lustre/target/tgt_handler.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2013, 2017, Intel Corporation.
  25  */
  26 /*
  27  * lustre/target/tgt_handler.c
  28  *
  29  * Lustre Unified Target request handler code
  30  *
  31  * Author: Brian Behlendorf <behlendorf1@llnl.gov>
  32  * Author: Mikhail Pershin <mike.pershin@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_CLASS
  36
  37 #include <linux/user_namespace.h>
  38 #ifdef HAVE_UIDGID_HEADER
  39 # include <linux/uidgid.h>
  40 #endif
  41
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47 #include <lustre_acl.h>
  48
  49 #include "tgt_internal.h"
  50
  51 char *tgt_name(struct lu_target *tgt)
  52 {
  53         LASSERT(tgt->lut_obd != NULL);
  54         return tgt->lut_obd->obd_name;
  55 }
  56 EXPORT_SYMBOL(tgt_name);
  57
  58 /*
  59  * Generic code handling requests that have struct mdt_body passed in:
  60  *
  61  *  - extract mdt_body from request and save it in @tsi, if present;
  62  *
  63  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  64  *  @tsi;
  65  *
  66  *  - if HAS_BODY flag is set for this request type check whether object
  67  *  actually exists on storage (lu_object_exists()).
  68  *
  69  */
  70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  71 {
  72         const struct mdt_body   *body;
  73         struct lu_object        *obj;
  74         struct req_capsule      *pill = tsi->tsi_pill;
  75         int                      rc;
  76
  77         ENTRY;
  78
  79         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EFAULT);
  82
  83         tsi->tsi_mdt_body = body;
  84
  85         if (!(body->mbo_valid & OBD_MD_FLID))
  86                 RETURN(0);
  87
  88         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  89          * in any case in pre-2.5 clients. Fix that here if needed */
  90         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  91                 RETURN(0);
  92
  93         if (!fid_is_sane(&body->mbo_fid1)) {
  94                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  95                        PFID(&body->mbo_fid1));
  96                 RETURN(-EINVAL);
  97         }
  98
  99         obj = lu_object_find(tsi->tsi_env,
 100                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 101                              &body->mbo_fid1, NULL);
 102         if (!IS_ERR(obj)) {
 103                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 104                         lu_object_put(tsi->tsi_env, obj);
 105                         rc = -ENOENT;
 106                 } else {
 107                         tsi->tsi_corpus = obj;
 108                         rc = 0;
 109                 }
 110         } else {
 111                 rc = PTR_ERR(obj);
 112         }
 113
 114         tsi->tsi_fid = body->mbo_fid1;
 115
 116         RETURN(rc);
 117 }
 118
 119 /**
 120  * Validate oa from client.
 121  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 122  * req are valid.
 123  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 124  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 125  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 126  *       whether this is oi_fid or real ostid. So it will check
 127  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 128  *    c. Old FID-disable osc will send IDIF.
 129  *    d. new FID-enable osc/osp will send normal FID.
 130  *
 131  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 132  * be used for LAST_ID file, and only being accessed inside OST now.
 133  */
 134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 135 {
 136         struct ost_id   *oi     = &oa->o_oi;
 137         u64              seq    = ostid_seq(oi);
 138         u64              id     = ostid_id(oi);
 139         int              rc;
 140         ENTRY;
 141
 142         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 143                      fid_seq_is_echo(seq))) {
 144                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 145                  * during create, and we will reset this to 1, since this
 146                  * oi_id is basically useless in the following create process,
 147                  * but oi_id == 0 will make it difficult to tell whether it is
 148                  * real FID or ost_id. */
 149                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 150                 oi->oi_fid.f_oid = id ?: 1;
 151                 oi->oi_fid.f_ver = 0;
 152         } else {
 153                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 154
 155                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 156                         GOTO(out, rc = -EPROTO);
 157
 158                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 159                  * all of the requests are required to setup FLGROUP */
 160                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
 161                         ostid_set_seq_mdt0(oi);
 162                         oa->o_valid |= OBD_MD_FLGROUP;
 163                         seq = ostid_seq(oi);
 164                 }
 165
 166                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 167                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 168                         GOTO(out, rc = -EPROTO);
 169
 170                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 171                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 172                 if (unlikely(rc != 0))
 173                         GOTO(out, rc);
 174
 175                 oi->oi_fid = tti->tti_fid1;
 176         }
 177
 178         RETURN(0);
 179
 180 out:
 181         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 182                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 183                seq, id, rc);
 184         return rc;
 185 }
 186 EXPORT_SYMBOL(tgt_validate_obdo);
 187
 188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 189 {
 190         unsigned                 max_brw;
 191         struct niobuf_remote    *rnb;
 192         struct obd_ioobj        *ioo;
 193         int                      obj_count;
 194
 195         ENTRY;
 196
 197         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 198         if (ioo == NULL)
 199                 RETURN(-EPROTO);
 200
 201         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 202         if (rnb == NULL)
 203                 RETURN(-EPROTO);
 204
 205         max_brw = ioobj_max_brw_get(ioo);
 206         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 207                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 208                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 209                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 210                        POSTID(oi), -EPROTO);
 211                 RETURN(-EPROTO);
 212         }
 213         ioo->ioo_oid = *oi;
 214
 215         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 216                                         RCL_CLIENT) / sizeof(*ioo);
 217         if (obj_count == 0) {
 218                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 219                 RETURN(-EPROTO);
 220         } else if (obj_count > 1) {
 221                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 222                        obj_count);
 223                 RETURN(-EPROTO);
 224         }
 225
 226         if (ioo->ioo_bufcnt == 0) {
 227                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 228                 RETURN(-EPROTO);
 229         }
 230
 231         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 232                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 233                           "bulk has too many pages (%d)",
 234                           ioo->ioo_bufcnt);
 235                 RETURN(-EPROTO);
 236         }
 237
 238         RETURN(0);
 239 }
 240
 241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 242 {
 243         struct ost_body         *body;
 244         struct req_capsule      *pill = tsi->tsi_pill;
 245         struct lu_nodemap       *nodemap;
 246         int                      rc;
 247
 248         ENTRY;
 249
 250         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 251         if (body == NULL)
 252                 RETURN(-EFAULT);
 253
 254         rc = tgt_validate_obdo(tsi, &body->oa);
 255         if (rc)
 256                 RETURN(rc);
 257
 258         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 259         if (IS_ERR(nodemap))
 260                 RETURN(PTR_ERR(nodemap));
 261
 262         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_uid);
 265         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 266                                         NODEMAP_CLIENT_TO_FS,
 267                                         body->oa.o_gid);
 268         nodemap_putref(nodemap);
 269
 270         tsi->tsi_ost_body = body;
 271         tsi->tsi_fid = body->oa.o_oi.oi_fid;
 272
 273         if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
 274                 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
 275                 if (rc < 0)
 276                         RETURN(rc);
 277         }
 278
 279         if (!(body->oa.o_valid & OBD_MD_FLID)) {
 280                 if (flags & HAS_BODY) {
 281                         CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\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 & (HAS_BODY | HAS_REPLY),
 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 & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 336                 RETURN(-EROFS);
 337
 338         if (flags & HAS_KEY) {
 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                         union ldlm_wire_policy_data *policy =
 346                                         &dlm_req->lock_desc.l_policy_data;
 347
 348                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 349                                      LDLM_IBITS &&
 350                                      (policy->l_inodebits.bits |
 351                                       policy->l_inodebits.try_bits) == 0)) {
 352                                 /*
 353                                  * Lock without inodebits makes no sense and
 354                                  * will oops later in ldlm. If client miss to
 355                                  * set such bits, do not trigger ASSERTION.
 356                                  *
 357                                  * For liblustre flock case, it maybe zero.
 358                                  */
 359                                 rc = -EPROTO;
 360                         } else {
 361                                 tsi->tsi_dlm_req = dlm_req;
 362                         }
 363                 } else {
 364                         rc = -EFAULT;
 365                 }
 366         }
 367         tsi->tsi_preprocessed = 1;
 368         RETURN(rc);
 369 }
 370
 371 /*
 372  * Invoke handler for this request opc. Also do necessary preprocessing
 373  * (according to handler ->th_flags), and post-processing (setting of
 374  * ->last_{xid,committed}).
 375  */
 376 static int tgt_handle_request0(struct tgt_session_info *tsi,
 377                                struct tgt_handler *h,
 378                                struct ptlrpc_request *req)
 379 {
 380         int      serious = 0;
 381         int      rc;
 382         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 383
 384         ENTRY;
 385
 386
 387         /* When dealing with sec context requests, no export is associated yet,
 388          * because these requests are sent before *_CONNECT requests.
 389          * A NULL req->rq_export means the normal *_common_slice handlers will
 390          * not be called, because there is no reference to the target.
 391          * So deal with them by hand and jump directly to target_send_reply().
 392          */
 393         switch (opc) {
 394         case SEC_CTX_INIT:
 395         case SEC_CTX_INIT_CONT:
 396         case SEC_CTX_FINI:
 397                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 398                 GOTO(out, rc = 0);
 399         }
 400
 401         /*
 402          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 403          * to put same checks into handlers like mdt_close(), mdt_reint(),
 404          * etc., without talking to mdt authors first. Checking same thing
 405          * there again is useless and returning 0 error without packing reply
 406          * is buggy! Handlers either pack reply or return error.
 407          *
 408          * We return 0 here and do not send any reply in order to emulate
 409          * network failure. Do not send any reply in case any of NET related
 410          * fail_id has occured.
 411          */
 412         if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 413                 RETURN(0);
 414         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 415                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 416                 RETURN(0);
 417
 418         rc = tgt_request_preprocess(tsi, h, req);
 419         /* pack reply if reply format is fixed */
 420         if (rc == 0 && h->th_flags & HAS_REPLY) {
 421                 /* Pack reply */
 422                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 423                                           RCL_SERVER))
 424                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 425                                              RCL_SERVER,
 426                                              tsi->tsi_mdt_body->mbo_eadatasize);
 427                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 428                                           RCL_SERVER))
 429                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 430                                              RCL_SERVER, 0);
 431                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 432                         req_capsule_set_size(tsi->tsi_pill,
 433                                              &RMF_ACL, RCL_SERVER,
 434                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 435
 436                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 437                                           RCL_SERVER)) {
 438                         struct niobuf_remote *remote_nb =
 439                                 req_capsule_client_get(tsi->tsi_pill,
 440                                                        &RMF_NIOBUF_REMOTE);
 441                         struct ost_body *body = tsi->tsi_ost_body;
 442
 443                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 444                                          RCL_SERVER,
 445                                          (body->oa.o_flags & OBD_FL_SHORT_IO) ?
