lustre/target/tgt_handler.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2013, 2017, Intel Corporation.
  25  */
  26 /*
  27  * lustre/target/tgt_handler.c
  28  *
  29  * Lustre Unified Target request handler code
  30  *
  31  * Author: Brian Behlendorf <behlendorf1@llnl.gov>
  32  * Author: Mikhail Pershin <mike.pershin@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_CLASS
  36
  37 #include <linux/user_namespace.h>
  38 #include <linux/delay.h>
  39 #include <linux/uidgid.h>
  40
  41 #include <libcfs/linux/linux-mem.h>
  42 #include <obd.h>
  43 #include <obd_class.h>
  44 #include <obd_cksum.h>
  45 #include <lustre_lfsck.h>
  46 #include <lustre_nodemap.h>
  47 #include <lustre_acl.h>
  48
  49 #include "tgt_internal.h"
  50
  51 char *tgt_name(struct lu_target *tgt)
  52 {
  53         LASSERT(tgt->lut_obd != NULL);
  54         return tgt->lut_obd->obd_name;
  55 }
  56 EXPORT_SYMBOL(tgt_name);
  57
  58 /*
  59  * Generic code handling requests that have struct mdt_body passed in:
  60  *
  61  *  - extract mdt_body from request and save it in @tsi, if present;
  62  *
  63  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
  64  *  @tsi;
  65  *
  66  *  - if HAS_BODY flag is set for this request type check whether object
  67  *  actually exists on storage (lu_object_exists()).
  68  *
  69  */
  70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
  71 {
  72         const struct mdt_body   *body;
  73         struct lu_object        *obj;
  74         struct req_capsule      *pill = tsi->tsi_pill;
  75         int                      rc;
  76
  77         ENTRY;
  78
  79         body = req_capsule_client_get(pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EFAULT);
  82
  83         tsi->tsi_mdt_body = body;
  84
  85         if (!(body->mbo_valid & OBD_MD_FLID))
  86                 RETURN(0);
  87
  88         /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
  89          * in any case in pre-2.5 clients. Fix that here if needed */
  90         if (unlikely(fid_is_zero(&body->mbo_fid1)))
  91                 RETURN(0);
  92
  93         if (!fid_is_sane(&body->mbo_fid1)) {
  94                 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
  95                        PFID(&body->mbo_fid1));
  96                 RETURN(-EINVAL);
  97         }
  98
  99         obj = lu_object_find(tsi->tsi_env,
 100                              &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
 101                              &body->mbo_fid1, NULL);
 102         if (!IS_ERR(obj)) {
 103                 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
 104                         lu_object_put(tsi->tsi_env, obj);
 105                         rc = -ENOENT;
 106                 } else {
 107                         tsi->tsi_corpus = obj;
 108                         rc = 0;
 109                 }
 110         } else {
 111                 rc = PTR_ERR(obj);
 112         }
 113
 114         tsi->tsi_fid = body->mbo_fid1;
 115
 116         RETURN(rc);
 117 }
 118
 119 /**
 120  * Validate oa from client.
 121  * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
 122  * req are valid.
 123  *    a. objects in Single MDT FS  seq = FID_SEQ_OST_MDT0, oi_id != 0
 124  *    b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
 125  *       pack ost_id. Because non-zero oi_seq will make it diffcult to tell
 126  *       whether this is oi_fid or real ostid. So it will check
 127  *       OBD_CONNECT_FID, then convert the ostid to FID for old client.
 128  *    c. Old FID-disable osc will send IDIF.
 129  *    d. new FID-enable osc/osp will send normal FID.
 130  *
 131  * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
 132  * be used for LAST_ID file, and only being accessed inside OST now.
 133  */
 134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
 135 {
 136         struct ost_id   *oi     = &oa->o_oi;
 137         u64              seq    = ostid_seq(oi);
 138         u64              id     = ostid_id(oi);
 139         int              rc;
 140         ENTRY;
 141
 142         if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
 143                      fid_seq_is_echo(seq))) {
 144                 /* Sigh 2.[123] client still sends echo req with oi_id = 0
 145                  * during create, and we will reset this to 1, since this
 146                  * oi_id is basically useless in the following create process,
 147                  * but oi_id == 0 will make it difficult to tell whether it is
 148                  * real FID or ost_id. */
 149                 oi->oi_fid.f_seq = FID_SEQ_ECHO;
 150                 oi->oi_fid.f_oid = id ?: 1;
 151                 oi->oi_fid.f_ver = 0;
 152         } else {
 153                 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
 154
 155                 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
 156                         GOTO(out, rc = -EPROTO);
 157
 158                 /* Note: this check might be forced in 2.5 or 2.6, i.e.
 159                  * all of the requests are required to setup FLGROUP */
 160                 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP)))
 161                         oa->o_valid |= OBD_MD_FLGROUP;
 162
 163                 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
 164                                fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
 165                         GOTO(out, rc = -EPROTO);
 166
 167                 rc = ostid_to_fid(&tti->tti_fid1, oi,
 168                                   tsi->tsi_tgt->lut_lsd.lsd_osd_index);
 169                 if (unlikely(rc != 0))
 170                         GOTO(out, rc);
 171
 172                 oi->oi_fid = tti->tti_fid1;
 173         }
 174
 175         RETURN(0);
 176
 177 out:
 178         CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
 179                tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
 180                seq, id, rc);
 181         return rc;
 182 }
 183 EXPORT_SYMBOL(tgt_validate_obdo);
 184
 185 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
 186 {
 187         unsigned                 max_brw;
 188         struct niobuf_remote    *rnb;
 189         struct obd_ioobj        *ioo;
 190         int                      obj_count;
 191
 192         ENTRY;
 193
 194         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
 195         if (ioo == NULL)
 196                 RETURN(-EPROTO);
 197
 198         rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
 199         if (rnb == NULL)
 200                 RETURN(-EPROTO);
 201
 202         max_brw = ioobj_max_brw_get(ioo);
 203         if (unlikely((max_brw & (max_brw - 1)) != 0)) {
 204                 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
 205                        ": rc = %d\n", tgt_name(tsi->tsi_tgt),
 206                        obd_export_nid2str(tsi->tsi_exp), max_brw,
 207                        POSTID(oi), -EPROTO);
 208                 RETURN(-EPROTO);
 209         }
 210         ioo->ioo_oid = *oi;
 211
 212         obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
 213                                         RCL_CLIENT) / sizeof(*ioo);
 214         if (obj_count == 0) {
 215                 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
 216                 RETURN(-EPROTO);
 217         } else if (obj_count > 1) {
 218                 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
 219                        obj_count);
 220                 RETURN(-EPROTO);
 221         }
 222
 223         if (ioo->ioo_bufcnt == 0) {
 224                 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
 225                 RETURN(-EPROTO);
 226         }
 227
 228         if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
 229                 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
 230                           "bulk has too many pages (%d)",
 231                           ioo->ioo_bufcnt);
 232                 RETURN(-EPROTO);
 233         }
 234
 235         RETURN(0);
 236 }
 237
 238 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
 239 {
 240         struct ost_body         *body;
 241         struct req_capsule      *pill = tsi->tsi_pill;
 242         struct lu_nodemap       *nodemap;
 243         int                      rc;
 244
 245         ENTRY;
 246
 247         body = req_capsule_client_get(pill, &RMF_OST_BODY);
 248         if (body == NULL)
 249                 RETURN(-EFAULT);
 250
 251         rc = tgt_validate_obdo(tsi, &body->oa);
 252         if (rc)
 253                 RETURN(rc);
 254
 255         nodemap = nodemap_get_from_exp(tsi->tsi_exp);
 256         if (IS_ERR(nodemap))
 257                 RETURN(PTR_ERR(nodemap));
 258
 259         body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
 260                                         NODEMAP_CLIENT_TO_FS,
 261                                         body->oa.o_uid);
 262         body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
 263                                         NODEMAP_CLIENT_TO_FS,
 264                                         body->oa.o_gid);
 265         body->oa.o_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
 266                                            NODEMAP_CLIENT_TO_FS,
 267                                            body->oa.o_projid);
 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                     req_capsule_field_present(pill, &RMF_MDT_BODY,
 325                                               RCL_CLIENT)) {
 326                         rc = tgt_mdt_body_unpack(tsi, flags);
 327                         if (rc < 0)
 328                                 RETURN(rc);
 329                 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
 330                                                  RCL_CLIENT) &&
 331                            req_capsule_field_present(pill, &RMF_OST_BODY,
 332                                                  RCL_CLIENT)) {
 333                         rc = tgt_ost_body_unpack(tsi, flags);
 334                         if (rc < 0)
 335                                 RETURN(rc);
 336                 }
 337         }
 338
 339         if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
 340                 RETURN(-EROFS);
 341
 342         if (flags & HAS_KEY) {
 343                 struct ldlm_request *dlm_req;
 344
 345                 LASSERT(h->th_fmt != NULL);
 346
 347                 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
 348                 if (dlm_req != NULL) {
 349                         union ldlm_wire_policy_data *policy =
 350                                         &dlm_req->lock_desc.l_policy_data;
 351
 352                         if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 353                                      LDLM_IBITS &&
 354                                      (policy->l_inodebits.bits |
 355                                       policy->l_inodebits.try_bits) == 0)) {
 356                                 /*
 357                                  * Lock without inodebits makes no sense and
 358                                  * will oops later in ldlm. If client miss to
 359                                  * set such bits, do not trigger ASSERTION.
 360                                  *
 361                                  * For liblustre flock case, it maybe zero.
 362                                  */
 363                                 rc = -EPROTO;
 364                         } else {
 365                                 tsi->tsi_dlm_req = dlm_req;
 366                         }
 367                 } else {
 368                         rc = -EFAULT;
 369                 }
 370         }
 371         tsi->tsi_preprocessed = 1;
 372         RETURN(rc);
 373 }
 374
 375 /*
 376  * Invoke handler for this request opc. Also do necessary preprocessing
 377  * (according to handler ->th_flags), and post-processing (setting of
 378  * ->last_{xid,committed}).
 379  */
 380 static int tgt_handle_request0(struct tgt_session_info *tsi,
 381                                struct tgt_handler *h,
 382                                struct ptlrpc_request *req)
 383 {
 384         int      serious = 0;
 385         int      rc;
 386         __u32    opc = lustre_msg_get_opc(req->rq_reqmsg);
 387
 388         ENTRY;
 389
 390
 391         /* When dealing with sec context requests, no export is associated yet,
 392          * because these requests are sent before *_CONNECT requests.
 393          * A NULL req->rq_export means the normal *_common_slice handlers will
 394          * not be called, because there is no reference to the target.
 395          * So deal with them by hand and jump directly to target_send_reply().
 396          */
 397         switch (opc) {
 398         case SEC_CTX_INIT:
 399         case SEC_CTX_INIT_CONT:
 400         case SEC_CTX_FINI:
 401                 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
 402                 GOTO(out, rc = 0);
 403         }
 404
 405         /*
 406          * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
 407          * to put same checks into handlers like mdt_close(), mdt_reint(),
 408          * etc., without talking to mdt authors first. Checking same thing
 409          * there again is useless and returning 0 error without packing reply
 410          * is buggy! Handlers either pack reply or return error.
 411          *
 412          * We return 0 here and do not send any reply in order to emulate
 413          * network failure. Do not send any reply in case any of NET related
 414          * fail_id has occured.
 415          */
 416         if (CFS_FAIL_CHECK_ORSET(h->th_fail_id, CFS_FAIL_ONCE))
 417                 RETURN(0);
 418         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 419                      CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
 420                 RETURN(0);
 421
 422         /* drop OUT_UPDATE rpc */
 423         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == OUT_UPDATE &&
 424                      CFS_FAIL_CHECK(OBD_FAIL_OUT_UPDATE_DROP)))
 425                 RETURN(0);
 426
 427         rc = tgt_request_preprocess(tsi, h, req);
 428         /* pack reply if reply format is fixed */
 429         if (rc == 0 && h->th_flags & HAS_REPLY) {
 430                 /* Pack reply */
 431                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
 432                                           RCL_SERVER))
 433                         req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
 434                                              RCL_SERVER,
 435                                              tsi->tsi_mdt_body->mbo_eadatasize);
 436                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
 437                                           RCL_SERVER))
 438                         req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
 439                                              RCL_SERVER, 0);
 440                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
 441                         req_capsule_set_size(tsi->tsi_pill,
 442                                              &RMF_ACL, RCL_SERVER,
 443                                              LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
 444
 445                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
 446                                           RCL_SERVER)) {
 447                         struct niobuf_remote *remote_nb =
 448                                 req_capsule_client_get(tsi->tsi_pill,
 449                                                        &RMF_NIOBUF_REMOTE);
 450                         struct ost_body *body = tsi->tsi_ost_body;
 451
 452                         req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
 453                                          RCL_SERVER,
 454                                          (body->oa.o_valid & OBD_MD_FLFLAGS &&
 455                                           body->oa.o_flags & OBD_FL_SHORT_IO) ?
