc);
2059 }
2060 EXPORT_SYMBOL(tgt_brw_read);
2061
2062 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2063                               struct ptlrpc_bulk_desc *desc,
2064                               struct niobuf_local *local_nb, int npages,
2065                               u32 client_cksum, u32 server_cksum,
2066                               bool mmap)
2067 {
2068         struct obd_export *exp = req->rq_export;
2069         struct ost_body *body;
2070         char *router = "";
2071         char *via = "";
2072
2073         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2074         LASSERT(body != NULL);
2075
2076         if (req->rq_peer.nid != desc->bd_sender) {
2077                 via = " via ";
2078                 router = libcfs_nid2str(desc->bd_sender);
2079         }
2080
2081         if (exp->exp_obd->obd_checksum_dump)
2082                 dump_all_bulk_pages(&body->oa, desc->bd_iov_count,
2083                                     &BD_GET_KIOV(desc, 0), server_cksum,
2084                                     client_cksum);
2085
2086         if (mmap) {
2087                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2088                              client_cksum, server_cksum);
2089                 return;
2090         }
2091
2092         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2093                            DFID" object "DOSTID" extent [%llu-%llu"
2094                            "]: client csum %x, server csum %x\n",
2095                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2096                            via, router,
2097                            body->oa.o_valid & OBD_MD_FLFID ?
2098                            body->oa.o_parent_seq : (__u64)0,
2099                            body->oa.o_valid & OBD_MD_FLFID ?
2100                            body->oa.o_parent_oid : 0,
2101                            body->oa.o_valid & OBD_MD_FLFID ?
2102                            body->oa.o_parent_ver : 0,
2103                            POSTID(&body->oa.o_oi),
2104                            local_nb[0].lnb_file_offset,
2105                            local_nb[npages-1].lnb_file_offset +
2106                            local_nb[npages - 1].lnb_len - 1,
2107                            client_cksum, server_cksum);
2108 }
2109
2110 int tgt_brw_write(struct tgt_session_info *tsi)
2111 {
2112         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2113         struct ptlrpc_bulk_desc *desc = NULL;
2114         struct obd_export       *exp = req->rq_export;
2115         struct niobuf_remote    *remote_nb;
2116         struct niobuf_local     *local_nb;
2117         struct obd_ioobj        *ioo;
2118         struct ost_body         *body, *repbody;
2119         struct l_wait_info       lwi;
2120         struct lustre_handle     lockh = {0};
2121         __u32                   *rcs;
2122         int                      objcount, niocount, npages;
2123         int                      rc, i, j;
2124         cksum_type_t             cksum_type = OBD_CKSUM_CRC32;
2125         bool                     no_reply = false, mmap;
2126         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2127         bool wait_sync = false;
2128
2129         ENTRY;
2130
2131         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
2132                 CERROR("%s: deny write request from %s to portal %u\n",
2133                        tgt_name(tsi->tsi_tgt),
2134                        obd_export_nid2str(req->rq_export),
2135                        ptlrpc_req2svc(req)->srv_req_portal);
2136                 RETURN(err_serious(-EPROTO));
2137         }
2138
2139         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2140                 RETURN(err_serious(-ENOSPC));
2141         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2142                 RETURN(err_serious(-EROFS));
2143
2144         req->rq_bulk_write = 1;
2145
2146         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2147                 RETURN(err_serious(-EIO));
2148         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2149                 RETURN(err_serious(-EFAULT));
2150
2151         /* pause before transaction has been started */
2152         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2153                          cfs_fail_val : (obd_timeout + 1) / 4);
2154
2155         /* There must be big cache in current thread to process this request
2156          * if it is NULL then something went wrong and it wasn't allocated,
2157          * report -ENOMEM in that case */
2158         if (tbc == NULL)
2159                 RETURN(-ENOMEM);
2160
2161         body = tsi->tsi_ost_body;
2162         LASSERT(body != NULL);
2163
2164         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2165         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2166
2167         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2168                                         RCL_CLIENT) / sizeof(*ioo);
2169
2170         for (niocount = i = 0; i < objcount; i++)
2171                 niocount += ioo[i].ioo_bufcnt;
2172
2173         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2174         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2175         if (niocount != req_capsule_get_size(&req->rq_pill,
2176                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2177                         sizeof(*remote_nb))
2178                 RETURN(err_serious(-EPROTO));
2179
2180         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2181             (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
2182                 memory_pressure_set();
2183
2184         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2185                              niocount * sizeof(*rcs));
2186         rc = req_capsule_server_pack(&req->rq_pill);
2187         if (rc != 0)
2188                 GOTO(out, rc = err_serious(rc));
2189
2190         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2191         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2192
2193         local_nb = tbc->local;
2194
2195         rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
2196                           remote_nb, &lockh, LCK_PW);
2197         if (rc != 0)
2198                 GOTO(out, rc);
2199
2200         /*
2201          * If getting the lock took more time than
2202          * client was willing to wait, drop it. b=11330
2203          */
2204         if (ktime_get_real_seconds() > req->rq_deadline ||
2205             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2206                 no_reply = true;
2207                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2208                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2209                        POSTID(&ioo->ioo_oid),
2210                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2211                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2212                 GOTO(out_lock, rc = -ETIMEDOUT);
2213         }
2214
2215         /* Because we already sync grant info with client when reconnect,
2216          * grant info will be cleared for resent req, then fed_grant and
2217          * total_grant will not be modified in following preprw_write */
2218         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2219                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2220                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2221         }
2222
2223         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2224         if (repbody == NULL)
2225                 GOTO(out_lock, rc = -ENOMEM);