 446                                          remote_nb[0].rnb_len : 0);
 447                 }
 448
 449                 rc = req_capsule_server_pack(tsi->tsi_pill);
 450         }
 451
 452         if (likely(rc == 0)) {
 453                 /*
 454                  * Process request, there can be two types of rc:
 455                  * 1) errors with msg unpack/pack, other failures outside the
 456                  * operation itself. This is counted as serious errors;
 457                  * 2) errors during fs operation, should be placed in rq_status
 458                  * only
 459                  */
 460                 rc = h->th_act(tsi);
 461                 if (!is_serious(rc) &&
 462                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 463                         DEBUG_REQ(D_ERROR, req,
 464                                   "%s: %s handler did not pack reply but returned no error",
 465                                   tgt_name(tsi->tsi_tgt), h->th_name);
 466                         LBUG();
 467                 }
 468                 serious = is_serious(rc);
 469                 rc = clear_serious(rc);
 470         } else {
 471                 serious = 1;
 472         }
 473
 474         req->rq_status = rc;
 475
 476         /*
 477          * ELDLM_* codes which > 0 should be in rq_status only as well as
 478          * all non-serious errors.
 479          */
 480         if (rc > 0 || !serious)
 481                 rc = 0;
 482
 483         LASSERT(current->journal_info == NULL);
 484
 485         if (likely(rc == 0 && req->rq_export))
 486                 target_committed_to_req(req);
 487
 488 out:
 489         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 490         RETURN(0);
 491 }
 492
 493 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 494                                        struct obd_device *obd, int *process)
 495 {
 496         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 497         case MDS_DISCONNECT:
 498         case OST_DISCONNECT:
 499         case OBD_IDX_READ:
 500                 *process = 1;
 501                 RETURN(0);
 502         case MDS_CLOSE:
 503         case MDS_SYNC: /* used in unmounting */
 504         case OBD_PING:
 505         case MDS_REINT:
 506         case OUT_UPDATE:
 507         case SEQ_QUERY:
 508         case FLD_QUERY:
 509         case FLD_READ:
 510         case LDLM_ENQUEUE:
 511         case OST_CREATE:
 512         case OST_DESTROY:
 513         case OST_PUNCH:
 514         case OST_SETATTR:
 515         case OST_SYNC:
 516         case OST_WRITE:
 517         case MDS_HSM_PROGRESS:
 518         case MDS_HSM_STATE_SET:
 519         case MDS_HSM_REQUEST:
 520                 *process = target_queue_recovery_request(req, obd);
 521                 RETURN(0);
 522
 523         default:
 524                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 525                 *process = -EAGAIN;
 526                 RETURN(0);
 527         }
 528 }
 529
 530 /*
 531  * Handle recovery. Return:
 532  *        +1: continue request processing;
 533  *       -ve: abort immediately with the given error code;
 534  *         0: send reply with error code in req->rq_status;
 535  */
 536 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 537 {
 538         ENTRY;
 539
 540         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 541         case MDS_CONNECT:
 542         case OST_CONNECT:
 543         case MGS_CONNECT:
 544         case SEC_CTX_INIT:
 545         case SEC_CTX_INIT_CONT:
 546         case SEC_CTX_FINI:
 547                 RETURN(+1);
 548         }
 549
 550         if (!req->rq_export->exp_obd->obd_replayable)
 551                 RETURN(+1);
 552
 553         /* sanity check: if the xid matches, the request must be marked as a
 554          * resent or replayed */
 555         if (req_can_reconstruct(req, NULL)) {
 556                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 557                       (MSG_RESENT | MSG_REPLAY))) {
 558                         DEBUG_REQ(D_WARNING, req,
 559                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 560                                   req->rq_xid,
 561                                   lustre_msg_get_flags(req->rq_reqmsg));
 562                         req->rq_status = -ENOTCONN;
 563                         RETURN(-ENOTCONN);
 564                 }
 565         }
 566         /* else: note the opposite is not always true; a RESENT req after a
 567          * failover will usually not match the last_xid, since it was likely
 568          * never committed. A REPLAYed request will almost never match the
 569          * last xid, however it could for a committed, but still retained,
 570          * open. */
 571
 572         /* Check for aborted recovery... */
 573         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 574                 int rc;
 575                 int should_process;
 576
 577                 DEBUG_REQ(D_INFO, req, "Got new replay");
 578                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 579                                                  &should_process);
 580                 if (rc != 0 || !should_process)
 581                         RETURN(rc);
 582                 else if (should_process < 0) {
 583                         req->rq_status = should_process;
 584                         rc = ptlrpc_error(req);
 585                         RETURN(rc);
 586                 }
 587         }
 588         RETURN(+1);
 589 }
 590
 591 /* Initial check for request, it is validation mostly */
 592 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 593 {
 594         struct tgt_handler      *h;
 595         struct tgt_opc_slice    *s;
 596         struct lu_target        *tgt;
 597         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 598
 599         ENTRY;
 600
 601         tgt = class_exp2tgt(req->rq_export);
 602         if (unlikely(tgt == NULL)) {
 603                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 604                           class_exp2obd(req->rq_export)->obd_name);
 605                 RETURN(ERR_PTR(-EINVAL));
 606         }
 607
 608         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 609                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 610                         break;
 611
 612         /* opcode was not found in slice */
 613         if (unlikely(s->tos_hs == NULL)) {
 614                 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
 615                        opc);
 616                 RETURN(ERR_PTR(-ENOTSUPP));
 617         }
 618
 619         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 620         h = s->tos_hs + (opc - s->tos_opc_start);
 621         if (unlikely(h->th_opc == 0)) {
 622                 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
 623                 RETURN(ERR_PTR(-ENOTSUPP));
 624         }
 625
 626         RETURN(h);
 627 }
 628
 629 static int process_req_last_xid(struct ptlrpc_request *req)
 630 {
 631         __u64   last_xid;
 632         ENTRY;
 633
 634         /* check request's xid is consistent with export's last_xid */
 635         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 636         if (last_xid > req->rq_export->exp_last_xid)
 637                 req->rq_export->exp_last_xid = last_xid;
 638
 639         if (req->rq_xid == 0 ||
 640             (req->rq_xid <= req->rq_export->exp_last_xid)) {
 641                 DEBUG_REQ(D_WARNING, req,
 642                           "unexpected rq_xid=%llx != exp_last_xid=%llx",
 643                           req->rq_xid, req->rq_export->exp_last_xid);
 644                 /* Some request is allowed to be sent during replay,
 645                  * such as OUT update requests, FLD requests, so it
 646                  * is possible that replay requests has smaller XID
 647                  * than the exp_last_xid.
 648                  *
 649                  * Some non-replay requests may have smaller XID as
 650                  * well:
 651                  *
 652                  * - Client send a no_resend RPC, like statfs;
 653                  * - The RPC timedout (or some other error) on client,
 654                  *   then it's removed from the unreplied list;
 655                  * - Client send some other request to bump the
 656                  *   exp_last_xid on server;
 657                  * - The former RPC got chance to be processed;
 658                  */
 659                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 660                         RETURN(-EPROTO);
 661         }
 662
 663         /* The "last_xid" is the minimum xid among unreplied requests,
 664          * if the request is from the previous connection, its xid can
 665          * still be larger than "exp_last_xid", then the above check of
 666          * xid is not enough to determine whether the request is delayed.