 456                                          remote_nb[0].rnb_len : 0);
 457                 }
 458                 if (req_capsule_has_field(tsi->tsi_pill, &RMF_FILE_ENCCTX,
 459                                           RCL_SERVER))
 460                         req_capsule_set_size(tsi->tsi_pill, &RMF_FILE_ENCCTX,
 461                                              RCL_SERVER, 0);
 462
 463                 rc = req_capsule_server_pack(tsi->tsi_pill);
 464         }
 465
 466         if (likely(rc == 0)) {
 467                 /*
 468                  * Process request, there can be two types of rc:
 469                  * 1) errors with msg unpack/pack, other failures outside the
 470                  * operation itself. This is counted as serious errors;
 471                  * 2) errors during fs operation, should be placed in rq_status
 472                  * only
 473                  */
 474                 rc = h->th_act(tsi);
 475                 if (!is_serious(rc) &&
 476                     !req->rq_no_reply && req->rq_reply_state == NULL) {
 477                         DEBUG_REQ(D_ERROR, req,
 478                                   "%s: %s handler did not pack reply but returned no error",
 479                                   tgt_name(tsi->tsi_tgt), h->th_name);
 480                         LBUG();
 481                 }
 482                 serious = is_serious(rc);
 483                 rc = clear_serious(rc);
 484         } else {
 485                 serious = 1;
 486         }
 487
 488         req->rq_status = rc;
 489
 490         /*
 491          * ELDLM_* codes which > 0 should be in rq_status only as well as
 492          * all non-serious errors.
 493          */
 494         if (rc > 0 || !serious)
 495                 rc = 0;
 496
 497         LASSERT(current->journal_info == NULL);
 498
 499         if (likely(rc == 0 && req->rq_export))
 500                 target_committed_to_req(req);
 501
 502 out:
 503         target_send_reply(req, rc, tsi->tsi_reply_fail_id);
 504         RETURN(0);
 505 }
 506
 507 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
 508                                        struct obd_device *obd, int *process)
 509 {
 510         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 511         case MDS_DISCONNECT:
 512         case OST_DISCONNECT:
 513         case OBD_IDX_READ:
 514                 *process = 1;
 515                 RETURN(0);
 516         case MDS_CLOSE:
 517         case MDS_SYNC: /* used in unmounting */
 518         case OBD_PING:
 519         case MDS_REINT:
 520         case OUT_UPDATE:
 521         case MDS_BATCH:
 522         case SEQ_QUERY:
 523         case FLD_QUERY:
 524         case FLD_READ:
 525         case LDLM_ENQUEUE:
 526         case OST_CREATE:
 527         case OST_DESTROY:
 528         case OST_PUNCH:
 529         case OST_SETATTR:
 530         case OST_SYNC:
 531         case OST_WRITE:
 532         case MDS_HSM_PROGRESS:
 533         case MDS_HSM_STATE_SET:
 534         case MDS_HSM_REQUEST:
 535         case OST_FALLOCATE:
 536                 *process = target_queue_recovery_request(req, obd);
 537                 RETURN(0);
 538
 539         default:
 540                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 541                 *process = -EAGAIN;
 542                 RETURN(0);
 543         }
 544 }
 545
 546 /*
 547  * Handle recovery. Return:
 548  *        +1: continue request processing;
 549  *       -ve: abort immediately with the given error code;
 550  *         0: send reply with error code in req->rq_status;
 551  */
 552 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
 553 {
 554         ENTRY;
 555
 556         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
 557         case MDS_CONNECT:
 558         case OST_CONNECT:
 559         case MGS_CONNECT:
 560         case SEC_CTX_INIT:
 561         case SEC_CTX_INIT_CONT:
 562         case SEC_CTX_FINI:
 563                 RETURN(+1);
 564         }
 565
 566         if (!req->rq_export->exp_obd->obd_replayable)
 567                 RETURN(+1);
 568
 569         /* sanity check: if the xid matches, the request must be marked as a
 570          * resent or replayed */
 571         if (req_can_reconstruct(req, NULL) == 1) {
 572                 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
 573                       (MSG_RESENT | MSG_REPLAY))) {
 574                         DEBUG_REQ(D_WARNING, req,
 575                                   "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
 576                                   req->rq_xid,
 577                                   lustre_msg_get_flags(req->rq_reqmsg));
 578                         req->rq_status = -ENOTCONN;
 579                         RETURN(-ENOTCONN);
 580                 }
 581         }
 582         /* else: note the opposite is not always true; a RESENT req after a
 583          * failover will usually not match the last_xid, since it was likely
 584          * never committed. A REPLAYed request will almost never match the
 585          * last xid, however it could for a committed, but still retained,
 586          * open. */
 587
 588         /* Check for aborted recovery... */
 589         if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
 590                 int rc;
 591                 int should_process;
 592
 593                 DEBUG_REQ(D_INFO, req, "Got new replay");
 594                 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
 595                                                  &should_process);
 596                 if (rc != 0 || !should_process)
 597                         RETURN(rc);
 598                 else if (should_process < 0) {
 599                         req->rq_status = should_process;
 600                         rc = ptlrpc_error(req);
 601                         RETURN(rc);
 602                 }
 603         }
 604         RETURN(+1);
 605 }
 606
 607 /* Initial check for request, it is validation mostly */
 608 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
 609 {
 610         struct tgt_handler      *h;
 611         struct tgt_opc_slice    *s;
 612         struct lu_target        *tgt;
 613         __u32                    opc = lustre_msg_get_opc(req->rq_reqmsg);
 614         /* don't spew error messages for unhandled RPCs */
 615         static bool              printed;
 616
 617         ENTRY;
 618
 619         tgt = class_exp2tgt(req->rq_export);
 620         if (unlikely(tgt == NULL)) {
 621                 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
 622                           class_exp2obd(req->rq_export)->obd_name);
 623                 RETURN(ERR_PTR(-EINVAL));
 624         }
 625
 626         for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
 627                 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
 628                         break;
 629
 630         /* opcode was not found in slice */
 631         if (unlikely(s->tos_hs == NULL)) {
 632                 if (!printed) {
 633                         CERROR("%s: no handler for opcode 0x%x from %s\n",
 634                                tgt_name(tgt), opc, libcfs_idstr(&req->rq_peer));
 635                         printed = true;
 636                 }
 637                 goto err_unsupported;
 638         }
 639
 640         LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
 641         h = s->tos_hs + (opc - s->tos_opc_start);
 642         if (unlikely(h->th_opc == 0)) {
 643                 if (!printed) {
 644                         CERROR("%s: unsupported opcode 0x%x\n",
 645                                tgt_name(tgt), opc);
 646                         printed = true;
 647                 }
 648                 goto err_unsupported;
 649         }
 650
 651         if (CFS_FAIL_CHECK(OBD_FAIL_OST_OPCODE) && opc == cfs_fail_val)
 652                 goto err_unsupported;
 653
 654         RETURN(h);
 655 err_unsupported:
 656         /*
 657          * Unknown opcode does not necessarily means insane client. A new
 658          * client might send RPCs with new opcodes to an old server. The
 659          * client might desperately stuck there waiting for a reply. So,
 660          * send an error back here.
 661          *
 662          * An old client might also send RPCs with deprecated opcodes (e.g.
 663          * OST_QUOTACHECK).
 664          *
 665          * Error in ptlrpc_send_error() is ignored.
 666          */
 667         req->rq_status = -EOPNOTSUPP;
 668         ptlrpc_send_error(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
 669         RETURN(ERR_PTR(-EOPNOTSUPP));
 670 }
 671
 672 static int process_req_last_xid(struct ptlrpc_request *req)
 673 {
 674         __u64   last_xid;
 675         int rc = 0;
 676         struct obd_export *exp = req->rq_export;
 677         struct tg_export_data *ted = &exp->exp_target_data;
 678         bool need_lock = tgt_is_multimodrpcs_client(exp);
 679         ENTRY;
 680
 681         if (need_lock)
 682                 mutex_lock(&ted->ted_lcd_lock);
 683         /* check request's xid is consistent with export's last_xid */
 684         last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
 685         if (last_xid > exp->exp_last_xid)
 686                 exp->exp_last_xid = last_xid;
 687
 688         if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
 689                 /* Some request is allowed to be sent during replay,
 690                  * such as OUT update requests, FLD requests, so it
 691                  * is possible that replay requests has smaller XID
 692                  * than the exp_last_xid.
 693                  *
 694                  * Some non-replay requests may have smaller XID as
 695                  * well:
 696                  *
 697                  * - Client send a no_resend RPC, like statfs;
 698                  * - The RPC timedout (or some other error) on client,
 699                  *   then it's removed from the unreplied list;
 700                  * - Client send some other request to bump the
 701                  *   exp_last_xid on server;
 702                  * - The former RPC got chance to be processed;
 703                  */
 704                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 705                         rc = -EPROTO;
 706
 707                 DEBUG_REQ(D_WARNING, req,
 708                           "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
 709                           req->rq_xid, exp->exp_last_xid, rc);
 710                 if (rc)
 711                         GOTO(out, rc);
 712         }
 713
 714         /* The "last_xid" is the minimum xid among unreplied requests,
 715          * if the request is from the previous connection, its xid can
 716          * still be larger than "exp_last_xid", then the above check of
 717          * xid is not enough to determine whether the request is delayed.