2226         repbody->oa = body->oa;
2227
2228         npages = PTLRPC_MAX_BRW_PAGES;
2229         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2230                         objcount, ioo, remote_nb, &npages, local_nb);
2231         if (rc < 0)
2232                 GOTO(out_lock, rc);
2233
2234         desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2235                                     PTLRPC_BULK_GET_SINK | PTLRPC_BULK_BUF_KIOV,
2236                                     OST_BULK_PORTAL,
2237                                     &ptlrpc_bulk_kiov_nopin_ops);
2238         if (desc == NULL)
2239                 GOTO(skip_transfer, rc = -ENOMEM);
2240
2241         /* NB Having prepped, we must commit... */
2242         for (i = 0; i < npages; i++)
2243                 desc->bd_frag_ops->add_kiov_frag(desc,
2244                                                  local_nb[i].lnb_page,
2245                                                  local_nb[i].lnb_page_offset,
2246                                                  local_nb[i].lnb_len);
2247
2248         rc = sptlrpc_svc_prep_bulk(req, desc);
2249         if (rc != 0)
2250                 GOTO(skip_transfer, rc);
2251
2252         rc = target_bulk_io(exp, desc, &lwi);
2253         no_reply = rc != 0;
2254
2255 skip_transfer:
2256         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2257                 static int cksum_counter;
2258
2259                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2260                         cksum_type = cksum_type_unpack(body->oa.o_flags);
2261
2262                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2263                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2264                 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
2265                 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
2266                                                         OST_WRITE, cksum_type);
2267                 cksum_counter++;
2268
2269                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2270                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2271                                 body->oa.o_flags & OBD_FL_MMAP);
2272
2273                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2274                                           body->oa.o_cksum,
2275                                           repbody->oa.o_cksum, mmap);
2276                         cksum_counter = 0;
2277                 } else if ((cksum_counter & (-cksum_counter)) ==
2278                            cksum_counter) {
2279                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2280                                cksum_counter, libcfs_id2str(req->rq_peer),
2281                                repbody->oa.o_cksum);
2282                 }
2283         }
2284
2285         /* Must commit after prep above in all cases */
2286         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2287                           objcount, ioo, remote_nb, npages, local_nb, rc);
2288         if (rc == -ENOTCONN)
2289                 /* quota acquire process has been given up because
2290                  * either the client has been evicted or the client
2291                  * has timed out the request already */
2292                 no_reply = true;
2293
2294         for (i = 0; i < niocount; i++) {
2295                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2296                         wait_sync = true;
2297                         break;
2298                 }
2299         }
2300         /*
2301          * Disable sending mtime back to the client. If the client locked the
2302          * whole object, then it has already updated the mtime on its side,
2303          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2304          */
2305         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2306
2307         if (rc == 0) {
2308                 int nob = 0;
2309
2310                 /* set per-requested niobuf return codes */
2311                 for (i = j = 0; i < niocount; i++) {
2312                         int len = remote_nb[i].rnb_len;
2313
2314                         nob += len;
2315                         rcs[i] = 0;
2316                         do {
2317                                 LASSERT(j < npages);
2318                                 if (local_nb[j].lnb_rc < 0)
2319                                         rcs[i] = local_nb[j].lnb_rc;
2320                                 len -= local_nb[j].lnb_len;
2321                                 j++;
2322                         } while (len > 0);
2323                         LASSERT(len == 0);
2324                 }
2325                 LASSERT(j == npages);
2326                 ptlrpc_lprocfs_brw(req, nob);
2327         }
2328 out_lock:
2329         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2330         if (desc)
2331                 ptlrpc_free_bulk(desc);
2332 out:
2333         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2334                 req->rq_no_reply = 1;
2335                 /* reply out callback would free */
2336                 ptlrpc_req_drop_rs(req);
2337                 if (!exp->exp_obd->obd_no_transno)
2338                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2339                                       " client will retry: rc = %d\n",
2340                                       exp->exp_obd->obd_name,
2341                                       obd_uuid2str(&exp->exp_client_uuid),
2342                                       obd_export_nid2str(exp), rc);
2343         }
2344         memory_pressure_clr();
2345         RETURN(rc);
2346 }
2347 EXPORT_SYMBOL(tgt_brw_write);
2348
2349 /* Check if request can be reconstructed from saved reply data
2350  * A copy of the reply data is returned in @trd if the pointer is not NULL
2351  */
2352 bool req_can_reconstruct(struct ptlrpc_request *req,
2353                          struct tg_reply_data *trd)
2354 {
2355         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2356         struct lsd_client_data *lcd = ted->ted_lcd;
2357         bool found;
2358
2359         if (tgt_is_multimodrpcs_client(req->rq_export))
2360                 return tgt_lookup_reply(req, trd);
2361
2362         mutex_lock(&ted->ted_lcd_lock);
2363         found = req->rq_xid == lcd->lcd_last_xid ||
2364                 req->rq_xid == lcd->lcd_last_close_xid;
2365
2366         if (found && trd != NULL) {
2367                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2368                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2369                         trd->trd_reply.lrd_transno =
2370                                                 lcd->lcd_last_close_transno;
2371                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2372                 } else {
2373                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2374                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2375                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2376                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2377                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2378                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2379                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2380                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2381                 }
2382         }
2383         mutex_unlock(&ted->ted_lcd_lock);
2384
2385         return found;
2386 }
2387 EXPORT_SYMBOL(req_can_reconstruct);
2388