 667          *
 668          * For example, if some replay request was delayed and caused
 669          * timeout at client and the replay is restarted, the delayed
 670          * replay request will have the larger xid than "exp_last_xid"
 671          */
 672         if (req->rq_export->exp_conn_cnt >
 673             lustre_msg_get_conn_cnt(req->rq_reqmsg))
 674                 RETURN(-ESTALE);
 675
 676         /* try to release in-memory reply data */
 677         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 678                 tgt_handle_received_xid(req->rq_export,
 679                                 lustre_msg_get_last_xid(req->rq_reqmsg));
 680                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 681                       (MSG_RESENT | MSG_REPLAY)))
 682                         tgt_handle_tag(req->rq_export,
 683                                        lustre_msg_get_tag(req->rq_reqmsg));
 684         }
 685         RETURN(0);
 686 }
 687
 688 int tgt_request_handle(struct ptlrpc_request *req)
 689 {
 690         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 691
 692         struct lustre_msg       *msg = req->rq_reqmsg;
 693         struct tgt_handler      *h;
 694         struct lu_target        *tgt;
 695         int                      request_fail_id = 0;
 696         __u32                    opc = lustre_msg_get_opc(msg);
 697         struct obd_device       *obd;
 698         int                      rc;
 699         bool                     is_connect = false;
 700         ENTRY;
 701
 702         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 703                 if (cfs_fail_val == 0 &&
 704                     lustre_msg_get_opc(msg) != OBD_PING &&
 705                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 706                         struct l_wait_info lwi =  { 0 };
 707
 708                         cfs_fail_val = 1;
 709                         cfs_race_state = 0;
 710                         l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
 711                                      &lwi);
 712                 }
 713         }
 714
 715         /* Refill the context, to make sure all thread keys are allocated */
 716         lu_env_refill(req->rq_svc_thread->t_env);
 717
 718         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 719         tsi->tsi_pill = &req->rq_pill;
 720         tsi->tsi_env = req->rq_svc_thread->t_env;
 721
 722         /* if request has export then get handlers slice from corresponding
 723          * target, otherwise that should be connect operation */
 724         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 725             opc == MGS_CONNECT) {
 726                 is_connect = true;
 727                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 728                 rc = target_handle_connect(req);
 729                 if (rc != 0) {
 730                         rc = ptlrpc_error(req);
 731                         GOTO(out, rc);
 732                 }
 733                 /* recovery-small test 18c asks to drop connect reply */
 734                 if (unlikely(opc == OST_CONNECT &&
 735                              OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 736                         GOTO(out, rc = 0);
 737         }
 738
 739         if (unlikely(!class_connected_export(req->rq_export))) {
 740                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 741                     opc == SEC_CTX_FINI) {
 742                         /* sec context initialization has to be handled
 743                          * by hand in tgt_handle_request0() */
 744                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 745                         h = NULL;
 746                         GOTO(handle_recov, rc = 0);
 747                 }
 748                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 749                        opc, libcfs_id2str(req->rq_peer));
 750                 req->rq_status = -ENOTCONN;
 751                 rc = ptlrpc_error(req);
 752                 GOTO(out, rc);
 753         }
 754
 755         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 756         tsi->tsi_exp = req->rq_export;
 757         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 758                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 759         else
 760                 tsi->tsi_jobid = NULL;
 761
 762         if (tgt == NULL) {
 763                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 764                           class_exp2obd(req->rq_export)->obd_name);
 765                 req->rq_status = -EINVAL;
 766                 rc = ptlrpc_error(req);
 767                 GOTO(out, rc);
 768         }
 769
 770         /* Skip last_xid processing for the recovery thread, otherwise, the
 771          * last_xid on same request could be processed twice: first time when
 772          * processing the incoming request, second time when the request is
 773          * being processed by recovery thread. */
 774         obd = class_exp2obd(req->rq_export);
 775         if (is_connect) {
 776                 /* reset the exp_last_xid on each connection. */
 777                 req->rq_export->exp_last_xid = 0;
 778         } else if (obd->obd_recovery_data.trd_processing_task !=
 779                    current_pid()) {
 780                 rc = process_req_last_xid(req);
 781                 if (rc) {
 782                         req->rq_status = rc;
 783                         rc = ptlrpc_error(req);
 784                         GOTO(out, rc);
 785                 }
 786         }
 787
 788         request_fail_id = tgt->lut_request_fail_id;
 789         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 790
 791         h = tgt_handler_find_check(req);
 792         if (IS_ERR(h)) {
 793                 req->rq_status = PTR_ERR(h);
 794                 rc = ptlrpc_error(req);
 795                 GOTO(out, rc);
 796         }
 797
 798         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 799                  h->th_opc, opc);
 800
 801         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 802              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 803                 GOTO(out, rc = 0);
 804
 805         rc = lustre_msg_check_version(msg, h->th_version);
 806         if (unlikely(rc)) {
 807                 DEBUG_REQ(D_ERROR, req,
 808                           "%s: drop malformed request version=%08x expect=%08x",
 809                           tgt_name(tgt), lustre_msg_get_version(msg),
 810                           h->th_version);
 811                 req->rq_status = -EINVAL;
 812                 rc = ptlrpc_error(req);
 813                 GOTO(out, rc);
 814         }
 815
 816 handle_recov:
 817         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 818         if (likely(rc == 1)) {
 819                 rc = tgt_handle_request0(tsi, h, req);
 820                 if (rc)
 821                         GOTO(out, rc);
 822         }
 823         EXIT;
 824 out:
 825         req_capsule_fini(tsi->tsi_pill);
 826         if (tsi->tsi_corpus != NULL) {
 827                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 828                 tsi->tsi_corpus = NULL;
 829         }
 830         return rc;
 831 }
 832 EXPORT_SYMBOL(tgt_request_handle);
 833
 834 /** Assign high priority operations to the request if needed. */
 835 int tgt_hpreq_handler(struct ptlrpc_request *req)
 836 {
 837         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 838         struct tgt_handler      *h;
 839         int                      rc;
 840
 841         ENTRY;
 842
 843         if (req->rq_export == NULL)
 844                 RETURN(0);
 845
 846         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 847         tsi->tsi_pill = &req->rq_pill;
 848         tsi->tsi_env = req->rq_svc_thread->t_env;
 849         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 850         tsi->tsi_exp = req->rq_export;
 851
 852         h = tgt_handler_find_check(req);
 853         if (IS_ERR(h)) {
 854                 rc = PTR_ERR(h);
 855                 RETURN(rc);
 856         }
 857
 858         rc = tgt_request_preprocess(tsi, h, req);
 859         if (unlikely(rc != 0))
 860                 RETURN(rc);
 861
 862         if (h->th_hp != NULL)
 863                 h->th_hp(tsi);
 864         RETURN(0);
 865 }
 866 EXPORT_SYMBOL(tgt_hpreq_handler);
 867
 868 void tgt_counter_incr(struct obd_export *exp, int opcode)
 869 {
 870         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 871         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 872                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 873 }
 874 EXPORT_SYMBOL(tgt_counter_incr);
 875
 876 /*
 877  * Unified target generic handlers.
 878  */
 879
 880 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 881 {
 882         struct lu_target        *tgt = class_exp2tgt(exp);
 883         struct sptlrpc_flavor    flvr;
 884         int                      rc = 0;
 885
 886         LASSERT(tgt);
 887         LASSERT(tgt->lut_obd);
 888         LASSERT(tgt->lut_slice);
 889
 890         /* always allow ECHO client */
 891         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 892                             LUSTRE_ECHO_NAME) == 0)) {
 893                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 894                 return 0;
 895         }
 896
 897         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 898                 read_lock(&tgt->lut_sptlrpc_lock);
 899                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 900                                              req->rq_sp_from,
 901                                              req->rq_peer.nid,
 902                                              &flvr);
 903                 read_unlock(&tgt->lut_sptlrpc_lock);
 904
 905                 spin_lock(&exp->exp_lock);
 906                 exp->exp_sp_peer = req->rq_sp_from;
 907                 exp->exp_flvr = flvr;
 908
 909                 /* when on mgs, if no restriction is set, or if client
 910                  * is loopback, allow any flavor */
 911                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 912                            LUSTRE_MGS_NAME) == 0) &&
 913                      (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 914                       LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
 915                       == LOLND))
 916                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 917
 918                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 919                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 920                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 921                                "expect %x\n", tgt_name(tgt),
 922                                req->rq_flvr.sf_rpc,
 923                                libcfs_nid2str(req->rq_peer.nid),
 924                                exp->exp_flvr.sf_rpc);
 925                         rc = -EACCES;
 926                 }
 927                 spin_unlock(&exp->exp_lock);
 928         } else {
 929                 if (exp->exp_sp_peer != req->rq_sp_from) {
 930                         CERROR("%s: RPC source %s doesn't match %s\n",
 931                                tgt_name(tgt),
 932                                sptlrpc_part2name(req->rq_sp_from),
 933                                sptlrpc_part2name(exp->exp_sp_peer));
 934                         rc = -EACCES;
 935                 } else {
 936                         rc = sptlrpc_target_export_check(exp, req);
 937                 }
 938         }
 939
 940         return rc;
 941 }
 942
 943 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 944 {
 945         struct sptlrpc_rule_set  tmp_rset;
 946         int                      rc;
 947
 948         if (unlikely(tgt == NULL)) {
 949                 CERROR("No target passed");
 950                 return -EINVAL;
 951         }
 952
 953         sptlrpc_rule_set_init(&tmp_rset);
 954         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
 955         if (rc) {
 956                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
 957                        tgt_name(tgt), rc);
 958                 return rc;
 959         }
 960
 961         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
 962
 963         write_lock(&tgt->lut_sptlrpc_lock);
 964         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
 965         tgt->lut_sptlrpc_rset = tmp_rset;
 966         write_unlock(&tgt->lut_sptlrpc_lock);
 967
 968         return 0;
 969 }
 970 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
 971
 972 int tgt_connect(struct tgt_session_info *tsi)
 973 {
 974         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 975         struct obd_connect_data *reply;
 976         int                      rc;
 977
 978         ENTRY;
 979
 980         /* XXX: better to call this check right after getting new export but
 981          * before last_rcvd slot allocation to avoid server load upon insecure
 982          * connects. This is to be fixed after unifiyng all targets.
 983          */
 984         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
 985         if (rc)
 986                 GOTO(out, rc);
 987
 988         /* To avoid exposing partially initialized connection flags, changes up
 989          * to this point have been staged in reply->ocd_connect_flags. Now that
 990          * connection handling has completed successfully, atomically update
 991          * the connect flags in the shared export data structure. LU-1623 */
 992         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
 993         spin_lock(&tsi->tsi_exp->exp_lock);
 994         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
 995         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
 996                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
 997                         reply->ocd_connect_flags2;
 998         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
 999         spin_unlock(&tsi->tsi_exp->exp_lock);
1000
1001         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1002                    LUSTRE_MDT_NAME) == 0) {
1003                 rc = req_check_sepol(tsi->tsi_pill);
1004                 if (rc)
1005                         GOTO(out, rc);
1006         }
1007
1008         RETURN(0);
1009 out:
1010         obd_disconnect(class_export_get(tsi->tsi_exp));
1011         return rc;
1012 }
1013 EXPORT_SYMBOL(tgt_connect);
1014
1015 int tgt_disconnect(struct tgt_session_info *tsi)
1016 {
1017         int rc;
1018
1019         ENTRY;
1020
1021         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1022
1023         rc = target_handle_disconnect(tgt_ses_req(tsi));
1024         if (rc)
1025                 RETURN(err_serious(rc));
1026
1027         RETURN(rc);
1028 }
1029 EXPORT_SYMBOL(tgt_disconnect);
1030
1031 /*
1032  * Unified target OBD handlers
1033  */
1034 int tgt_obd_ping(struct tgt_session_info *tsi)
1035 {
1036         int rc;
1037
1038         ENTRY;
1039
1040         /* The target-specific part of OBD_PING request handling.
1041          * It controls Filter Modification Data (FMD) expiration each time
1042          * PING is received.