 718          *
 719          * For example, if some replay request was delayed and caused
 720          * timeout at client and the replay is restarted, the delayed
 721          * replay request will have the larger xid than "exp_last_xid"
 722          */
 723         if (req->rq_export->exp_conn_cnt >
 724             lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 725                 CDEBUG(D_RPCTRACE,
 726                        "Dropping request %llu from an old epoch %u/%u\n",
 727                        req->rq_xid,
 728                        lustre_msg_get_conn_cnt(req->rq_reqmsg),
 729                        req->rq_export->exp_conn_cnt);
 730                 req->rq_no_reply = 1;
 731                 GOTO(out, rc = -ESTALE);
 732         }
 733
 734         /* try to release in-memory reply data */
 735         if (tgt_is_multimodrpcs_client(exp)) {
 736                 tgt_handle_received_xid(exp, last_xid);
 737                 rc = tgt_handle_tag(req);
 738         }
 739
 740 out:
 741         if (need_lock)
 742                 mutex_unlock(&ted->ted_lcd_lock);
 743
 744         RETURN(rc);
 745 }
 746
 747 int tgt_request_handle(struct ptlrpc_request *req)
 748 {
 749         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 750
 751         struct lustre_msg       *msg = req->rq_reqmsg;
 752         struct tgt_handler      *h;
 753         struct lu_target        *tgt;
 754         int                      request_fail_id = 0;
 755         __u32                    opc = lustre_msg_get_opc(msg);
 756         struct obd_device       *obd;
 757         int                      rc;
 758         bool                     is_connect = false;
 759         ENTRY;
 760
 761         if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
 762                 if (cfs_fail_val == 0 &&
 763                     lustre_msg_get_opc(msg) != OBD_PING &&
 764                     lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
 765                         cfs_fail_val = 1;
 766                         cfs_race_state = 0;
 767                         wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
 768                 }
 769         }
 770
 771         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 772         tsi->tsi_pill = &req->rq_pill;
 773         tsi->tsi_env = req->rq_svc_thread->t_env;
 774
 775         /* if request has export then get handlers slice from corresponding
 776          * target, otherwise that should be connect operation */
 777         if (opc == MDS_CONNECT || opc == OST_CONNECT ||
 778             opc == MGS_CONNECT) {
 779                 is_connect = true;
 780                 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
 781                 rc = target_handle_connect(req);
 782                 if (rc != 0) {
 783                         rc = ptlrpc_error(req);
 784                         GOTO(out, rc);
 785                 }
 786                 /* recovery-small test 18c asks to drop connect reply */
 787                 if (unlikely(opc == OST_CONNECT &&
 788                              CFS_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
 789                         GOTO(out, rc = 0);
 790         }
 791
 792         if (unlikely(!class_connected_export(req->rq_export))) {
 793                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
 794                     opc == SEC_CTX_FINI) {
 795                         /* sec context initialization has to be handled
 796                          * by hand in tgt_handle_request0() */
 797                         tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
 798                         h = NULL;
 799                         GOTO(handle_recov, rc = 0);
 800                 }
 801                 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
 802                        opc, libcfs_idstr(&req->rq_peer));
 803                 req->rq_status = -ENOTCONN;
 804                 rc = ptlrpc_error(req);
 805                 GOTO(out, rc);
 806         }
 807
 808         tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
 809         tsi->tsi_exp = req->rq_export;
 810         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
 811                 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 812         else
 813                 tsi->tsi_jobid = NULL;
 814
 815         if (tgt == NULL) {
 816                 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
 817                           class_exp2obd(req->rq_export)->obd_name);
 818                 req->rq_status = -EINVAL;
 819                 rc = ptlrpc_error(req);
 820                 GOTO(out, rc);
 821         }
 822
 823         /* Skip last_xid processing for the recovery thread, otherwise, the
 824          * last_xid on same request could be processed twice: first time when
 825          * processing the incoming request, second time when the request is
 826          * being processed by recovery thread. */
 827         obd = class_exp2obd(req->rq_export);
 828         if (is_connect) {
 829                 /* reset the exp_last_xid on each connection. */
 830                 req->rq_export->exp_last_xid = 0;
 831         } else if (obd->obd_recovery_data.trd_processing_task !=
 832                    current->pid) {
 833                 rc = process_req_last_xid(req);
 834                 if (rc) {
 835                         req->rq_status = rc;
 836                         rc = ptlrpc_error(req);
 837                         GOTO(out, rc);
 838                 }
 839         }
 840
 841         request_fail_id = tgt->lut_request_fail_id;
 842         tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
 843
 844         h = tgt_handler_find_check(req);
 845         if (IS_ERR(h)) {
 846                 req->rq_status = PTR_ERR(h);
 847                 rc = ptlrpc_error(req);
 848                 GOTO(out, rc);
 849         }
 850
 851         LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 852                  h->th_opc, opc);
 853
 854         if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
 855              CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
 856                 GOTO(out, rc = 0);
 857
 858         rc = lustre_msg_check_version(msg, h->th_version);
 859         if (unlikely(rc)) {
 860                 DEBUG_REQ(D_ERROR, req,
 861                           "%s: drop malformed request version=%08x expect=%08x",
 862                           tgt_name(tgt), lustre_msg_get_version(msg),
 863                           h->th_version);
 864                 req->rq_status = -EINVAL;
 865                 rc = ptlrpc_error(req);
 866                 GOTO(out, rc);
 867         }
 868
 869 handle_recov:
 870         rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
 871         if (likely(rc == 1)) {
 872                 rc = tgt_handle_request0(tsi, h, req);
 873                 if (rc)
 874                         GOTO(out, rc);
 875         }
 876         EXIT;
 877 out:
 878         req_capsule_fini(tsi->tsi_pill);
 879         if (tsi->tsi_corpus != NULL) {
 880                 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
 881                 tsi->tsi_corpus = NULL;
 882         }
 883         return rc;
 884 }
 885 EXPORT_SYMBOL(tgt_request_handle);
 886
 887 /** Assign high priority operations to the request if needed. */
 888 int tgt_hpreq_handler(struct ptlrpc_request *req)
 889 {
 890         struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
 891         struct tgt_handler      *h;
 892         int                      rc;
 893
 894         ENTRY;
 895
 896         if (req->rq_export == NULL)
 897                 RETURN(0);
 898
 899         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
 900         tsi->tsi_pill = &req->rq_pill;
 901         tsi->tsi_env = req->rq_svc_thread->t_env;
 902         tsi->tsi_tgt = class_exp2tgt(req->rq_export);
 903         tsi->tsi_exp = req->rq_export;
 904
 905         h = tgt_handler_find_check(req);
 906         if (IS_ERR(h)) {
 907                 rc = PTR_ERR(h);
 908                 RETURN(rc);
 909         }
 910
 911         rc = tgt_request_preprocess(tsi, h, req);
 912         if (unlikely(rc != 0))
 913                 RETURN(rc);
 914
 915         if (h->th_hp != NULL)
 916                 h->th_hp(tsi);
 917         RETURN(0);
 918 }
 919 EXPORT_SYMBOL(tgt_hpreq_handler);
 920
 921 void tgt_counter_incr(struct obd_export *exp, int opcode)
 922 {
 923         lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
 924         if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
 925                 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
 926 }
 927 EXPORT_SYMBOL(tgt_counter_incr);
 928
 929 /*
 930  * Unified target generic handlers.
 931  */
 932
 933 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
 934 {
 935         struct lu_target *tgt = class_exp2tgt(exp);
 936         struct sptlrpc_flavor flvr;
 937         int rc = 0;
 938
 939         LASSERT(tgt);
 940         LASSERT(tgt->lut_obd);
 941         LASSERT(tgt->lut_slice);
 942
 943         /* always allow ECHO client */
 944         if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
 945                             LUSTRE_ECHO_NAME) == 0)) {
 946                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 947                 return 0;
 948         }
 949
 950         if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
 951                 read_lock(&tgt->lut_sptlrpc_lock);
 952                 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
 953                                              req->rq_sp_from,
 954                                              &req->rq_peer.nid,
 955                                              &flvr);
 956                 read_unlock(&tgt->lut_sptlrpc_lock);
 957
 958                 spin_lock(&exp->exp_lock);
 959                 exp->exp_sp_peer = req->rq_sp_from;
 960                 exp->exp_flvr = flvr;
 961
 962                 /* when on mgs, if no restriction is set, or if the client
 963                  * NID is on the local node, allow any flavor
 964                  */
 965                 if ((strcmp(exp->exp_obd->obd_type->typ_name,
 966                             LUSTRE_MGS_NAME) == 0) &&
 967                     (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
 968                      LNetIsPeerLocal(&exp->exp_connection->c_peer.nid)))
 969                         exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
 970
 971                 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
 972                     exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
 973                         CERROR("%s: unauthorized rpc flavor %x from %s, "
 974                                "expect %x\n", tgt_name(tgt),
 975                                req->rq_flvr.sf_rpc,
 976                                libcfs_nidstr(&req->rq_peer.nid),
 977                                exp->exp_flvr.sf_rpc);
 978                         rc = -EACCES;
 979                 }
 980                 spin_unlock(&exp->exp_lock);
 981         } else {
 982                 if (exp->exp_sp_peer != req->rq_sp_from) {
 983                         CERROR("%s: RPC source %s doesn't match %s\n",
 984                                tgt_name(tgt),
 985                                sptlrpc_part2name(req->rq_sp_from),
 986                                sptlrpc_part2name(exp->exp_sp_peer));
 987                         rc = -EACCES;
 988                 } else {
 989                         rc = sptlrpc_target_export_check(exp, req);
 990                 }
 991         }
 992
 993         return rc;
 994 }
 995
 996 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
 997 {
 998         struct sptlrpc_rule_set  tmp_rset;
 999         int                      rc;
1000
1001         if (unlikely(tgt == NULL)) {
1002                 CERROR("No target passed\n");
1003                 return -EINVAL;
1004         }
1005
1006         sptlrpc_rule_set_init(&tmp_rset);
1007         rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
1008         if (rc) {
1009                 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
1010                        tgt_name(tgt), rc);
1011                 return rc;
1012         }
1013
1014         sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
1015
1016         write_lock(&tgt->lut_sptlrpc_lock);
1017         sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
1018         tgt->lut_sptlrpc_rset = tmp_rset;
1019         write_unlock(&tgt->lut_sptlrpc_lock);
1020
1021         return 0;
1022 }
1023 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
1024
1025 int tgt_connect(struct tgt_session_info *tsi)
1026 {
1027         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1028         struct obd_connect_data *reply;
1029         int                      rc;
1030
1031         ENTRY;
1032
1033         /* XXX: better to call this check right after getting new export but
1034          * before last_rcvd slot allocation to avoid server load upon insecure
1035          * connects. This is to be fixed after unifiyng all targets.
1036          */
1037         rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1038         if (rc)
1039                 GOTO(out, rc);
1040
1041         /* To avoid exposing partially initialized connection flags, changes up
1042          * to this point have been staged in reply->ocd_connect_flags. Now that
1043          * connection handling has completed successfully, atomically update
1044          * the connect flags in the shared export data structure. LU-1623 */
1045         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1046         spin_lock(&tsi->tsi_exp->exp_lock);
1047         *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1048         if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1049                 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1050                         reply->ocd_connect_flags2;
1051         tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1052         spin_unlock(&tsi->tsi_exp->exp_lock);
1053
1054         if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1055                    LUSTRE_MDT_NAME) == 0) {
1056                 struct lu_nodemap *nm = NULL;
1057
1058                 rc = req_check_sepol(tsi->tsi_pill);
1059                 if (rc)
1060                         GOTO(out, rc);
1061
1062                 if (tsi->tsi_pill->rc_req->rq_export)
1063                         nm =
1064                          nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1065
1066                 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1067                     reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT) {
1068                         bool forbid_encrypt = true;
1069
1070                         if (!nm)
1071                                 /* nodemap_get_from_exp returns NULL in case
1072                                  * nodemap is not active, so we do not forbid
1073                                  */
1074                                 forbid_encrypt = false;
1075                         else if (!IS_ERR(nm))
1076                                 forbid_encrypt = nm->nmf_forbid_encryption;
1077                         if (forbid_encrypt)
1078                                 GOTO(put_nm, rc = -EACCES);
1079                 }
1080
1081                 if (!(reply->ocd_connect_flags & OBD_CONNECT_RDONLY)) {
1082                         bool readonly = false;
1083
1084                         if (!IS_ERR_OR_NULL(nm))
1085                                 readonly = nm->nmf_readonly_mount;
1086                         if (unlikely(readonly))
1087                                 GOTO(put_nm, rc = -EROFS);
1088                 }
1089
1090 put_nm:
1091                 if (!IS_ERR_OR_NULL(nm))
1092                         nodemap_putref(nm);
1093                 if (rc)
1094                         GOTO(out, rc);
1095         }
1096
1097         RETURN(0);
1098 out:
1099         obd_disconnect(class_export_get(tsi->tsi_exp));
1100         return rc;
1101 }
1102 EXPORT_SYMBOL(tgt_connect);
1103
1104 int tgt_disconnect(struct tgt_session_info *tsi)
1105 {
1106         int rc;
1107
1108         ENTRY;
1109
1110         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1111
1112         rc = target_handle_disconnect(tgt_ses_req(tsi));
1113         if (rc)
1114                 RETURN(err_serious(rc));
1115
1116         RETURN(rc);
1117 }
1118 EXPORT_SYMBOL(tgt_disconnect);
1119
1120 /*
1121  * Unified target OBD handlers
1122  */
1123 int tgt_obd_ping(struct tgt_session_info *tsi)
1124 {
1125         int rc;
1126
1127         ENTRY;
1128
1129         /* The target-specific part of OBD_PING request handling.
1130          * It controls Filter Modification Data (FMD) expiration each time
1131          * PING is received.
1132          *
1133          * Valid only for replayable targets, e.g. MDT and OFD
1134          */
1135         if (tsi->tsi_exp->exp_obd->obd_replayable)
1136                 tgt_fmd_expire(tsi->tsi_exp);
1137
1138         rc = req_capsule_server_pack(tsi->tsi_pill);
1139         if (rc)
1140                 RETURN(err_serious(rc));
1141
1142         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_VS_EVICT)) {
1143                 if (strstr(tsi->tsi_exp->exp_obd->obd_name, "MDT0000") &&
1144                     (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_MDS_MDS))
1145                         tsi->tsi_pill->rc_req->rq_no_reply = 1;
1146         }
1147
1148         RETURN(rc);
1149 }
1150 EXPORT_SYMBOL(tgt_obd_ping);
1151
1152 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1153 {
1154         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1155         struct obd_export       *exp = req->rq_export;
1156         struct ptlrpc_bulk_desc *desc;
1157         int                      i;
1158         int                      rc;
1159         int                      pages = 0;
1160
1161         ENTRY;
1162
1163         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1164                 unsigned int offset;
1165
1166                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1167                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1168                                       PAGE_SIZE);
1169         }
1170
1171         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1172                                   PTLRPC_BULK_PUT_SOURCE,
1173                                     MDS_BULK_PORTAL,
1174                                     &ptlrpc_bulk_kiov_nopin_ops);
1175         if (desc == NULL)
1176                 RETURN(-ENOMEM);
1177
1178         for (i = 0; i < rdbuf->rb_nbufs; i++)
1179                 desc->bd_frag_ops->add_iov_frag(desc,
1180                                         rdbuf->rb_bufs[i].lb_buf,
1181                                         rdbuf->rb_bufs[i].lb_len);
1182
1183         rc = target_bulk_io(exp, desc);
1184         ptlrpc_free_bulk(desc);
1185         RETURN(rc);
1186 }
1187 EXPORT_SYMBOL(tgt_send_buffer);
1188
1189 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1190 {
1191         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1192         struct obd_export       *exp = req->rq_export;
1193         struct ptlrpc_bulk_desc *desc;
1194         int                      tmpcount;
1195         int                      tmpsize;
1196         int                      i;
1197         int                      rc;
1198
1199         ENTRY;
1200
1201         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1202                                     PTLRPC_BULK_PUT_SOURCE,
1203                                     MDS_BULK_PORTAL,
1204                                     &ptlrpc_bulk_kiov_pin_ops);
1205         if (desc == NULL)
1206                 RETURN(-ENOMEM);
1207
1208         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1209                 /* old client requires reply size in it's PAGE_SIZE,
1210                  * which is rdpg->rp_count */
1211                 nob = rdpg->rp_count;
1212
1213         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1214              i++, tmpcount -= tmpsize) {
1215                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1216                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1217                                                  tmpsize);
1218         }
1219
1220         LASSERT(desc->bd_nob == nob);
1221         rc = target_bulk_io(exp, desc);
1222         ptlrpc_free_bulk(desc);
1223         RETURN(rc);
1224 }
1225 EXPORT_SYMBOL(tgt_sendpage);
1226
1227 /*
1228  * OBD_IDX_READ handler
1229  */
1230 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1231 {
1232         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1233         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1234         struct idx_info         *req_ii, *rep_ii;
1235         int                      rc, i;
1236
1237         ENTRY;
1238
1239         memset(rdpg, 0, sizeof(*rdpg));
1240         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1241
1242         /* extract idx_info buffer from request & reply */
1243         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1244         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1245                 RETURN(err_serious(-EPROTO));
1246
1247         rc = req_capsule_server_pack(tsi->tsi_pill);
1248         if (rc)
1249                 RETURN(err_serious(rc));
1250
1251         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1252         if (rep_ii == NULL)
1253                 RETURN(err_serious(-EFAULT));
1254         rep_ii->ii_magic = IDX_INFO_MAGIC;
1255
1256         /* extract hash to start with */
1257         rdpg->rp_hash = req_ii->ii_hash_start;
1258
1259         /* extract requested attributes */
1260         rdpg->rp_attrs = req_ii->ii_attrs;
1261
1262         /* check that fid packed in request is valid and supported */
1263         if (!fid_is_sane(&req_ii->ii_fid))
1264                 RETURN(-EINVAL);
1265         rep_ii->ii_fid = req_ii->ii_fid;
1266
1267         /* copy flags */
1268         rep_ii->ii_flags = req_ii->ii_flags;
1269
1270         /* compute number of pages to allocate, ii_count is the number of 4KB
1271          * containers */
1272         if (req_ii->ii_count <= 0)
1273                 GOTO(out, rc = -EFAULT);
1274         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1275                                exp_max_brw_size(tsi->tsi_exp));
1276         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1277
1278         /* allocate pages to store the containers */
1279         OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1280         if (rdpg->rp_pages == NULL)
1281                 GOTO(out, rc = -ENOMEM);
1282         for (i = 0; i < rdpg->rp_npages; i++) {
1283                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1284                 if (rdpg->rp_pages[i] == NULL)
1285                         GOTO(out, rc = -ENOMEM);
1286         }
1287
1288         /* populate pages with key/record pairs */
1289         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1290         if (rc < 0)
1291                 GOTO(out, rc);
1292
1293         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1294                  "asked %d > %d\n", rc, rdpg->rp_count);
1295
1296         /* send pages to client */
1297         rc = tgt_sendpage(tsi, rdpg, rc);
1298         if (rc)
1299                 GOTO(out, rc);
1300         EXIT;
1301 out:
1302         if (rdpg->rp_pages) {
1303                 for (i = 0; i < rdpg->rp_npages; i++)
1304                         if (rdpg->rp_pages[i])
1305                                 __free_page(rdpg->rp_pages[i]);
1306                 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1307         }
1308         return rc;
1309 }
1310
1311 struct tgt_handler tgt_obd_handlers[] = {
1312 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1313 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1314 };
1315 EXPORT_SYMBOL(tgt_obd_handlers);
1316
1317 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1318              struct dt_object *obj, __u64 start, __u64 end)
1319 {
1320         int rc = 0;
1321
1322         ENTRY;
1323
1324         /* if no objid is specified, it means "sync whole filesystem" */
1325         if (obj == NULL) {
1326                 rc = dt_sync(env, tgt->lut_bottom);
1327         } else if (dt_version_get(env, obj) >
1328                    tgt->lut_obd->obd_last_committed) {
1329                 rc = dt_object_sync(env, obj, start, end);
1330         }
1331         atomic_inc(&tgt->lut_sync_count);
1332
1333         RETURN(rc);
1334 }
1335 EXPORT_SYMBOL(tgt_sync);
1336 /*
1337  * Unified target DLM handlers.
1338  */
1339
1340 /**
1341  * Unified target BAST
1342  *
1343  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1344  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1345  * but if it's cross-MDT lock, because remote object version is not set, a
1346  * filesystem sync is needed.
1347  *
1348  * \param lock server side lock
1349  * \param desc lock desc
1350  * \param data ldlm_cb_set_arg
1351  * \param flag  indicates whether this cancelling or blocking callback
1352  * \retval      0 on success
1353  * \retval      negative number on error
1354  */
1355 int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1356                      void *data, int flag)
1357 {
1358         struct lu_env            env;
1359         struct lu_target        *tgt;
1360         struct dt_object        *obj = NULL;
1361         struct lu_fid            fid;
1362         int                      rc = 0;
1363
1364         ENTRY;
1365
1366         tgt = class_exp2tgt(lock->l_export);
1367
1368         if (unlikely(tgt == NULL)) {
1369                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1370                        class_exp2obd(lock->l_export)->obd_name);
1371                 RETURN(-EINVAL);
1372         }
1373
1374         if (flag == LDLM_CB_CANCELING &&
1375             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1376             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1377              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1378               ldlm_is_cbpending(lock))) &&
1379             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1380              lock->l_resource->lr_type == LDLM_EXTENT)) {
1381                 __u64 start = 0;
1382                 __u64 end = OBD_OBJECT_EOF;
1383
1384                 rc = lu_env_init(&env, LCT_DT_THREAD);
1385                 if (unlikely(rc != 0))
1386                         GOTO(err, rc);
1387
1388                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1389                                    tgt->lut_lsd.lsd_osd_index);
1390
1391                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1392                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1393                         if (IS_ERR(obj))
1394                                 GOTO(err_env, rc = PTR_ERR(obj));
1395
1396                         if (!dt_object_exists(obj))
1397                                 GOTO(err_put, rc = -ENOENT);
1398
1399                         start = lock->l_policy_data.l_extent.start;
1400                         end = lock->l_policy_data.l_extent.end;
1401                 }
1402
1403                 rc = tgt_sync(&env, tgt, obj, start, end);
1404                 if (rc < 0) {
1405                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1406                                "cancel: rc = %d\n",
1407                                tgt_name(tgt), PFID(&fid),
1408                                lock->l_policy_data.l_extent.start,
1409                                lock->l_policy_data.l_extent.end, rc);
1410                 }
1411 err_put:
1412                 if (obj != NULL)
1413                         dt_object_put(&env, obj);
1414 err_env:
1415                 lu_env_fini(&env);
1416         }
1417 err:
1418         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1419         RETURN(rc);
1420 }
1421 EXPORT_SYMBOL(tgt_blocking_ast);
1422
1423 static struct ldlm_callback_suite tgt_dlm_cbs = {
1424         .lcs_completion = ldlm_server_completion_ast,
1425         .lcs_blocking   = tgt_blocking_ast,
1426         .lcs_glimpse    = ldlm_server_glimpse_ast
1427 };
1428
1429 int tgt_enqueue(struct tgt_session_info *tsi)
1430 {
1431         struct ptlrpc_request *req = tgt_ses_req(tsi);
1432         int rc;
1433
1434         ENTRY;
1435         /*
1436          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1437          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1438          */
1439         LASSERT(tsi->tsi_dlm_req != NULL);
1440         rc = ldlm_handle_enqueue(tsi->tsi_exp->exp_obd->obd_namespace,
1441                                  &req->rq_pill, tsi->tsi_dlm_req, &tgt_dlm_cbs);
1442         if (rc)
1443                 RETURN(err_serious(rc));
1444
1445         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1446         case LUT_FAIL_MDT:
1447                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1448                 break;
1449         case LUT_FAIL_OST:
1450                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1451                 break;
1452         case LUT_FAIL_MGT:
1453                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1454                 break;
1455         default:
1456                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1457                 break;
1458         }
1459         RETURN(req->rq_status);
1460 }
1461 EXPORT_SYMBOL(tgt_enqueue);
1462
1463 int tgt_convert(struct tgt_session_info *tsi)
1464 {
1465         struct ptlrpc_request *req = tgt_ses_req(tsi);
1466         int rc;
1467
1468         ENTRY;
1469         LASSERT(tsi->tsi_dlm_req);
1470         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1471         if (rc)
1472                 RETURN(err_serious(rc));
1473
1474         RETURN(req->rq_status);
1475 }
1476
1477 int tgt_bl_callback(struct tgt_session_info *tsi)
1478 {
1479         return err_serious(-EOPNOTSUPP);
1480 }
1481
1482 int tgt_cp_callback(struct tgt_session_info *tsi)
1483 {
1484         return err_serious(-EOPNOTSUPP);
1485 }
1486
1487 /* generic LDLM target handler */
1488 struct tgt_handler tgt_dlm_handlers[] = {
1489 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1490 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1491 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1492 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1493 };
1494 EXPORT_SYMBOL(tgt_dlm_handlers);
1495
1496 /*
1497  * Unified target LLOG handlers.