1043          *
1044          * Valid only for replayable targets, e.g. MDT and OFD
1045          */
1046         if (tsi->tsi_exp->exp_obd->obd_replayable)
1047                 tgt_fmd_expire(tsi->tsi_exp);
1048
1049         rc = req_capsule_server_pack(tsi->tsi_pill);
1050         if (rc)
1051                 RETURN(err_serious(rc));
1052
1053         RETURN(rc);
1054 }
1055 EXPORT_SYMBOL(tgt_obd_ping);
1056
1057 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1058 {
1059         return err_serious(-EOPNOTSUPP);
1060 }
1061
1062 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1063 {
1064         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1065         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1066         struct obd_export       *exp = req->rq_export;
1067         struct ptlrpc_bulk_desc *desc;
1068         struct l_wait_info      *lwi = &tti->tti_u.update.tti_wait_info;
1069         int                      i;
1070         int                      rc;
1071
1072         ENTRY;
1073
1074         desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1075                                   PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1076                                     MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1077         if (desc == NULL)
1078                 RETURN(-ENOMEM);
1079
1080         for (i = 0; i < rdbuf->rb_nbufs; i++)
1081                 desc->bd_frag_ops->add_iov_frag(desc,
1082                                         rdbuf->rb_bufs[i].lb_buf,
1083                                         rdbuf->rb_bufs[i].lb_len);
1084
1085         rc = target_bulk_io(exp, desc, lwi);
1086         ptlrpc_free_bulk(desc);
1087         RETURN(rc);
1088 }
1089 EXPORT_SYMBOL(tgt_send_buffer);
1090
1091 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1092 {
1093         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1094         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1095         struct obd_export       *exp = req->rq_export;
1096         struct ptlrpc_bulk_desc *desc;
1097         struct l_wait_info      *lwi = &tti->tti_u.rdpg.tti_wait_info;
1098         int                      tmpcount;
1099         int                      tmpsize;
1100         int                      i;
1101         int                      rc;
1102
1103         ENTRY;
1104
1105         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1106                                     PTLRPC_BULK_PUT_SOURCE |
1107                                         PTLRPC_BULK_BUF_KIOV,
1108                                     MDS_BULK_PORTAL,
1109                                     &ptlrpc_bulk_kiov_pin_ops);
1110         if (desc == NULL)
1111                 RETURN(-ENOMEM);
1112
1113         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1114                 /* old client requires reply size in it's PAGE_SIZE,
1115                  * which is rdpg->rp_count */
1116                 nob = rdpg->rp_count;
1117
1118         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1119              i++, tmpcount -= tmpsize) {
1120                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1121                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1122                                                  tmpsize);
1123         }
1124
1125         LASSERT(desc->bd_nob == nob);
1126         rc = target_bulk_io(exp, desc, lwi);
1127         ptlrpc_free_bulk(desc);
1128         RETURN(rc);
1129 }
1130 EXPORT_SYMBOL(tgt_sendpage);
1131
1132 /*
1133  * OBD_IDX_READ handler
1134  */
1135 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1136 {
1137         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1138         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1139         struct idx_info         *req_ii, *rep_ii;
1140         int                      rc, i;
1141
1142         ENTRY;
1143
1144         memset(rdpg, 0, sizeof(*rdpg));
1145         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1146
1147         /* extract idx_info buffer from request & reply */
1148         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1149         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1150                 RETURN(err_serious(-EPROTO));
1151
1152         rc = req_capsule_server_pack(tsi->tsi_pill);
1153         if (rc)
1154                 RETURN(err_serious(rc));
1155
1156         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1157         if (rep_ii == NULL)
1158                 RETURN(err_serious(-EFAULT));
1159         rep_ii->ii_magic = IDX_INFO_MAGIC;
1160
1161         /* extract hash to start with */
1162         rdpg->rp_hash = req_ii->ii_hash_start;
1163
1164         /* extract requested attributes */
1165         rdpg->rp_attrs = req_ii->ii_attrs;
1166
1167         /* check that fid packed in request is valid and supported */
1168         if (!fid_is_sane(&req_ii->ii_fid))
1169                 RETURN(-EINVAL);
1170         rep_ii->ii_fid = req_ii->ii_fid;
1171
1172         /* copy flags */
1173         rep_ii->ii_flags = req_ii->ii_flags;
1174
1175         /* compute number of pages to allocate, ii_count is the number of 4KB
1176          * containers */
1177         if (req_ii->ii_count <= 0)
1178                 GOTO(out, rc = -EFAULT);
1179         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1180                                exp_max_brw_size(tsi->tsi_exp));
1181         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1182
1183         /* allocate pages to store the containers */
1184         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1185         if (rdpg->rp_pages == NULL)
1186                 GOTO(out, rc = -ENOMEM);
1187         for (i = 0; i < rdpg->rp_npages; i++) {
1188                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1189                 if (rdpg->rp_pages[i] == NULL)
1190                         GOTO(out, rc = -ENOMEM);
1191         }
1192
1193         /* populate pages with key/record pairs */
1194         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1195         if (rc < 0)
1196                 GOTO(out, rc);
1197
1198         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1199                  "asked %d > %d\n", rc, rdpg->rp_count);
1200
1201         /* send pages to client */
1202         rc = tgt_sendpage(tsi, rdpg, rc);
1203         if (rc)
1204                 GOTO(out, rc);
1205         EXIT;
1206 out:
1207         if (rdpg->rp_pages) {
1208                 for (i = 0; i < rdpg->rp_npages; i++)
1209                         if (rdpg->rp_pages[i])
1210                                 __free_page(rdpg->rp_pages[i]);
1211                 OBD_FREE(rdpg->rp_pages,
1212                          rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1213         }
1214         return rc;
1215 }
1216
1217 struct tgt_handler tgt_obd_handlers[] = {
1218 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1219 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1220 };
1221 EXPORT_SYMBOL(tgt_obd_handlers);
1222
1223 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1224              struct dt_object *obj, __u64 start, __u64 end)
1225 {
1226         int rc = 0;
1227
1228         ENTRY;
1229
1230         /* if no objid is specified, it means "sync whole filesystem" */
1231         if (obj == NULL) {
1232                 rc = dt_sync(env, tgt->lut_bottom);
1233         } else if (dt_version_get(env, obj) >
1234                    tgt->lut_obd->obd_last_committed) {
1235                 rc = dt_object_sync(env, obj, start, end);
1236         }
1237         atomic_inc(&tgt->lut_sync_count);
1238
1239         RETURN(rc);
1240 }
1241 EXPORT_SYMBOL(tgt_sync);
1242 /*
1243  * Unified target DLM handlers.
1244  */
1245
1246 /**
1247  * Unified target BAST
1248  *
1249  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1250  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1251  * but if it's cross-MDT lock, because remote object version is not set, a
1252  * filesystem sync is needed.
1253  *
1254  * \param lock server side lock
1255  * \param desc lock desc
1256  * \param data ldlm_cb_set_arg
1257  * \param flag  indicates whether this cancelling or blocking callback
1258  * \retval      0 on success
1259  * \retval      negative number on error
1260  */
1261 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1262                             void *data, int flag)
1263 {
1264         struct lu_env            env;
1265         struct lu_target        *tgt;
1266         struct dt_object        *obj = NULL;
1267         struct lu_fid            fid;
1268         int                      rc = 0;
1269
1270         ENTRY;
1271
1272         tgt = class_exp2tgt(lock->l_export);
1273
1274         if (unlikely(tgt == NULL)) {
1275                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1276                        class_exp2obd(lock->l_export)->obd_name);
1277                 RETURN(-EINVAL);
1278         }
1279
1280         if (flag == LDLM_CB_CANCELING &&
1281             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1282             (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1283              (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1284               ldlm_is_cbpending(lock))) &&
1285             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1286              lock->l_resource->lr_type == LDLM_EXTENT)) {
1287                 __u64 start = 0;
1288                 __u64 end = OBD_OBJECT_EOF;
1289
1290                 rc = lu_env_init(&env, LCT_DT_THREAD);
1291                 if (unlikely(rc != 0))
1292                         GOTO(err, rc);
1293
1294                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1295                                    tgt->lut_lsd.lsd_osd_index);
1296
1297                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1298                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1299                         if (IS_ERR(obj))
1300                                 GOTO(err_env, rc = PTR_ERR(obj));
1301
1302                         if (!dt_object_exists(obj))
1303                                 GOTO(err_put, rc = -ENOENT);
1304
1305                         start = lock->l_policy_data.l_extent.start;
1306                         end = lock->l_policy_data.l_extent.end;
1307                 }
1308
1309                 rc = tgt_sync(&env, tgt, obj, start, end);
1310                 if (rc < 0) {
1311                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1312                                "cancel: rc = %d\n",
1313                                tgt_name(tgt), PFID(&fid),
1314                                lock->l_policy_data.l_extent.start,
1315                                lock->l_policy_data.l_extent.end, rc);
1316                 }
1317 err_put:
1318                 if (obj != NULL)
1319                         dt_object_put(&env, obj);
1320 err_env:
1321                 lu_env_fini(&env);
1322         }
1323 err:
1324         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1325         RETURN(rc);
1326 }
1327
1328 static struct ldlm_callback_suite tgt_dlm_cbs = {
1329         .lcs_completion = ldlm_server_completion_ast,
1330         .lcs_blocking   = tgt_blocking_ast,
1331         .lcs_glimpse    = ldlm_server_glimpse_ast
1332 };
1333
1334 int tgt_enqueue(struct tgt_session_info *tsi)
1335 {
1336         struct ptlrpc_request *req = tgt_ses_req(tsi);
1337         int rc;
1338
1339         ENTRY;
1340         /*
1341          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1342          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1343          */
1344         LASSERT(tsi->tsi_dlm_req != NULL);
1345         rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1346                                   tsi->tsi_dlm_req, &tgt_dlm_cbs);
1347         if (rc)
1348                 RETURN(err_serious(rc));
1349
1350         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1351         case LUT_FAIL_MDT:
1352                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1353                 break;
1354         case LUT_FAIL_OST:
1355                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1356                 break;
1357         case LUT_FAIL_MGT:
1358                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1359                 break;
1360         default:
1361                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1362                 break;
1363         }
1364         RETURN(req->rq_status);
1365 }
1366 EXPORT_SYMBOL(tgt_enqueue);
1367
1368 int tgt_convert(struct tgt_session_info *tsi)
1369 {
1370         struct ptlrpc_request *req = tgt_ses_req(tsi);
1371         int rc;
1372
1373         ENTRY;
1374         LASSERT(tsi->tsi_dlm_req);
1375         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1376         if (rc)
1377                 RETURN(err_serious(rc));
1378
1379         RETURN(req->rq_status);
1380 }
1381
1382 int tgt_bl_callback(struct tgt_session_info *tsi)
1383 {
1384         return err_serious(-EOPNOTSUPP);
1385 }
1386
1387 int tgt_cp_callback(struct tgt_session_info *tsi)
1388 {
1389         return err_serious(-EOPNOTSUPP);
1390 }
1391
1392 /* generic LDLM target handler */
1393 struct tgt_handler tgt_dlm_handlers[] = {
1394 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1395 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1396 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1397 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1398 };
1399 EXPORT_SYMBOL(tgt_dlm_handlers);
1400
1401 /*
1402  * Unified target LLOG handlers.