1498  */
1499 int tgt_llog_open(struct tgt_session_info *tsi)
1500 {
1501         int rc;
1502
1503         ENTRY;
1504
1505         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1506
1507         RETURN(rc);
1508 }
1509 EXPORT_SYMBOL(tgt_llog_open);
1510
1511 int tgt_llog_read_header(struct tgt_session_info *tsi)
1512 {
1513         int rc;
1514
1515         ENTRY;
1516
1517         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1518
1519         RETURN(rc);
1520 }
1521 EXPORT_SYMBOL(tgt_llog_read_header);
1522
1523 int tgt_llog_next_block(struct tgt_session_info *tsi)
1524 {
1525         int rc;
1526
1527         ENTRY;
1528
1529         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1530
1531         RETURN(rc);
1532 }
1533 EXPORT_SYMBOL(tgt_llog_next_block);
1534
1535 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1536 {
1537         int rc;
1538
1539         ENTRY;
1540
1541         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1542
1543         RETURN(rc);
1544 }
1545 EXPORT_SYMBOL(tgt_llog_prev_block);
1546
1547 /* generic llog target handler */
1548 struct tgt_handler tgt_llog_handlers[] = {
1549 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1550 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1551 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1552 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1553 };
1554 EXPORT_SYMBOL(tgt_llog_handlers);
1555
1556 /*
1557  * sec context handlers
1558  */
1559 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1560 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1561 {
1562         return 0;
1563 }
1564
1565 struct tgt_handler tgt_sec_ctx_handlers[] = {
1566 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1567 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1568 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1569 };
1570 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1571
1572 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1573                                  struct dt_device *key,
1574                                  struct lfsck_req_local *lrl,
1575                                  struct thandle *th) = NULL;
1576
1577 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1578                                                       struct dt_device *,
1579                                                       struct lfsck_req_local *,
1580                                                       struct thandle *))
1581 {
1582         tgt_lfsck_in_notify_local = notify;
1583 }
1584 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1585
1586 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1587                            struct dt_device *key,
1588                            struct lfsck_request *lr) = NULL;
1589
1590 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1591                                                 struct dt_device *,
1592                                                 struct lfsck_request *))
1593 {
1594         tgt_lfsck_in_notify = notify;
1595 }
1596 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1597
1598 static int (*tgt_lfsck_query)(const struct lu_env *env,
1599                               struct dt_device *key,
1600                               struct lfsck_request *req,
1601                               struct lfsck_reply *rep,
1602                               struct lfsck_query *que) = NULL;
1603
1604 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1605                                            struct dt_device *,
1606                                            struct lfsck_request *,
1607                                            struct lfsck_reply *,
1608                                            struct lfsck_query *))
1609 {
1610         tgt_lfsck_query = query;
1611 }
1612 EXPORT_SYMBOL(tgt_register_lfsck_query);
1613
1614 /* LFSCK request handlers */
1615 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1616 {
1617         const struct lu_env     *env = tsi->tsi_env;
1618         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1619         struct lfsck_request    *lr;
1620         int                      rc;
1621         ENTRY;
1622
1623         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1624         if (lr == NULL)
1625                 RETURN(-EPROTO);
1626
1627         rc = tgt_lfsck_in_notify(env, key, lr);
1628
1629         RETURN(rc);
1630 }
1631
1632 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1633 {
1634         struct lfsck_request    *request;
1635         struct lfsck_reply      *reply;
1636         int                      rc;
1637         ENTRY;
1638
1639         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1640         if (request == NULL)
1641                 RETURN(-EPROTO);
1642
1643         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1644         if (reply == NULL)
1645                 RETURN(-ENOMEM);
1646
1647         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1648                              request, reply, NULL);
1649
1650         RETURN(rc < 0 ? rc : 0);
1651 }
1652
1653 struct tgt_handler tgt_lfsck_handlers[] = {
1654 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1655 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1656 };
1657 EXPORT_SYMBOL(tgt_lfsck_handlers);
1658
1659 /*
1660  * initialize per-thread page pool (bug 5137).
1661  */
1662 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1663 {
1664         struct tgt_thread_big_cache *tbc;
1665
1666         ENTRY;
1667
1668         LASSERT(thread != NULL);
1669         LASSERT(thread->t_data == NULL);
1670
1671         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1672         if (tbc == NULL)
1673                 RETURN(-ENOMEM);
1674         thread->t_data = tbc;
1675         RETURN(0);
1676 }
1677 EXPORT_SYMBOL(tgt_io_thread_init);
1678
1679 /*
1680  * free per-thread pool created by tgt_thread_init().
1681  */
1682 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1683 {
1684         struct tgt_thread_big_cache *tbc;
1685
1686         ENTRY;
1687
1688         LASSERT(thread != NULL);
1689
1690         /*
1691          * be prepared to handle partially-initialized pools (because this is
1692          * called from ost_io_thread_init() for cleanup.
1693          */
1694         tbc = thread->t_data;
1695         if (tbc != NULL) {
1696                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1697                 thread->t_data = NULL;
1698         }
1699         EXIT;
1700 }
1701 EXPORT_SYMBOL(tgt_io_thread_done);
1702
1703 /**
1704  * Helper function for getting Data-on-MDT file server DLM lock
1705  * if asked by client.
1706  */
1707 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1708                       struct lustre_handle *lh, int mode, __u64 *flags)
1709 {
1710         union ldlm_policy_data policy = {
1711                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1712         };
1713         int rc;
1714
1715         ENTRY;
1716
1717         LASSERT(lh != NULL);
1718         LASSERT(ns != NULL);
1719         LASSERT(!lustre_handle_is_used(lh));
1720
1721         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1722                                     flags, ldlm_blocking_ast,
1723                                     ldlm_completion_ast, ldlm_glimpse_ast,
1724                                     NULL, 0, LVB_T_NONE, NULL, lh);
1725
1726         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1727 }
1728 EXPORT_SYMBOL(tgt_mdt_data_lock);
1729
1730 /**
1731  * Helper function for getting server side [start, start+count] DLM lock
1732  * if asked by client.
1733  */
1734 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1735                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1736                     struct lustre_handle *lh, int mode, __u64 *flags)
1737 {
1738         union ldlm_policy_data policy;
1739         int rc;
1740
1741         ENTRY;
1742
1743         LASSERT(lh != NULL);
1744         LASSERT(ns != NULL);
1745         LASSERT(!lustre_handle_is_used(lh));
1746
1747         policy.l_extent.gid = 0;
1748         policy.l_extent.start = start & PAGE_MASK;
1749
1750         /*
1751          * If ->o_blocks is EOF it means "lock till the end of the file".
1752          * Otherwise, it's size of an extent or hole being punched (in bytes).
1753          */
1754         if (end == OBD_OBJECT_EOF || end < start)
1755                 policy.l_extent.end = OBD_OBJECT_EOF;
1756         else
1757                 policy.l_extent.end = end | ~PAGE_MASK;
1758
1759         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1760                                     mode, flags, ldlm_blocking_ast,
1761                                     ldlm_completion_ast, ldlm_glimpse_ast,
1762                                     NULL, 0, LVB_T_NONE, NULL, lh);
1763         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1764 }
1765 EXPORT_SYMBOL(tgt_extent_lock);
1766
1767 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1768                          struct ldlm_res_id *res_id, __u64 start, __u64 end,
1769                          struct lustre_handle *lh, enum ldlm_mode mode)
1770 {
1771         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1772         __u64 flags = 0;
1773
1774         /* MDT IO for data-on-mdt */
1775         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1776                 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1777
1778         return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1779 }
1780
1781 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1782 {
1783         LASSERT(lustre_handle_is_used(lh));
1784         ldlm_lock_decref(lh, mode);
1785 }
1786 EXPORT_SYMBOL(tgt_data_unlock);
1787
1788 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1789                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1790                         struct niobuf_remote *nb, struct lustre_handle *lh,
1791                         enum ldlm_mode mode)
1792 {
1793         int nrbufs = obj->ioo_bufcnt;
1794         int i;
1795
1796         ENTRY;
1797
1798         LASSERT(mode == LCK_PR || mode == LCK_PW);
1799         LASSERT(!lustre_handle_is_used(lh));
1800
1801         if (exp->exp_obd->obd_recovering)
1802                 RETURN(0);
1803
1804         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1805                 RETURN(0);
1806
1807         for (i = 1; i < nrbufs; i++)
1808                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1809                         RETURN(-EFAULT);
1810
1811         return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1812                              nb[nrbufs - 1].rnb_offset +
1813                              nb[nrbufs - 1].rnb_len - 1, lh, mode);
1814 }
1815
1816 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1817                            struct niobuf_remote *niob,
1818                            struct lustre_handle *lh, enum ldlm_mode mode)
1819 {
1820         ENTRY;
1821
1822         LASSERT(mode == LCK_PR || mode == LCK_PW);
1823         LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1824                  niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1825                 lustre_handle_is_used(lh));
1826
1827         if (lustre_handle_is_used(lh))
1828                 tgt_data_unlock(lh, mode);
1829         EXIT;
1830 }
1831
1832 static int tgt_checksum_niobuf(struct lu_target *tgt,
1833                                  struct niobuf_local *local_nb, int npages,
1834                                  int opc, enum cksum_types cksum_type,
1835                                  __u32 *cksum)
1836 {
1837         struct ahash_request           *req;
1838         unsigned int                    bufsize;
1839         int                             i, err;
1840         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1841
1842         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1843         if (IS_ERR(req)) {
1844                 CERROR("%s: unable to initialize checksum hash %s\n",
1845                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1846                 return PTR_ERR(req);
1847         }
1848
1849         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1850         for (i = 0; i < npages; i++) {
1851                 /* corrupt the data before we compute the checksum, to
1852                  * simulate a client->OST data error */
1853                 if (i == 0 && opc == OST_WRITE &&
1854                     CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1855                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1856                         int len = local_nb[i].lnb_len;
1857                         struct page *np = tgt_page_to_corrupt;
1858
1859                         if (np) {
1860                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1861                                 char *ptr2 = page_address(np);
1862
1863                                 memcpy(ptr2 + off, ptr + off, len);
1864                                 memcpy(ptr2 + off, "bad3", min(4, len));
1865                                 kunmap_atomic(ptr);
1866
1867                                 /* LU-8376 to preserve original index for
1868                                  * display in dump_all_bulk_pages() */
1869                                 np->index = i;
1870
1871                                 cfs_crypto_hash_update_page(req, np, off,
1872                                                             len);
1873                                 continue;
1874                         } else {
1875                                 CERROR("%s: can't alloc page for corruption\n",
1876                                        tgt_name(tgt));
1877                         }
1878                 }
1879                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1880                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1881                                   local_nb[i].lnb_len);
1882
1883                  /* corrupt the data after we compute the checksum, to
1884                  * simulate an OST->client data error */
1885                 if (i == 0 && opc == OST_READ &&
1886                     CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1887                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1888                         int len = local_nb[i].lnb_len;
1889                         struct page *np = tgt_page_to_corrupt;
1890
1891                         if (np) {
1892                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1893                                 char *ptr2 = page_address(np);
1894
1895                                 memcpy(ptr2 + off, ptr + off, len);
1896                                 memcpy(ptr2 + off, "bad4", min(4, len));
1897                                 kunmap_atomic(ptr);
1898
1899                                 /* LU-8376 to preserve original index for
1900                                  * display in dump_all_bulk_pages() */
1901                                 np->index = i;
1902
1903                                 cfs_crypto_hash_update_page(req, np, off,
1904                                                             len);
1905                                 continue;
1906                         } else {
1907                                 CERROR("%s: can't alloc page for corruption\n",
1908                                        tgt_name(tgt));
1909                         }
1910                 }
1911         }
1912
1913         bufsize = sizeof(*cksum);
1914         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1915
1916         return 0;
1917 }
1918
1919 char dbgcksum_file_name[PATH_MAX];
1920
1921 static void dump_all_bulk_pages(struct obdo *oa, int count,
1922                                 struct niobuf_local *local_nb,
1923                                 __u32 server_cksum, __u32 client_cksum)
1924 {
1925         struct file *filp;
1926         int rc, i;
1927         unsigned int len;
1928         char *buf;
1929
1930         /* will only keep dump of pages on first error for the same range in
1931          * file/fid, not during the resends/retries. */
1932         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1933                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1934                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1935                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1936                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1937                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1938                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1939                  local_nb[0].lnb_file_offset,
1940                  local_nb[count-1].lnb_file_offset +
1941                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1942         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1943         filp = filp_open(dbgcksum_file_name,
1944                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1945         if (IS_ERR(filp)) {
1946                 rc = PTR_ERR(filp);
1947                 if (rc == -EEXIST)
1948                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1949                                "checksum error: rc = %d\n", dbgcksum_file_name,
1950                                rc);
1951                 else
1952                         CERROR("%s: can't open to dump pages with checksum "
1953                                "error: rc = %d\n", dbgcksum_file_name, rc);
1954                 return;
1955         }
1956
1957         for (i = 0; i < count; i++) {
1958                 len = local_nb[i].lnb_len;
1959                 buf = kmap(local_nb[i].lnb_page);
1960                 while (len != 0) {