1403  */
1404 int tgt_llog_open(struct tgt_session_info *tsi)
1405 {
1406         int rc;
1407
1408         ENTRY;
1409
1410         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1411
1412         RETURN(rc);
1413 }
1414 EXPORT_SYMBOL(tgt_llog_open);
1415
1416 int tgt_llog_read_header(struct tgt_session_info *tsi)
1417 {
1418         int rc;
1419
1420         ENTRY;
1421
1422         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1423
1424         RETURN(rc);
1425 }
1426 EXPORT_SYMBOL(tgt_llog_read_header);
1427
1428 int tgt_llog_next_block(struct tgt_session_info *tsi)
1429 {
1430         int rc;
1431
1432         ENTRY;
1433
1434         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1435
1436         RETURN(rc);
1437 }
1438 EXPORT_SYMBOL(tgt_llog_next_block);
1439
1440 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1441 {
1442         int rc;
1443
1444         ENTRY;
1445
1446         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1447
1448         RETURN(rc);
1449 }
1450 EXPORT_SYMBOL(tgt_llog_prev_block);
1451
1452 /* generic llog target handler */
1453 struct tgt_handler tgt_llog_handlers[] = {
1454 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1455 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1456 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1457 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1458 };
1459 EXPORT_SYMBOL(tgt_llog_handlers);
1460
1461 /*
1462  * sec context handlers
1463  */
1464 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1465 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1466 {
1467         return 0;
1468 }
1469
1470 struct tgt_handler tgt_sec_ctx_handlers[] = {
1471 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1472 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1473 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1474 };
1475 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1476
1477 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1478                                  struct dt_device *key,
1479                                  struct lfsck_req_local *lrl,
1480                                  struct thandle *th) = NULL;
1481
1482 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1483                                                       struct dt_device *,
1484                                                       struct lfsck_req_local *,
1485                                                       struct thandle *))
1486 {
1487         tgt_lfsck_in_notify_local = notify;
1488 }
1489 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1490
1491 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1492                            struct dt_device *key,
1493                            struct lfsck_request *lr) = NULL;
1494
1495 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1496                                                 struct dt_device *,
1497                                                 struct lfsck_request *))
1498 {
1499         tgt_lfsck_in_notify = notify;
1500 }
1501 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1502
1503 static int (*tgt_lfsck_query)(const struct lu_env *env,
1504                               struct dt_device *key,
1505                               struct lfsck_request *req,
1506                               struct lfsck_reply *rep,
1507                               struct lfsck_query *que) = NULL;
1508
1509 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1510                                            struct dt_device *,
1511                                            struct lfsck_request *,
1512                                            struct lfsck_reply *,
1513                                            struct lfsck_query *))
1514 {
1515         tgt_lfsck_query = query;
1516 }
1517 EXPORT_SYMBOL(tgt_register_lfsck_query);
1518
1519 /* LFSCK request handlers */
1520 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1521 {
1522         const struct lu_env     *env = tsi->tsi_env;
1523         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1524         struct lfsck_request    *lr;
1525         int                      rc;
1526         ENTRY;
1527
1528         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1529         if (lr == NULL)
1530                 RETURN(-EPROTO);
1531
1532         rc = tgt_lfsck_in_notify(env, key, lr);
1533
1534         RETURN(rc);
1535 }
1536
1537 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1538 {
1539         struct lfsck_request    *request;
1540         struct lfsck_reply      *reply;
1541         int                      rc;
1542         ENTRY;
1543
1544         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1545         if (request == NULL)
1546                 RETURN(-EPROTO);
1547
1548         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1549         if (reply == NULL)
1550                 RETURN(-ENOMEM);
1551
1552         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1553                              request, reply, NULL);
1554
1555         RETURN(rc < 0 ? rc : 0);
1556 }
1557
1558 struct tgt_handler tgt_lfsck_handlers[] = {
1559 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1560 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1561 };
1562 EXPORT_SYMBOL(tgt_lfsck_handlers);
1563
1564 /*
1565  * initialize per-thread page pool (bug 5137).
1566  */
1567 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1568 {
1569         struct tgt_thread_big_cache *tbc;
1570
1571         ENTRY;
1572
1573         LASSERT(thread != NULL);
1574         LASSERT(thread->t_data == NULL);
1575
1576         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1577         if (tbc == NULL)
1578                 RETURN(-ENOMEM);
1579         thread->t_data = tbc;
1580         RETURN(0);
1581 }
1582 EXPORT_SYMBOL(tgt_io_thread_init);
1583
1584 /*
1585  * free per-thread pool created by tgt_thread_init().
1586  */
1587 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1588 {
1589         struct tgt_thread_big_cache *tbc;
1590
1591         ENTRY;
1592
1593         LASSERT(thread != NULL);
1594
1595         /*
1596          * be prepared to handle partially-initialized pools (because this is
1597          * called from ost_io_thread_init() for cleanup.
1598          */
1599         tbc = thread->t_data;
1600         if (tbc != NULL) {
1601                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1602                 thread->t_data = NULL;
1603         }
1604         EXIT;
1605 }
1606 EXPORT_SYMBOL(tgt_io_thread_done);
1607
1608 /**
1609  * Helper function for getting Data-on-MDT file server DLM lock
1610  * if asked by client.
1611  */
1612 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1613                       struct lustre_handle *lh, int mode, __u64 *flags)
1614 {
1615         union ldlm_policy_data policy = {
1616                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1617         };
1618         int rc;
1619
1620         ENTRY;
1621
1622         LASSERT(lh != NULL);
1623         LASSERT(ns != NULL);
1624         LASSERT(!lustre_handle_is_used(lh));
1625
1626         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1627                                     flags, ldlm_blocking_ast,
1628                                     ldlm_completion_ast, ldlm_glimpse_ast,
1629                                     NULL, 0, LVB_T_NONE, NULL, lh);
1630
1631         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1632 }
1633 EXPORT_SYMBOL(tgt_mdt_data_lock);
1634
1635 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1636 {
1637         LASSERT(lustre_handle_is_used(lh));
1638         ldlm_lock_decref(lh, mode);
1639 }
1640 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1641
1642 /**
1643  * Helper function for getting server side [start, start+count] DLM lock
1644  * if asked by client.
1645  */
1646 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1647                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1648                     struct lustre_handle *lh, int mode, __u64 *flags)
1649 {
1650         union ldlm_policy_data policy;
1651         int rc;
1652
1653         ENTRY;
1654
1655         LASSERT(lh != NULL);
1656         LASSERT(ns != NULL);
1657         LASSERT(!lustre_handle_is_used(lh));
1658
1659         policy.l_extent.gid = 0;
1660         policy.l_extent.start = start & PAGE_MASK;
1661
1662         /*
1663          * If ->o_blocks is EOF it means "lock till the end of the file".
1664          * Otherwise, it's size of an extent or hole being punched (in bytes).
1665          */
1666         if (end == OBD_OBJECT_EOF || end < start)
1667                 policy.l_extent.end = OBD_OBJECT_EOF;
1668         else
1669                 policy.l_extent.end = end | ~PAGE_MASK;
1670
1671         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1672                                     mode, flags, ldlm_blocking_ast,
1673                                     ldlm_completion_ast, ldlm_glimpse_ast,
1674                                     NULL, 0, LVB_T_NONE, NULL, lh);
1675         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1676 }
1677 EXPORT_SYMBOL(tgt_extent_lock);
1678
1679 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1680 {
1681         LASSERT(lustre_handle_is_used(lh));
1682         ldlm_lock_decref(lh, mode);
1683 }
1684 EXPORT_SYMBOL(tgt_extent_unlock);
1685
1686 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1687                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1688                         struct niobuf_remote *nb, struct lustre_handle *lh,
1689                         enum ldlm_mode mode)
1690 {
1691         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
1692         __u64                    flags = 0;
1693         int                      nrbufs = obj->ioo_bufcnt;
1694         int                      i;
1695         int                      rc;
1696
1697         ENTRY;
1698
1699         LASSERT(mode == LCK_PR || mode == LCK_PW);
1700         LASSERT(!lustre_handle_is_used(lh));
1701
1702         if (ns->ns_obd->obd_recovering)
1703                 RETURN(0);
1704
1705         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1706                 RETURN(0);
1707
1708         for (i = 1; i < nrbufs; i++)
1709                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1710                         RETURN(-EFAULT);
1711
1712         /* MDT IO for data-on-mdt */
1713         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1714                 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1715         else
1716                 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1717                                      nb[nrbufs - 1].rnb_offset +
1718                                      nb[nrbufs - 1].rnb_len - 1,
1719                                      lh, mode, &flags);
1720         RETURN(rc);
1721 }
1722
1723 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1724                            struct lustre_handle *lh, enum ldlm_mode mode)
1725 {
1726         ENTRY;
1727
1728         LASSERT(mode == LCK_PR || mode == LCK_PW);
1729         LASSERT((obj->ioo_bufcnt > 0 &&
1730                  (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1731                 lustre_handle_is_used(lh));
1732
1733         if (lustre_handle_is_used(lh))
1734                 tgt_extent_unlock(lh, mode);
1735         EXIT;
1736 }
1737 static int tgt_checksum_niobuf(struct lu_target *tgt,
1738                                  struct niobuf_local *local_nb, int npages,
1739                                  int opc, enum cksum_types cksum_type,
1740                                  __u32 *cksum)
1741 {
1742         struct ahash_request           *req;
1743         unsigned int                    bufsize;
1744         int                             i, err;
1745         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1746
1747         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1748         if (IS_ERR(req)) {
1749                 CERROR("%s: unable to initialize checksum hash %s\n",
1750                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1751                 return PTR_ERR(req);
1752         }
1753
1754         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1755         for (i = 0; i < npages; i++) {
1756                 /* corrupt the data before we compute the checksum, to
1757                  * simulate a client->OST data error */
1758                 if (i == 0 && opc == OST_WRITE &&
1759                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1760                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1761                         int len = local_nb[i].lnb_len;
1762                         struct page *np = tgt_page_to_corrupt;
1763
1764                         if (np) {
1765                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1766                                                         KM_USER0);
1767                                 char *ptr2 = page_address(np);
1768
1769                                 memcpy(ptr2 + off, ptr + off, len);
1770                                 memcpy(ptr2 + off, "bad3", min(4, len));
1771                                 ll_kunmap_atomic(ptr, KM_USER0);
1772
1773                                 /* LU-8376 to preserve original index for
1774                                  * display in dump_all_bulk_pages() */
1775                                 np->index = i;
1776
1777                                 cfs_crypto_hash_update_page(req, np, off,
1778                                                             len);
1779                                 continue;
1780                         } else {
1781                                 CERROR("%s: can't alloc page for corruption\n",
1782                                        tgt_name(tgt));
1783                         }
1784                 }
1785                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1786                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1787                                   local_nb[i].lnb_len);
1788
1789                  /* corrupt the data after we compute the checksum, to
1790                  * simulate an OST->client data error */
1791                 if (i == 0 && opc == OST_READ &&
1792                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1793                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1794                         int len = local_nb[i].lnb_len;
1795                         struct page *np = tgt_page_to_corrupt;
1796
1797                         if (np) {
1798                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1799                                                         KM_USER0);
1800                                 char *ptr2 = page_address(np);
1801
1802                                 memcpy(ptr2 + off, ptr + off, len);
1803                                 memcpy(ptr2 + off, "bad4", min(4, len));
1804                                 ll_kunmap_atomic(ptr, KM_USER0);
1805
1806                                 /* LU-8376 to preserve original index for
1807                                  * display in dump_all_bulk_pages() */
1808                                 np->index = i;
1809
1810                                 cfs_crypto_hash_update_page(req, np, off,
1811                                                             len);
1812                                 continue;
1813                         } else {
1814                                 CERROR("%s: can't alloc page for corruption\n",
1815                                        tgt_name(tgt));
1816                         }
1817                 }
1818         }
1819
1820         bufsize = sizeof(*cksum);
1821         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1822
1823         return 0;
1824 }
1825
1826 char dbgcksum_file_name[PATH_MAX];
1827
1828 static void dump_all_bulk_pages(struct obdo *oa, int count,
1829                                 struct niobuf_local *local_nb,
1830                                 __u32 server_cksum, __u32 client_cksum)
1831 {
1832         struct file *filp;
1833         int rc, i;
1834         unsigned int len;
1835         char *buf;
1836
1837         /* will only keep dump of pages on first error for the same range in
1838          * file/fid, not during the resends/retries. */
1839         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1840                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1841                  (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1842                   libcfs_debug_file_path_arr :
1843                   LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1844                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1845                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1846                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1847                  local_nb[0].lnb_file_offset,
1848                  local_nb[count-1].lnb_file_offset +
1849                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1850         filp = filp_open(dbgcksum_file_name,
1851                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1852         if (IS_ERR(filp)) {
1853                 rc = PTR_ERR(filp);
1854                 if (rc == -EEXIST)
1855                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1856                                "checksum error: rc = %d\n", dbgcksum_file_name,
1857                                rc);
1858                 else
1859                         CERROR("%s: can't open to dump pages with checksum "
1860                                "error: rc = %d\n", dbgcksum_file_name, rc);
1861                 return;
1862         }
1863
1864         for (i = 0; i < count; i++) {
1865                 len = local_nb[i].lnb_len;
1866                 buf = kmap(local_nb[i].lnb_page);
1867                 while (len != 0) {
1868                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1869                         if (rc < 0) {
1870                                 CERROR("%s: wanted to write %u but got %d "
1871                                        "error\n", dbgcksum_file_name, len, rc);
1872                                 break;
1873                         }
1874                         len -= rc;
1875                         buf += rc;
1876                         CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1877                                dbgcksum_file_name, rc);
1878                 }
1879                 kunmap(local_nb[i].lnb_page);
1880         }
1881
1882         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1883         if (rc)
1884                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1885         filp_close(filp, NULL);
1886         return;
1887 }
1888
1889 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1890                                struct obd_export *exp, struct obdo *oa,
1891                                const struct lnet_process_id *peer,
1892                                __u32 client_cksum, __u32 server_cksum,
1893                                enum cksum_types server_cksum_type)
1894 {
1895         char *msg;
1896         enum cksum_types cksum_type;
1897         loff_t start, end;
1898
1899         /* unlikely to happen and only if resend does not occur due to cksum
1900          * control failure on Client */
1901         if (unlikely(server_cksum == client_cksum)) {
1902                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1903                        client_cksum);
1904                 return 0;
1905         }
1906
1907         if (exp->exp_obd->obd_checksum_dump)
1908                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1909                                     client_cksum);
1910
1911         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1912                                            oa->o_flags : 0);
1913
1914         if (cksum_type != server_cksum_type)
1915                 msg = "the server may have not used the checksum type specified"
1916                       " in the original request - likely a protocol problem";
1917         else
1918                 msg = "should have changed on the client or in transit";
1919
1920         start = local_nb[0].lnb_file_offset;
1921         end = local_nb[npages-1].lnb_file_offset +
1922                                         local_nb[npages-1].lnb_len - 1;
1923
1924         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1925                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1926                 " %x (type %x), server csum %x (type %x)\n",
1927                 exp->exp_obd->obd_name,
1928                 msg, libcfs_nid2str(peer->nid),
1929                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1930                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1931                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1932                 POSTID(&oa->o_oi),
1933                 start, end, client_cksum, cksum_type, server_cksum,
1934                 server_cksum_type);
1935
1936         return 1;
1937 }
1938
1939 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1940                              unsigned char *buf, int size)
1941 {
1942         int     i, off, len, copied = size;
1943         char    *ptr;
1944
1945         for (i = 0; i < npages; i++) {
1946                 off = local[i].lnb_page_offset & ~PAGE_MASK;
1947                 len = local[i].lnb_len;
1948
1949                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1950                        i, off, len, size);
1951                 if (len > size)
1952                         return -EINVAL;
1953
1954                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1955                 memcpy(buf + off, ptr, len);
1956                 ll_kunmap_atomic(ptr, KM_USER0);
1957                 buf += len;
1958                 size -= len;
1959         }
1960         return copied - size;
1961 }
1962
1963 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1964                                      struct niobuf_local *local_nb,
1965                                      int npages, int opc,
1966                                      obd_dif_csum_fn *fn,
1967                                      int sector_size,
1968                                      u32 *check_sum)
1969 {
1970         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1971         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1972         const char *obd_name = tgt->lut_obd->obd_name;
1973         struct ahash_request *req;
1974         unsigned int bufsize;
1975         unsigned char *buffer;
1976         struct page *__page;
1977         __u16 *guard_start;
1978         int guard_number;
1979         int used_number = 0;
1980         __u32 cksum;
1981         int rc = 0;
1982         int used;
1983         int i;
1984
1985         __page = alloc_page(GFP_KERNEL);
1986         if (__page == NULL)
1987                 return -ENOMEM;
1988
1989         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1990         if (IS_ERR(req)) {
1991                 CERROR("%s: unable to initialize checksum hash %s\n",
1992                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1993                 return PTR_ERR(req);
1994         }
1995
1996         buffer = kmap(__page);
1997         guard_start = (__u16 *)buffer;
1998         guard_number = PAGE_SIZE / sizeof(*guard_start);
1999         for (i = 0; i < npages; i++) {
2000                 /* corrupt the data before we compute the checksum, to
2001                  * simulate a client->OST data error */
2002                 if (i == 0 && opc == OST_WRITE &&
2003                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2004                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2005                         int len = local_nb[i].lnb_len;
2006                         struct page *np = tgt_page_to_corrupt;
2007
2008                         if (np) {
2009                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2010                                                         KM_USER0);
2011                                 char *ptr2 = page_address(np);
2012
2013                                 memcpy(ptr2 + off, ptr + off, len);
2014                                 memcpy(ptr2 + off, "bad3", min(4, len));
2015                                 ll_kunmap_atomic(ptr, KM_USER0);
2016
2017                                 /* LU-8376 to preserve original index for
2018                                  * display in dump_all_bulk_pages() */
2019                                 np->index = i;
2020
2021                                 cfs_crypto_hash_update_page(req, np, off,
2022                                                             len);
2023                                 continue;
2024                         } else {
2025                                 CERROR("%s: can't alloc page for corruption\n",
2026                                        tgt_name(tgt));
2027                         }
2028                 }
2029
2030                 /*
2031                  * The left guard number should be able to hold checksums of a
2032                  * whole page
2033                  */
2034                 if (t10_cksum_type && opc == OST_READ &&
2035                     local_nb[i].lnb_guard_disk) {
2036                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2037                         if (used > (guard_number - used_number)) {
2038                                 rc = -E2BIG;
2039                                 break;
2040                         }
2041                         memcpy(guard_start + used_number,
2042                                local_nb[i].lnb_guards,
2043                                used * sizeof(*local_nb[i].lnb_guards));
2044                 } else {
2045                         rc = obd_page_dif_generate_buffer(obd_name,
2046                                 local_nb[i].lnb_page,
2047                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2048                                 local_nb[i].lnb_len, guard_start + used_number,
2049                                 guard_number - used_number, &used, sector_size,
2050                                 fn);
2051                         if (rc)
2052                                 break;
2053                 }
2054
2055                 LASSERT(used <= MAX_GUARD_NUMBER);
2056                 /*
2057                  * If disk support T10PI checksum, copy guards to local_nb.
2058                  * If the write is partial page, do not use the guards for bio
2059                  * submission since the data might not be full-sector. The bio
2060                  * guards will be generated later based on the full sectors. If
2061                  * the sector size is 512B rather than 4 KB, or the page size
2062                  * is larger than 4KB, this might drop some useful guards for
2063                  * partial page write, but it will only add minimal extra time
2064                  * of checksum calculation.