1961                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1962                         if (rc < 0) {
1963                                 CERROR("%s: wanted to write %u but got %d "
1964                                        "error\n", dbgcksum_file_name, len, rc);
1965                                 break;
1966                         }
1967                         len -= rc;
1968                         buf += rc;
1969                 }
1970                 kunmap(local_nb[i].lnb_page);
1971         }
1972
1973         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1974         if (rc)
1975                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1976         filp_close(filp, NULL);
1977
1978         libcfs_debug_dumplog();
1979 }
1980
1981 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1982                                struct obd_export *exp, struct obdo *oa,
1983                                const struct lnet_processid *peer,
1984                                __u32 client_cksum, __u32 server_cksum,
1985                                enum cksum_types server_cksum_type)
1986 {
1987         char *msg;
1988         enum cksum_types cksum_type;
1989         loff_t start, end;
1990
1991         if (unlikely(npages <= 0))
1992                 return 0;
1993
1994         /* unlikely to happen and only if resend does not occur due to cksum
1995          * control failure on Client */
1996         if (unlikely(server_cksum == client_cksum)) {
1997                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1998                        client_cksum);
1999                 return 0;
2000         }
2001
2002         if (exp->exp_obd->obd_checksum_dump)
2003                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
2004                                     client_cksum);
2005
2006         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
2007                                            oa->o_flags : 0);
2008
2009         if (cksum_type != server_cksum_type)
2010                 msg = "the server may have not used the checksum type specified"
2011                       " in the original request - likely a protocol problem";
2012         else
2013                 msg = "should have changed on the client or in transit";
2014
2015         start = local_nb[0].lnb_file_offset;
2016         end = local_nb[npages-1].lnb_file_offset +
2017                                         local_nb[npages-1].lnb_len - 1;
2018
2019         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
2020                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
2021                 " %x (type %x), server csum %x (type %x)\n",
2022                 exp->exp_obd->obd_name,
2023                 msg, libcfs_nidstr(&peer->nid),
2024                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
2025                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
2026                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
2027                 POSTID(&oa->o_oi),
2028                 start, end, client_cksum, cksum_type, server_cksum,
2029                 server_cksum_type);
2030
2031         return 1;
2032 }
2033
2034 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
2035                              unsigned char *buf, int size)
2036 {
2037         int     i, off, len, copied = size;
2038         char    *ptr;
2039
2040         for (i = 0; i < npages; i++) {
2041                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2042                 len = local[i].lnb_len;
2043
2044                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2045                        i, off, len, size);
2046                 if (len > size)
2047                         return -EINVAL;
2048
2049                 ptr = kmap_atomic(local[i].lnb_page);
2050                 memcpy(buf, ptr + off, len);
2051                 kunmap_atomic(ptr);
2052                 buf += len;
2053                 size -= len;
2054         }
2055         return copied - size;
2056 }
2057
2058 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2059                                      enum cksum_types cksum_type,
2060                                      struct niobuf_local *local_nb, int npages,
2061                                      int opc, obd_dif_csum_fn *fn,
2062                                      int sector_size, u32 *check_sum,
2063                                      bool resend)
2064 {
2065         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2066         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2067         const char *obd_name = tgt->lut_obd->obd_name;
2068         struct ahash_request *req;
2069         unsigned char *buffer;
2070         struct page *__page;
2071         __be16 *guard_start;
2072         int guard_number;
2073         int used_number = 0;
2074         __u32 cksum;
2075         unsigned int bufsize = sizeof(cksum);
2076         int rc = 0, rc2;
2077         int used;
2078         int i;
2079
2080         __page = alloc_page(GFP_KERNEL);
2081         if (__page == NULL)
2082                 return -ENOMEM;
2083
2084         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2085         if (IS_ERR(req)) {
2086                 rc = PTR_ERR(req);
2087                 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
2088                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg), rc);
2089                 goto out;
2090         }
2091
2092         buffer = kmap(__page);
2093         guard_start = (__be16 *)buffer;
2094         guard_number = PAGE_SIZE / sizeof(*guard_start);
2095         if (unlikely(resend))
2096                 CDEBUG(D_PAGE | D_HA, "GRD tags per page = %u\n", guard_number);
2097         for (i = 0; i < npages; i++) {
2098                 bool use_t10_grd;
2099                 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2100                 int len = local_nb[i].lnb_len;
2101                 int guards_needed = DIV_ROUND_UP(off + len, sector_size) -
2102                                         (off / sector_size);
2103
2104                 if (guards_needed > guard_number - used_number) {
2105                         cfs_crypto_hash_update_page(req, __page, 0,
2106                                 used_number * sizeof(*guard_start));
2107                         used_number = 0;
2108                 }
2109
2110                 /* corrupt the data before we compute the checksum, to
2111                  * simulate a client->OST data error */
2112                 if (i == 0 && opc == OST_WRITE &&
2113                     CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2114                         struct page *np = tgt_page_to_corrupt;
2115
2116                         if (np) {
2117                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2118                                 char *ptr2 = page_address(np);
2119
2120                                 memcpy(ptr2 + off, ptr + off, len);
2121                                 memcpy(ptr2 + off, "bad3", min(4, len));
2122                                 kunmap_atomic(ptr);
2123
2124                                 /* LU-8376 to preserve original index for
2125                                  * display in dump_all_bulk_pages() */
2126                                 np->index = i;
2127
2128                                 cfs_crypto_hash_update_page(req, np, off,
2129                                                             len);
2130                                 continue;
2131                         } else {
2132                                 CERROR("%s: can't alloc page for corruption\n",
2133                                        tgt_name(tgt));
2134                         }
2135                 }
2136
2137                 /*
2138                  * The left guard number should be able to hold checksums of a
2139                  * whole page
2140                  */
2141                 use_t10_grd = t10_cksum_type && t10_cksum_type == cksum_type &&
2142                               opc == OST_READ &&
2143                               local_nb[i].lnb_len == PAGE_SIZE &&
2144                               local_nb[i].lnb_guard_disk;
2145                 if (use_t10_grd) {
2146                         used = guards_needed;
2147                         if (used > (guard_number - used_number)) {
2148                                 rc = -E2BIG;
2149                                 CDEBUG(D_PAGE | D_HA,
2150                                        "%s: used %u, guard %u/%u, data size %u+%u, sector_size %u: rc = %d\n",
2151                                        obd_name, used, guard_number,
2152                                        used_number, local_nb[i].lnb_page_offset,
2153                                        local_nb[i].lnb_len, sector_size, rc);
2154                                 break;
2155                         }
2156                         memcpy(guard_start + used_number,
2157                                local_nb[i].lnb_guards,
2158                                used * sizeof(*guard_start));
2159                         if (unlikely(resend))
2160                                 CDEBUG(D_PAGE | D_HA,
2161                                        "lnb[%u]: used %u off %u+%u lnb checksum: %*phN\n",
2162                                        i, used,
2163                                        local_nb[i].lnb_page_offset,
2164                                        local_nb[i].lnb_len,
2165                                        (int)(used * sizeof(*guard_start)),
2166                                        guard_start + used_number);
2167                 }
2168                 if (!use_t10_grd || unlikely(resend)) {
2169                         __be16 guard_tmp[MAX_GUARD_NUMBER];
2170                         __be16 *guards = guard_start + used_number;
2171                         int used_tmp = -1, *usedp = &used;
2172
2173                         if (unlikely(use_t10_grd)) {
2174                                 guards = guard_tmp;
2175                                 usedp = &used_tmp;
2176                         }
2177                         rc = obd_page_dif_generate_buffer(obd_name,
2178                                 local_nb[i].lnb_page,
2179                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2180                                 local_nb[i].lnb_len, guards,
2181                                 guard_number - used_number, usedp, sector_size,
2182                                 fn);
2183                         if (unlikely(resend)) {
2184                                 bool bad = use_t10_grd &&
2185                                         memcmp(guard_tmp,
2186                                                local_nb[i].lnb_guards,
2187                                                used_tmp * sizeof(*guard_tmp));
2188
2189                                 if (bad)
2190                                         CERROR("lnb[%u]: used %u/%u off %u+%u tmp checksum: %*phN\n",
2191                                                i, used, used_tmp,
2192                                                local_nb[i].lnb_page_offset,
2193                                                local_nb[i].lnb_len,
2194                                                (int)(used_tmp * sizeof(*guard_start)),
2195                                                guard_tmp);
2196                                 CDEBUG_LIMIT(D_PAGE | D_HA | (bad ? D_ERROR : 0),
2197                                        "lnb[%u]: used %u/%u off %u+%u gen checksum: %*phN\n",
2198                                        i, used, used_tmp,
2199                                        local_nb[i].lnb_page_offset,
2200                                        local_nb[i].lnb_len,
2201                                        (int)(used * sizeof(*guard_start)),
2202                                        guard_start + used_number);
2203                         }
2204                         if (rc)
2205                                 break;
2206                 }
2207
2208                 LASSERT(used <= MAX_GUARD_NUMBER);
2209                 /*
2210                  * If disk support T10PI checksum, copy guards to local_nb.
2211                  * If the write is partial page, do not use the guards for bio
2212                  * submission since the data might not be full-sector. The bio
2213                  * guards will be generated later based on the full sectors. If
2214                  * the sector size is 512B rather than 4 KB, or the page size
2215                  * is larger than 4KB, this might drop some useful guards for
2216                  * partial page write, but it will only add minimal extra time
2217                  * of checksum calculation.
2218                  */
2219                 if (t10_cksum_type && t10_cksum_type == cksum_type &&
2220                     opc == OST_WRITE &&
2221                     local_nb[i].lnb_len == PAGE_SIZE) {
2222                         local_nb[i].lnb_guard_rpc = 1;
2223                         memcpy(local_nb[i].lnb_guards,
2224                                guard_start + used_number,
2225                                used * sizeof(*local_nb[i].lnb_guards));
2226                 }
2227
2228                 used_number += used;
2229
2230                  /* corrupt the data after we compute the checksum, to
2231                  * simulate an OST->client data error */
2232                 if (unlikely(i == 0 && opc == OST_READ &&
2233                              CFS_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2234                         struct page *np = tgt_page_to_corrupt;
2235
2236                         if (np) {
2237                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2238                                 char *ptr2 = page_address(np);
2239
2240                                 memcpy(ptr2 + off, ptr + off, len);
2241                                 memcpy(ptr2 + off, "bad4", min(4, len));
2242                                 kunmap_atomic(ptr);
2243
2244                                 /* LU-8376 to preserve original index for
2245                                  * display in dump_all_bulk_pages() */
2246                                 np->index = i;
2247
2248                                 cfs_crypto_hash_update_page(req, np, off,
2249                                                             len);
2250                                 continue;
2251                         } else {
2252                                 CERROR("%s: can't alloc page for corruption\n",
2253                                        tgt_name(tgt));
2254                         }
2255                 }
2256         }
2257         kunmap(__page);
2258         if (rc)
2259                 GOTO(out_hash, rc);
2260
2261         if (used_number != 0)
2262                 cfs_crypto_hash_update_page(req, __page, 0,
2263                                             used_number * sizeof(*guard_start));
2264 out_hash:
2265         rc2 = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2266         if (!rc)
2267                 rc = rc2;
2268         if (rc == 0)
2269                 *check_sum = cksum;
2270 out:
2271         __free_page(__page);
2272         return rc;
2273 }
2274
2275 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2276                                   enum cksum_types cksum_type,
2277                                   struct niobuf_local *local_nb,
2278                                   int npages, int opc, u32 *check_sum,
2279                                   bool resend)
2280 {
2281         obd_dif_csum_fn *fn = NULL;
2282         int sector_size = 0;
2283         int rc;
2284
2285         ENTRY;
2286         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2287
2288         if (fn)
2289                 rc = tgt_checksum_niobuf_t10pi(tgt, cksum_type,
2290                                                local_nb, npages,
2291                                                opc, fn, sector_size,
2292                                                check_sum, resend);
2293         else
2294                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2295                                          cksum_type, check_sum);
2296
2297         RETURN(rc);
2298 }
2299
2300 int tgt_brw_read(struct tgt_session_info *tsi)
2301 {
2302         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2303         struct ptlrpc_bulk_desc *desc = NULL;
2304         struct obd_export       *exp = tsi->tsi_exp;
2305         struct niobuf_remote    *remote_nb;
2306         struct niobuf_local     *local_nb;
2307         struct obd_ioobj        *ioo;
2308         struct ost_body         *body, *repbody;
2309         struct lustre_handle     lockh = { 0 };
2310         int                      npages, nob = 0, rc, i, no_reply = 0,
2311                                  npages_read;
2312         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2313         const char *obd_name = exp->exp_obd->obd_name;
2314         ktime_t kstart;
2315
2316         ENTRY;
2317
2318         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2319             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2320                 CERROR("%s: deny read request from %s to portal %u\n",
2321                        tgt_name(tsi->tsi_tgt),
2322                        obd_export_nid2str(req->rq_export),
2323                        ptlrpc_req2svc(req)->srv_req_portal);
2324                 RETURN(-EPROTO);
2325         }
2326
2327         req->rq_bulk_read = 1;
2328
2329         if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) {
2330                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2331                  * this server to fail requests.