2065                  */
2066                 if (t10_cksum_type && opc == OST_WRITE &&
2067                     local_nb[i].lnb_len == PAGE_SIZE) {
2068                         local_nb[i].lnb_guard_rpc = 1;
2069                         memcpy(local_nb[i].lnb_guards,
2070                                guard_start + used_number,
2071                                used * sizeof(*local_nb[i].lnb_guards));
2072                 }
2073
2074                 used_number += used;
2075                 if (used_number == guard_number) {
2076                         cfs_crypto_hash_update_page(req, __page, 0,
2077                                 used_number * sizeof(*guard_start));
2078                         used_number = 0;
2079                 }
2080
2081                  /* corrupt the data after we compute the checksum, to
2082                  * simulate an OST->client data error */
2083                 if (unlikely(i == 0 && opc == OST_READ &&
2084                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2085                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2086                         int len = local_nb[i].lnb_len;
2087                         struct page *np = tgt_page_to_corrupt;
2088
2089                         if (np) {
2090                                 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2091                                                         KM_USER0);
2092                                 char *ptr2 = page_address(np);
2093
2094                                 memcpy(ptr2 + off, ptr + off, len);
2095                                 memcpy(ptr2 + off, "bad4", min(4, len));
2096                                 ll_kunmap_atomic(ptr, KM_USER0);
2097
2098                                 /* LU-8376 to preserve original index for
2099                                  * display in dump_all_bulk_pages() */
2100                                 np->index = i;
2101
2102                                 cfs_crypto_hash_update_page(req, np, off,
2103                                                             len);
2104                                 continue;
2105                         } else {
2106                                 CERROR("%s: can't alloc page for corruption\n",
2107                                        tgt_name(tgt));
2108                         }
2109                 }
2110         }
2111         kunmap(__page);
2112         if (rc)
2113                 GOTO(out, rc);
2114
2115         if (used_number != 0)
2116                 cfs_crypto_hash_update_page(req, __page, 0,
2117                         used_number * sizeof(*guard_start));
2118
2119         bufsize = sizeof(cksum);
2120         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2121
2122         if (rc == 0)
2123                 *check_sum = cksum;
2124 out:
2125         __free_page(__page);
2126         return rc;
2127 }
2128
2129 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2130                                   enum cksum_types cksum_type,
2131                                   struct niobuf_local *local_nb,
2132                                   int npages, int opc, u32 *check_sum)
2133 {
2134         obd_dif_csum_fn *fn = NULL;
2135         int sector_size = 0;
2136         int rc;
2137
2138         ENTRY;
2139         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2140
2141         if (fn)
2142                 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2143                                                opc, fn, sector_size,
2144                                                check_sum);
2145         else
2146                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2147                                          cksum_type, check_sum);
2148         RETURN(rc);
2149 }
2150
2151 int tgt_brw_read(struct tgt_session_info *tsi)
2152 {
2153         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2154         struct ptlrpc_bulk_desc *desc = NULL;
2155         struct obd_export       *exp = tsi->tsi_exp;
2156         struct niobuf_remote    *remote_nb;
2157         struct niobuf_local     *local_nb;
2158         struct obd_ioobj        *ioo;
2159         struct ost_body         *body, *repbody;
2160         struct l_wait_info       lwi;
2161         struct lustre_handle     lockh = { 0 };
2162         int                      npages, nob = 0, rc, i, no_reply = 0,
2163                                  npages_read;
2164         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2165         const char *obd_name = exp->exp_obd->obd_name;
2166
2167         ENTRY;
2168
2169         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2170             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2171                 CERROR("%s: deny read request from %s to portal %u\n",
2172                        tgt_name(tsi->tsi_tgt),
2173                        obd_export_nid2str(req->rq_export),
2174                        ptlrpc_req2svc(req)->srv_req_portal);
2175                 RETURN(-EPROTO);
2176         }
2177
2178         req->rq_bulk_read = 1;
2179
2180         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2181                 RETURN(-EIO);
2182
2183         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2184                          cfs_fail_val : (obd_timeout + 1) / 4);
2185
2186         /* Check if there is eviction in progress, and if so, wait for it to
2187          * finish */
2188         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2189                 /* We do not care how long it takes */
2190                 lwi = LWI_INTR(NULL, NULL);
2191                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2192                          !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2193                          &lwi);
2194         }
2195
2196         /* There must be big cache in current thread to process this request
2197          * if it is NULL then something went wrong and it wasn't allocated,
2198          * report -ENOMEM in that case */
2199         if (tbc == NULL)
2200                 RETURN(-ENOMEM);
2201
2202         body = tsi->tsi_ost_body;
2203         LASSERT(body != NULL);
2204
2205         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2206         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2207
2208         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2209         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2210
2211         local_nb = tbc->local;
2212
2213         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2214                           &lockh, LCK_PR);
2215         if (rc != 0)
2216                 RETURN(rc);
2217
2218         /*
2219          * If getting the lock took more time than
2220          * client was willing to wait, drop it. b=11330
2221          */
2222         if (ktime_get_real_seconds() > req->rq_deadline ||
2223             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2224                 no_reply = 1;
2225                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2226                        libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2227                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2228                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2229                 GOTO(out_lock, rc = -ETIMEDOUT);
2230         }
2231
2232         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2233         repbody->oa = body->oa;
2234
2235         npages = PTLRPC_MAX_BRW_PAGES;
2236         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2237                         ioo, remote_nb, &npages, local_nb);
2238         if (rc != 0)
2239                 GOTO(out_lock, rc);
2240
2241         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2242                 desc = NULL;
2243         } else {
2244                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2245                                             PTLRPC_BULK_PUT_SOURCE |
2246                                                 PTLRPC_BULK_BUF_KIOV,
2247                                             OST_BULK_PORTAL,
2248                                             &ptlrpc_bulk_kiov_nopin_ops);
2249                 if (desc == NULL)
2250                         GOTO(out_commitrw, rc = -ENOMEM);
2251         }
2252
2253         nob = 0;
2254         npages_read = npages;
2255         for (i = 0; i < npages; i++) {
2256                 int page_rc = local_nb[i].lnb_rc;
2257
2258                 if (page_rc < 0) {
2259                         rc = page_rc;
2260                         npages_read = i;
2261                         break;
2262                 }
2263
2264                 nob += page_rc;
2265                 if (page_rc != 0 && desc != NULL) { /* some data! */
2266                         LASSERT(local_nb[i].lnb_page != NULL);
2267                         desc->bd_frag_ops->add_kiov_frag
2268                           (desc, local_nb[i].lnb_page,
2269                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2270                            page_rc);
2271                 }
2272
2273                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2274                         local_nb[i].lnb_len = page_rc;
2275                         npages_read = i + (page_rc != 0 ? 1 : 0);
2276                         /* All subsequent pages should be 0 */
2277                         while (++i < npages)
2278                                 LASSERT(local_nb[i].lnb_rc == 0);
2279                         break;
2280                 }
2281         }
2282         if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2283             nob != cfs_fail_val)
2284                 rc = -E2BIG;
2285
2286         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2287                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2288                            body->oa.o_flags : 0;
2289                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2290
2291                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2292                                                           cksum_type);
2293                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2294
2295                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2296                                             local_nb, npages_read, OST_READ,
2297                                             &repbody->oa.o_cksum);
2298                 if (rc < 0)
2299                         GOTO(out_commitrw, rc);
2300                 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2301                        repbody->oa.o_cksum);
2302
2303                 /* if a resend it could be for a cksum error, so check Server
2304                  * cksum with returned Client cksum (this should even cover
2305                  * zero-cksum case) */
2306                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2307                     (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2308                         check_read_checksum(local_nb, npages_read, exp,
2309                                             &body->oa, &req->rq_peer,
2310                                             body->oa.o_cksum,
2311                                             repbody->oa.o_cksum, cksum_type);
2312         } else {
2313                 repbody->oa.o_valid = 0;
2314         }
2315         /* We're finishing using body->oa as an input variable */
2316
2317         /* Check if client was evicted while we were doing i/o before touching
2318          * network */
2319         if (rc == 0) {
2320                 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2321                         unsigned char *short_io_buf;
2322                         int short_io_size;
2323
2324                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2325                                                               &RMF_SHORT_IO);
2326                         short_io_size = req_capsule_get_size(&req->rq_pill,
2327                                                              &RMF_SHORT_IO,
2328                                                              RCL_SERVER);
2329                         rc = tgt_pages2shortio(local_nb, npages_read,
2330                                                short_io_buf, short_io_size);
2331                         if (rc >= 0)
2332                                 req_capsule_shrink(&req->rq_pill,
2333                                                    &RMF_SHORT_IO, rc,
2334                                                    RCL_SERVER);
2335                         rc = rc > 0 ? 0 : rc;
2336                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2337                         rc = target_bulk_io(exp, desc, &lwi);
2338                 }
2339                 no_reply = rc != 0;
2340         } else {
2341                 if (body->oa.o_flags & OBD_FL_SHORT_IO)
2342                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2343                                            RCL_SERVER);
2344         }
2345
2346 out_commitrw:
2347         /* Must commit after prep above in all cases */
2348         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2349                           remote_nb, npages, local_nb, rc);
2350 out_lock:
2351         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2352
2353         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2354                 ptlrpc_free_bulk(desc);
2355
2356         LASSERT(rc <= 0);
2357         if (rc == 0) {
2358                 rc = nob;
2359                 ptlrpc_lprocfs_brw(req, nob);
2360         } else if (no_reply) {
2361                 req->rq_no_reply = 1;
2362                 /* reply out callback would free */
2363                 ptlrpc_req_drop_rs(req);
2364                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2365                               "client will retry: rc %d\n",
2366                               obd_name,
2367                               obd_uuid2str(&exp->exp_client_uuid),
2368                               obd_export_nid2str(exp), rc);
2369         }
2370         /* send a bulk after reply to simulate a network delay or reordering
2371          * by a router - Note that !desc implies short io, so there is no bulk
2372          * to reorder. */
2373         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2374             desc) {
2375                 wait_queue_head_t        waitq;
2376                 struct l_wait_info       lwi1;
2377
2378                 CDEBUG(D_INFO, "reorder BULK\n");
2379                 init_waitqueue_head(&waitq);
2380
2381                 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2382                 l_wait_event(waitq, 0, &lwi1);
2383                 target_bulk_io(exp, desc, &lwi);
2384                 ptlrpc_free_bulk(desc);
2385         }
2386
2387         RETURN(rc);
2388 }
2389 EXPORT_SYMBOL(tgt_brw_read);
2390
2391 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2392                              unsigned char *buf, unsigned int size)
2393 {
2394         int     i, off, len;
2395         char    *ptr;
2396
2397         for (i = 0; i < npages; i++) {
2398                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2399                 len = local[i].lnb_len;
2400
2401                 if (len == 0)
2402                         continue;
2403
2404                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2405                        i, off, len, size);
2406                 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2407                 if (ptr == NULL)
2408                         return -EINVAL;
2409                 memcpy(ptr + off, buf, len < size ? len : size);
2410                 ll_kunmap_atomic(ptr, KM_USER0);
2411                 buf += len;
2412                 size -= len;
2413         }
2414         return 0;
2415 }
2416
2417 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2418                               struct ptlrpc_bulk_desc *desc,
2419                               struct niobuf_local *local_nb, int npages,
2420                               u32 client_cksum, u32 server_cksum,
2421                               bool mmap)
2422 {
2423         struct obd_export *exp = req->rq_export;
2424         struct ost_body *body;
2425         char *router = "";
2426         char *via = "";
2427
2428         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2429         LASSERT(body != NULL);
2430
2431         if (desc && req->rq_peer.nid != desc->bd_sender) {
2432                 via = " via ";
2433                 router = libcfs_nid2str(desc->bd_sender);
2434         }
2435
2436         if (exp->exp_obd->obd_checksum_dump)
2437                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2438                                     client_cksum);
2439
2440         if (mmap) {
2441                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2442                              client_cksum, server_cksum);
2443                 return;
2444         }
2445
2446         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2447                            DFID" object "DOSTID" extent [%llu-%llu"
2448                            "]: client csum %x, server csum %x\n",
2449                            exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2450                            via, router,
2451                            body->oa.o_valid & OBD_MD_FLFID ?