2332                  */
2333                 char fail_ost_name[MAX_OBD_NAME];
2334
2335                 if (cfs_fail_val > 0) {
2336                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2337                                  cfs_fail_val - 1);
2338
2339                         if (strstr(obd_name, fail_ost_name))
2340                                 RETURN(err_serious(-EIO));
2341                 } else {
2342                         RETURN(err_serious(-EIO));
2343                 }
2344         }
2345
2346         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2347                          cfs_fail_val : (obd_timeout + 1) / 4);
2348
2349         /* There must be big cache in current thread to process this request
2350          * if it is NULL then something went wrong and it wasn't allocated,
2351          * report -ENOMEM in that case */
2352         if (tbc == NULL)
2353                 RETURN(-ENOMEM);
2354
2355         body = tsi->tsi_ost_body;
2356         LASSERT(body != NULL);
2357
2358         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2359             body->oa.o_flags & OBD_FL_NORPC)
2360                 RETURN(0);
2361
2362         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2363         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2364
2365         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2366         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2367
2368         local_nb = tbc->local;
2369
2370         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2371                           &lockh, LCK_PR);
2372         if (rc != 0)
2373                 RETURN(rc);
2374
2375         /*
2376          * If getting the lock took more time than
2377          * client was willing to wait, drop it. b=11330
2378          */
2379         if (ktime_get_real_seconds() > req->rq_deadline ||
2380             CFS_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2381                 no_reply = 1;
2382                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2383                        libcfs_idstr(&req->rq_peer), POSTID(&ioo->ioo_oid),
2384                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2385                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2386                 GOTO(out_lock, rc = -ETIMEDOUT);
2387         }
2388
2389         /*
2390          * Because we already sync grant info with client when
2391          * reconnect, grant info will be cleared for resent req,
2392          * otherwise, outdated grant count in the rpc would de-sync
2393          * grant counters in case of shrink
2394          */
2395         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2396                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2397                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2398         }
2399
2400         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2401         repbody->oa = body->oa;
2402
2403         npages = PTLRPC_MAX_BRW_PAGES;
2404         kstart = ktime_get();
2405         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2406                         ioo, remote_nb, &npages, local_nb);
2407         if (rc != 0)
2408                 GOTO(out_lock, rc);
2409
2410         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2411             body->oa.o_flags & OBD_FL_SHORT_IO) {
2412                 desc = NULL;
2413         } else {
2414                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2415                                             PTLRPC_BULK_PUT_SOURCE,
2416                                             OST_BULK_PORTAL,
2417                                             &ptlrpc_bulk_kiov_nopin_ops);
2418                 if (desc == NULL)
2419                         GOTO(out_commitrw, rc = -ENOMEM);
2420         }
2421
2422         npages_read = npages;
2423         for (i = 0; i < npages; i++) {
2424                 int page_rc = local_nb[i].lnb_rc;
2425
2426                 if (page_rc < 0) {
2427                         rc = page_rc;
2428                         npages_read = i;
2429                         break;
2430                 }
2431
2432                 nob += page_rc;
2433                 if (page_rc != 0 && desc != NULL) { /* some data! */
2434                         LASSERT(local_nb[i].lnb_page != NULL);
2435                         desc->bd_frag_ops->add_kiov_frag
2436                           (desc, local_nb[i].lnb_page,
2437                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2438                            page_rc);
2439                 }
2440
2441                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2442                         local_nb[i].lnb_len = page_rc;
2443                         npages_read = i + (page_rc != 0 ? 1 : 0);
2444                         /* All subsequent pages should be 0 */
2445                         while (++i < npages)
2446                                 LASSERT(local_nb[i].lnb_rc == 0);
2447                         break;
2448                 }
2449         }
2450
2451         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2452                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2453                            body->oa.o_flags : 0;
2454                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2455                 bool resend = (body->oa.o_valid & OBD_MD_FLFLAGS) &&
2456                         (body->oa.o_flags & OBD_FL_RECOV_RESEND);
2457
2458                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2459                                                           cksum_type);
2460                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2461
2462                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2463                                             local_nb, npages_read, OST_READ,
2464                                             &repbody->oa.o_cksum, resend);
2465                 if (rc < 0)
2466                         GOTO(out_commitrw, rc);
2467                 CDEBUG(D_PAGE | (resend ? D_HA : 0),
2468                        "checksum at read origin: %x (%x)\n",
2469                        repbody->oa.o_cksum, cksum_type);
2470
2471                 /* if a resend it could be for a cksum error, so check Server
2472                  * cksum with returned Client cksum (this should even cover
2473                  * zero-cksum case) */
2474                 if (resend)
2475                         check_read_checksum(local_nb, npages_read, exp,
2476                                             &body->oa, &req->rq_peer,
2477                                             body->oa.o_cksum,
2478                                             repbody->oa.o_cksum, cksum_type);
2479         } else {
2480                 repbody->oa.o_valid = 0;
2481         }
2482         if (body->oa.o_valid & OBD_MD_FLGRANT)
2483                 repbody->oa.o_valid |= OBD_MD_FLGRANT;
2484         /* We're finishing using body->oa as an input variable */
2485
2486         /* Check if client was evicted while we were doing i/o before touching
2487          * network */
2488         if (rc == 0) {
2489                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2490                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2491                         unsigned char *short_io_buf;
2492                         int short_io_size;
2493
2494                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2495                                                               &RMF_SHORT_IO);
2496                         short_io_size = req_capsule_get_size(&req->rq_pill,
2497                                                              &RMF_SHORT_IO,
2498                                                              RCL_SERVER);
2499                         rc = tgt_pages2shortio(local_nb, npages_read,
2500                                                short_io_buf, short_io_size);
2501                         if (rc >= 0)
2502                                 req_capsule_shrink(&req->rq_pill,
2503                                                    &RMF_SHORT_IO, rc,
2504                                                    RCL_SERVER);
2505                         rc = rc > 0 ? 0 : rc;
2506                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2507                         rc = target_bulk_io(exp, desc);
2508                 }
2509                 no_reply = rc != 0;
2510         } else {
2511                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2512                     body->oa.o_flags & OBD_FL_SHORT_IO)
2513                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2514                                            RCL_SERVER);
2515         }
2516
2517 out_commitrw:
2518         /* Must commit after prep above in all cases */
2519         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2520                           remote_nb, npages, local_nb, rc, nob, kstart);
2521 out_lock:
2522         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2523
2524         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2525                 ptlrpc_free_bulk(desc);
2526
2527         LASSERT(rc <= 0);
2528         if (rc == 0) {
2529                 rc = nob;
2530                 ptlrpc_lprocfs_brw(req, nob);
2531         } else if (no_reply) {
2532                 req->rq_no_reply = 1;
2533                 /* reply out callback would free */
2534                 ptlrpc_req_drop_rs(req);
2535                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2536                               "client will retry: rc %d\n",
2537                               obd_name,
2538                               obd_uuid2str(&exp->exp_client_uuid),
2539                               obd_export_nid2str(exp), rc);
2540         }
2541         /* send a bulk after reply to simulate a network delay or reordering
2542          * by a router - Note that !desc implies short io, so there is no bulk
2543          * to reorder. */
2544         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2545             desc) {
2546                 /* Calculate checksum before request transfer, original
2547                  * it is done by target_bulk_io() */
2548                 rc = sptlrpc_svc_wrap_bulk(req, desc);
2549                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
2550                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2551                 else /* old version, bulk matchbits is rq_xid */
2552                         req->rq_mbits = req->rq_xid;
2553
2554                 req->rq_status = rc;
2555                 target_committed_to_req(req);
2556                 target_send_reply(req, 0, 0);
2557
2558                 CDEBUG(D_INFO, "reorder BULK\n");
2559                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
2560                                  cfs_fail_val ? : 3);
2561
2562                 target_bulk_io(exp, desc);
2563                 ptlrpc_free_bulk(desc);
2564         }
2565
2566         RETURN(rc);
2567 }
2568 EXPORT_SYMBOL(tgt_brw_read);
2569
2570 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2571                              unsigned char *buf, unsigned int size)
2572 {
2573         int     i, off, len;
2574         char    *ptr;
2575
2576         for (i = 0; i < npages; i++) {
2577                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2578                 len = local[i].lnb_len;
2579
2580                 if (len == 0)
2581                         continue;
2582
2583                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2584                        i, off, len, size);
2585                 ptr = kmap_atomic(local[i].lnb_page);
2586                 if (ptr == NULL)
2587                         return -EINVAL;
2588                 memcpy(ptr + off, buf, len < size ? len : size);
2589                 kunmap_atomic(ptr);
2590                 buf += len;
2591                 size -= len;
2592         }
2593         return 0;
2594 }
2595
2596 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2597                               struct ptlrpc_bulk_desc *desc,
2598                               struct niobuf_local *local_nb, int npages,
2599                               u32 client_cksum, u32 server_cksum,
2600                               bool mmap)
2601 {
2602         struct obd_export *exp = req->rq_export;
2603         struct ost_body *body;
2604         char *router = "";
2605         char *via = "";
2606
2607         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2608         LASSERT(body != NULL);
2609
2610         if (desc && !nid_same(&req->rq_peer.nid, &desc->bd_sender)) {
2611                 via = " via ";
2612                 router = libcfs_nidstr(&desc->bd_sender);
2613         }
2614
2615         if (exp->exp_obd->obd_checksum_dump)
2616                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2617                                     client_cksum);
2618
2619         if (mmap) {
2620                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2621                              client_cksum, server_cksum);
2622                 return;
2623         }
2624
2625         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2626                            DFID" object "DOSTID" extent [%llu-%llu"
2627                            "]: client csum %x, server csum %x\n",
2628                            exp->exp_obd->obd_name, libcfs_idstr(&req->rq_peer),
2629                            via, router,
2630                            body->oa.o_valid & OBD_MD_FLFID ?
2631                            body->oa.o_parent_seq : (__u64)0,
2632                            body->oa.o_valid & OBD_MD_FLFID ?
2633                            body->oa.o_parent_oid : 0,
2634                            body->oa.o_valid & OBD_MD_FLFID ?
2635                            body->oa.o_parent_ver : 0,
2636                            POSTID(&body->oa.o_oi),
2637                            local_nb[0].lnb_file_offset,
2638                            local_nb[npages-1].lnb_file_offset +
2639                            local_nb[npages - 1].lnb_len - 1,
2640                            client_cksum, server_cksum);
2641 }
2642
2643 int tgt_brw_write(struct tgt_session_info *tsi)
2644 {
2645         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2646         struct ptlrpc_bulk_desc *desc = NULL;
2647         struct obd_export       *exp = req->rq_export;
2648         struct niobuf_remote    *remote_nb;
2649         struct niobuf_local     *local_nb;
2650         struct obd_ioobj        *ioo;
2651         struct ost_body         *body, *repbody;
2652         struct lustre_handle     lockh = {0};
2653         __u32                   *rcs;
2654         int                      objcount, niocount, npages;
2655         int                      rc = 0;
2656         int                      i, j;
2657         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2658         bool                     no_reply = false, mmap;
2659         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2660         bool wait_sync = false;
2661         const char *obd_name = exp->exp_obd->obd_name;
2662         /* '1' for consistency with code that checks !mpflag to restore */
2663         unsigned int mpflags = 1;
2664         ktime_t kstart;
2665         int nob = 0;
2666
2667         ENTRY;
2668
2669         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2670             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2671                 CERROR("%s: deny write request from %s to portal %u\n",
2672                        tgt_name(tsi->tsi_tgt),
2673                        obd_export_nid2str(req->rq_export),
2674                        ptlrpc_req2svc(req)->srv_req_portal);
2675                 RETURN(err_serious(-EPROTO));
2676         }
2677
2678         if (CFS_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2679                 RETURN(err_serious(-ENOSPC));
2680         if (CFS_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2681                 RETURN(err_serious(cfs_fail_val ? -cfs_fail_val : -EROFS));
2682
2683         req->rq_bulk_write = 1;
2684
2685         if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2686                 rc = -EIO;
2687         if (CFS_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2688                 rc = -EFAULT;
2689         if (rc < 0) {
2690                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2691                  * this server to fail requests.