2452                            body->oa.o_parent_seq : (__u64)0,
2453                            body->oa.o_valid & OBD_MD_FLFID ?
2454                            body->oa.o_parent_oid : 0,
2455                            body->oa.o_valid & OBD_MD_FLFID ?
2456                            body->oa.o_parent_ver : 0,
2457                            POSTID(&body->oa.o_oi),
2458                            local_nb[0].lnb_file_offset,
2459                            local_nb[npages-1].lnb_file_offset +
2460                            local_nb[npages - 1].lnb_len - 1,
2461                            client_cksum, server_cksum);
2462 }
2463
2464 int tgt_brw_write(struct tgt_session_info *tsi)
2465 {
2466         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2467         struct ptlrpc_bulk_desc *desc = NULL;
2468         struct obd_export       *exp = req->rq_export;
2469         struct niobuf_remote    *remote_nb;
2470         struct niobuf_local     *local_nb;
2471         struct obd_ioobj        *ioo;
2472         struct ost_body         *body, *repbody;
2473         struct l_wait_info       lwi;
2474         struct lustre_handle     lockh = {0};
2475         __u32                   *rcs;
2476         int                      objcount, niocount, npages;
2477         int                      rc, i, j;
2478         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2479         bool                     no_reply = false, mmap;
2480         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2481         bool wait_sync = false;
2482         const char *obd_name = exp->exp_obd->obd_name;
2483
2484         ENTRY;
2485
2486         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2487             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2488                 CERROR("%s: deny write request from %s to portal %u\n",
2489                        tgt_name(tsi->tsi_tgt),
2490                        obd_export_nid2str(req->rq_export),
2491                        ptlrpc_req2svc(req)->srv_req_portal);
2492                 RETURN(err_serious(-EPROTO));
2493         }
2494
2495         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2496                 RETURN(err_serious(-ENOSPC));
2497         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2498                 RETURN(err_serious(-EROFS));
2499
2500         req->rq_bulk_write = 1;
2501
2502         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2503                 RETURN(err_serious(-EIO));
2504         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2505                 RETURN(err_serious(-EFAULT));
2506
2507         /* pause before transaction has been started */
2508         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2509                          cfs_fail_val : (obd_timeout + 1) / 4);
2510
2511         /* Delay write commit to show stale size information */
2512         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2513
2514         /* There must be big cache in current thread to process this request
2515          * if it is NULL then something went wrong and it wasn't allocated,
2516          * report -ENOMEM in that case */
2517         if (tbc == NULL)
2518                 RETURN(-ENOMEM);
2519
2520         body = tsi->tsi_ost_body;
2521         LASSERT(body != NULL);
2522
2523         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2524         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2525
2526         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2527                                         RCL_CLIENT) / sizeof(*ioo);
2528
2529         for (niocount = i = 0; i < objcount; i++)
2530                 niocount += ioo[i].ioo_bufcnt;
2531
2532         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2533         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2534         if (niocount != req_capsule_get_size(&req->rq_pill,
2535                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2536                         sizeof(*remote_nb))
2537                 RETURN(err_serious(-EPROTO));
2538
2539         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2540             ptlrpc_connection_is_local(exp->exp_connection))
2541                 memory_pressure_set();
2542
2543         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2544                              niocount * sizeof(*rcs));
2545         rc = req_capsule_server_pack(&req->rq_pill);
2546         if (rc != 0)
2547                 GOTO(out, rc = err_serious(rc));
2548
2549         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2550         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2551
2552         local_nb = tbc->local;
2553
2554         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2555                           &lockh, LCK_PW);
2556         if (rc != 0)
2557                 GOTO(out, rc);
2558
2559         /*
2560          * If getting the lock took more time than
2561          * client was willing to wait, drop it. b=11330
2562          */
2563         if (ktime_get_real_seconds() > req->rq_deadline ||
2564             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2565                 no_reply = true;
2566                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2567                        tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2568                        POSTID(&ioo->ioo_oid),
2569                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2570                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2571                 GOTO(out_lock, rc = -ETIMEDOUT);
2572         }
2573
2574         /* Because we already sync grant info with client when reconnect,
2575          * grant info will be cleared for resent req, then fed_grant and
2576          * total_grant will not be modified in following preprw_write */
2577         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2578                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2579                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2580         }
2581
2582         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2583         if (repbody == NULL)
2584                 GOTO(out_lock, rc = -ENOMEM);
2585         repbody->oa = body->oa;
2586
2587         npages = PTLRPC_MAX_BRW_PAGES;
2588         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2589                         objcount, ioo, remote_nb, &npages, local_nb);
2590         if (rc < 0)
2591                 GOTO(out_lock, rc);
2592         if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2593                 unsigned int short_io_size;
2594                 unsigned char *short_io_buf;
2595
2596                 short_io_size = req_capsule_get_size(&req->rq_pill,
2597                                                      &RMF_SHORT_IO,
2598                                                      RCL_CLIENT);
2599                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2600                                                       &RMF_SHORT_IO);
2601                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2602                                " size = %d\n", short_io_size);
2603
2604                 /* Copy short io buf to pages */
2605                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2606                                        short_io_size);
2607                 desc = NULL;
2608         } else {
2609                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2610                                             PTLRPC_BULK_GET_SINK |
2611                                             PTLRPC_BULK_BUF_KIOV,
2612                                             OST_BULK_PORTAL,
2613                                             &ptlrpc_bulk_kiov_nopin_ops);
2614                 if (desc == NULL)
2615                         GOTO(skip_transfer, rc = -ENOMEM);
2616
2617                 /* NB Having prepped, we must commit... */
2618                 for (i = 0; i < npages; i++)
2619                         desc->bd_frag_ops->add_kiov_frag(desc,
2620                                         local_nb[i].lnb_page,
2621                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2622                                         local_nb[i].lnb_len);
2623
2624                 rc = sptlrpc_svc_prep_bulk(req, desc);
2625                 if (rc != 0)
2626                         GOTO(skip_transfer, rc);
2627
2628                 rc = target_bulk_io(exp, desc, &lwi);
2629         }
2630
2631         no_reply = rc != 0;
2632
2633 skip_transfer:
2634         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2635                 static int cksum_counter;
2636
2637                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2638                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2639
2640                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2641                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2642                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2643                                                            cksum_type);
2644
2645                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2646                                             local_nb, npages, OST_WRITE,
2647                                             &repbody->oa.o_cksum);
2648                 if (rc < 0)
2649                         GOTO(out_commitrw, rc);
2650
2651                 cksum_counter++;
2652
2653                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2654                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2655                                 body->oa.o_flags & OBD_FL_MMAP);
2656
2657                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2658                                           body->oa.o_cksum,
2659                                           repbody->oa.o_cksum, mmap);
2660                         cksum_counter = 0;
2661                 } else if ((cksum_counter & (-cksum_counter)) ==
2662                            cksum_counter) {
2663                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2664                                cksum_counter, libcfs_id2str(req->rq_peer),
2665                                repbody->oa.o_cksum);
2666                 }
2667         }
2668
2669 out_commitrw:
2670         /* Must commit after prep above in all cases */
2671         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2672                           objcount, ioo, remote_nb, npages, local_nb, rc);
2673         if (rc == -ENOTCONN)
2674                 /* quota acquire process has been given up because
2675                  * either the client has been evicted or the client
2676                  * has timed out the request already */
2677                 no_reply = true;
2678
2679         for (i = 0; i < niocount; i++) {
2680                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2681                         wait_sync = true;
2682                         break;
2683                 }
2684         }
2685         /*
2686          * Disable sending mtime back to the client. If the client locked the
2687          * whole object, then it has already updated the mtime on its side,
2688          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2689          */
2690         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2691
2692         if (rc == 0) {
2693                 int nob = 0;
2694
2695                 /* set per-requested niobuf return codes */
2696                 for (i = j = 0; i < niocount; i++) {
2697                         int len = remote_nb[i].rnb_len;
2698
2699                         nob += len;
2700                         rcs[i] = 0;
2701                         do {
2702                                 LASSERT(j < npages);
2703                                 if (local_nb[j].lnb_rc < 0)
2704                                         rcs[i] = local_nb[j].lnb_rc;
2705                                 len -= local_nb[j].lnb_len;
2706                                 j++;
2707                         } while (len > 0);
2708                         LASSERT(len == 0);
2709                 }
2710                 LASSERT(j == npages);
2711                 ptlrpc_lprocfs_brw(req, nob);
2712         }
2713 out_lock:
2714         tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2715         if (desc)
2716                 ptlrpc_free_bulk(desc);
2717 out:
2718         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2719                 req->rq_no_reply = 1;
2720                 /* reply out callback would free */
2721                 ptlrpc_req_drop_rs(req);
2722                 if (!exp->exp_obd->obd_no_transno)
2723                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2724                                       " client will retry: rc = %d\n",
2725                                       obd_name,
2726                                       obd_uuid2str(&exp->exp_client_uuid),
2727                                       obd_export_nid2str(exp), rc);
2728         }
2729         memory_pressure_clr();
2730         RETURN(rc);
2731 }
2732 EXPORT_SYMBOL(tgt_brw_write);
2733
2734 /* Check if request can be reconstructed from saved reply data
2735  * A copy of the reply data is returned in @trd if the pointer is not NULL
2736  */
2737 bool req_can_reconstruct(struct ptlrpc_request *req,
2738                          struct tg_reply_data *trd)
2739 {
2740         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2741         struct lsd_client_data *lcd = ted->ted_lcd;
2742         bool found;
2743
2744         if (tgt_is_multimodrpcs_client(req->rq_export))
2745                 return tgt_lookup_reply(req, trd);
2746
2747         mutex_lock(&ted->ted_lcd_lock);
2748         found = req->rq_xid == lcd->lcd_last_xid ||
2749                 req->rq_xid == lcd->lcd_last_close_xid;
2750
2751         if (found && trd != NULL) {
2752                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2753                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2754                         trd->trd_reply.lrd_transno =
2755                                                 lcd->lcd_last_close_transno;
2756                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2757                 } else {
2758                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2759                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2760                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
2761                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
2762                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2763                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2764                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2765                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2766                 }
2767         }
2768         mutex_unlock(&ted->ted_lcd_lock);
2769
2770         return found;
2771 }
2772 EXPORT_SYMBOL(req_can_reconstruct);
2773