2692                  */
2693                 char fail_ost_name[MAX_OBD_NAME];
2694
2695                 if (cfs_fail_val > 0) {
2696                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2697                                  cfs_fail_val - 1);
2698
2699                         if (strstr(obd_name, fail_ost_name))
2700                                 RETURN(err_serious(rc));
2701                 } else {
2702                         RETURN(err_serious(rc));
2703                 }
2704         }
2705
2706         /* pause before transaction has been started */
2707         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2708                          cfs_fail_val : (obd_timeout + 1) / 4);
2709
2710         /* Delay write commit to show stale size information */
2711         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2712
2713         /* There must be big cache in current thread to process this request
2714          * if it is NULL then something went wrong and it wasn't allocated,
2715          * report -ENOMEM in that case */
2716         if (tbc == NULL)
2717                 RETURN(-ENOMEM);
2718
2719         body = tsi->tsi_ost_body;
2720         LASSERT(body != NULL);
2721
2722         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2723             body->oa.o_flags & OBD_FL_NORPC)
2724                 RETURN(0);
2725
2726
2727         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2728         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2729
2730         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2731                                         RCL_CLIENT) / sizeof(*ioo);
2732
2733         for (niocount = i = 0; i < objcount; i++)
2734                 niocount += ioo[i].ioo_bufcnt;
2735
2736         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2737         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2738         if (niocount != req_capsule_get_size(&req->rq_pill,
2739                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2740                         sizeof(*remote_nb))
2741                 RETURN(err_serious(-EPROTO));
2742
2743         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2744             ptlrpc_connection_is_local(exp->exp_connection))
2745                 mpflags = memalloc_noreclaim_save();
2746
2747         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2748                              niocount * sizeof(*rcs));
2749         rc = req_capsule_server_pack(&req->rq_pill);
2750         if (rc != 0)
2751                 GOTO(out, rc = err_serious(rc));
2752
2753         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2754         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2755
2756         local_nb = tbc->local;
2757
2758         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2759                           &lockh, LCK_PW);
2760         if (rc != 0)
2761                 GOTO(out, rc);
2762
2763         /*
2764          * If getting the lock took more time than
2765          * client was willing to wait, drop it. b=11330
2766          */
2767         if (ktime_get_real_seconds() > req->rq_deadline ||
2768             CFS_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2769                 no_reply = true;
2770                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2771                        tgt_name(tsi->tsi_tgt), libcfs_idstr(&req->rq_peer),
2772                        POSTID(&ioo->ioo_oid),
2773                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2774                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2775                 GOTO(out_lock, rc = -ETIMEDOUT);
2776         }
2777
2778         /* Because we already sync grant info with client when reconnect,
2779          * grant info will be cleared for resent req, then fed_grant and
2780          * total_grant will not be modified in following preprw_write */
2781         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2782                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2783                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2784         }
2785
2786         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2787         if (repbody == NULL)
2788                 GOTO(out_lock, rc = -ENOMEM);
2789         repbody->oa = body->oa;
2790
2791         npages = PTLRPC_MAX_BRW_PAGES;
2792         kstart = ktime_get();
2793         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2794                         objcount, ioo, remote_nb, &npages, local_nb);
2795         if (rc < 0)
2796                 GOTO(out_lock, rc);
2797         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2798             body->oa.o_flags & OBD_FL_SHORT_IO) {
2799                 unsigned int short_io_size;
2800                 unsigned char *short_io_buf;
2801
2802                 short_io_size = req_capsule_get_size(&req->rq_pill,
2803                                                      &RMF_SHORT_IO,
2804                                                      RCL_CLIENT);
2805                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2806                                                       &RMF_SHORT_IO);
2807                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2808                                " size = %d\n", short_io_size);
2809
2810                 /* Copy short io buf to pages */
2811                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2812                                        short_io_size);
2813                 desc = NULL;
2814         } else {
2815                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2816                                             PTLRPC_BULK_GET_SINK,
2817                                             OST_BULK_PORTAL,
2818                                             &ptlrpc_bulk_kiov_nopin_ops);
2819                 if (desc == NULL)
2820                         GOTO(skip_transfer, rc = -ENOMEM);
2821
2822                 /* NB Having prepped, we must commit... */
2823                 for (i = 0; i < npages; i++)
2824                         desc->bd_frag_ops->add_kiov_frag(desc,
2825                                         local_nb[i].lnb_page,
2826                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2827                                         local_nb[i].lnb_len);
2828
2829                 rc = sptlrpc_svc_prep_bulk(req, desc);
2830                 if (rc != 0)
2831                         GOTO(skip_transfer, rc);
2832
2833                 rc = target_bulk_io(exp, desc);
2834         }
2835
2836         no_reply = rc != 0;
2837
2838 skip_transfer:
2839         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2840                 static int cksum_counter;
2841
2842                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2843                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2844
2845                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2846                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2847                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2848                                                            cksum_type);
2849
2850                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2851                                             local_nb, npages, OST_WRITE,
2852                                             &repbody->oa.o_cksum, false);
2853                 if (rc < 0)
2854                         GOTO(out_commitrw, rc);
2855
2856                 cksum_counter++;
2857
2858                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2859                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2860                                 body->oa.o_flags & OBD_FL_MMAP);
2861
2862                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2863                                           body->oa.o_cksum,
2864                                           repbody->oa.o_cksum, mmap);
2865                         cksum_counter = 0;
2866                 } else if ((cksum_counter & (-cksum_counter)) ==
2867                            cksum_counter) {
2868                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2869                                cksum_counter, libcfs_idstr(&req->rq_peer),
2870                                repbody->oa.o_cksum);
2871                 }
2872         }
2873
2874         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2875
2876 out_commitrw:
2877         /* calculate the expected actual write bytes (nob) for OFD stats.
2878          * Technically, if commit fails this would be wrong, but that should be
2879          * very rare
2880          */
2881         for (i = 0; i < niocount; i++) {
2882                 int len = remote_nb[i].rnb_len;
2883
2884                 nob += len;
2885         }
2886
2887         /* multiple transactions can be assigned during write commit */
2888         tsi->tsi_mult_trans = 1;
2889
2890         /* Must commit after prep above in all cases */
2891         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2892                           objcount, ioo, remote_nb, npages, local_nb, rc, nob,
2893                           kstart);
2894         if (rc == -ENOTCONN)
2895                 /* quota acquire process has been given up because
2896                  * either the client has been evicted or the client
2897                  * has timed out the request already
2898                  */
2899                 no_reply = true;
2900
2901         for (i = 0; i < niocount; i++) {
2902                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2903                         wait_sync = true;
2904                         break;
2905                 }
2906         }
2907         /*
2908          * Disable sending mtime back to the client. If the client locked the
2909          * whole object, then it has already updated the mtime on its side,
2910          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2911          */
2912         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2913
2914         if (rc == 0) {
2915                 /* set per-requested niobuf return codes */
2916                 for (i = j = 0; i < niocount; i++) {
2917                         int len = remote_nb[i].rnb_len;
2918
2919                         rcs[i] = 0;
2920                         do {
2921                                 LASSERT(j < npages);
2922                                 if (local_nb[j].lnb_rc < 0)
2923                                         rcs[i] = local_nb[j].lnb_rc;
2924                                 len -= local_nb[j].lnb_len;
2925                                 j++;
2926                         } while (len > 0);
2927                         LASSERT(len == 0);
2928                 }
2929                 LASSERT(j == npages);
2930                 ptlrpc_lprocfs_brw(req, nob);
2931         }
2932 out_lock:
2933         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2934         if (desc)
2935                 ptlrpc_free_bulk(desc);
2936 out:
2937         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2938                 req->rq_no_reply = 1;
2939                 /* reply out callback would free */
2940                 ptlrpc_req_drop_rs(req);
2941                 if (!exp->exp_obd->obd_no_transno)
2942                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2943                                       " client will retry: rc = %d\n",
2944                                       obd_name,
2945                                       obd_uuid2str(&exp->exp_client_uuid),
2946                                       obd_export_nid2str(exp), rc);
2947         }
2948
2949         if (mpflags)
2950                 memalloc_noreclaim_restore(mpflags);
2951
2952         RETURN(rc);
2953 }
2954 EXPORT_SYMBOL(tgt_brw_write);
2955
2956 /**
2957  * Common request handler for OST_SEEK RPC.
2958  *
2959  * Unified request handling for OST_SEEK RPC.
2960  * It takes object by its FID, does needed lseek and packs result
2961  * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2962  *
2963  * \param[in] tsi       target session environment for this request
2964  *
2965  * \retval              0 if successful
2966  * \retval              negative value on error
2967  */
2968 int tgt_lseek(struct tgt_session_info *tsi)
2969 {
2970         struct lustre_handle lh = { 0 };
2971         struct dt_object *dob;
2972         struct ost_body *repbody;
2973         loff_t offset = tsi->tsi_ost_body->oa.o_size;
2974         int whence = tsi->tsi_ost_body->oa.o_mode;
2975         bool srvlock;
2976         int rc = 0;
2977
2978         ENTRY;
2979
2980         if (whence != SEEK_HOLE && whence != SEEK_DATA)
2981                 RETURN(-EPROTO);
2982
2983         /* Negative offset is prohibited on wire and must be handled on client
2984          * prior sending RPC.
2985          */
2986         if (offset < 0)
2987                 RETURN(-EPROTO);
2988
2989         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2990         if (repbody == NULL)
2991                 RETURN(-ENOMEM);
2992         repbody->oa = tsi->tsi_ost_body->oa;
2993
2994         srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
2995                   tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
2996         if (srvlock) {
2997                 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
2998                                    offset, OBD_OBJECT_EOF, &lh, LCK_PR);
2999                 if (rc)
3000                         RETURN(rc);
3001         }
3002
3003         dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
3004         if (IS_ERR(dob))
3005                 GOTO(out, rc = PTR_ERR(dob));
3006
3007         if (!dt_object_exists(dob))
3008                 GOTO(obj_put, rc = -ENOENT);
3009
3010         repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
3011         rc = 0;
3012 obj_put:
3013         dt_object_put(tsi->tsi_env, dob);
3014 out:
3015         if (srvlock)
3016                 tgt_data_unlock(&lh, LCK_PR);
3017
3018         RETURN(rc);
3019 }
3020 EXPORT_SYMBOL(tgt_lseek);
3021
3022 /* Check if request can be reconstructed from saved reply data
3023  * A copy of the reply data is returned in @trd if the pointer is not NULL
3024  */
3025 int req_can_reconstruct(struct ptlrpc_request *req,
3026                          struct tg_reply_data *trd)
3027 {
3028         struct tg_export_data *ted = &req->rq_export->exp_target_data;
3029         struct lsd_client_data *lcd = ted->ted_lcd;
3030         int found;
3031
3032         if (tgt_is_multimodrpcs_client(req->rq_export))
3033                 return tgt_lookup_reply(req, trd);
3034
3035         mutex_lock(&ted->ted_lcd_lock);
3036         found = req->rq_xid == lcd->lcd_last_xid ||
3037                 req->rq_xid == lcd->lcd_last_close_xid;
3038
3039         if (found && trd != NULL) {
3040                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
3041                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
3042                         trd->trd_reply.lrd_transno =
3043                                                 lcd->lcd_last_close_transno;
3044                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
3045                 } else {
3046                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
3047                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
3048                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
3049                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
3050                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
3051                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
3052                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
3053                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
3054                 }
3055         }
3056         mutex_unlock(&ted->ted_lcd_lock);
3057
3058         return found;
3059 }
3060 EXPORT_SYMBOL(req_can_reconstruct);
3061
3062 bool tgt_check_resent(struct ptlrpc_request *req, struct tg_reply_data *trd)
3063 {
3064         struct lsd_reply_data *lrd;
3065         bool need_reconstruct = false;
3066
3067         ENTRY;
3068
3069         if (likely(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
3070                 return false;
3071
3072         if (req_can_reconstruct(req, trd)) {
3073                 lrd = &trd->trd_reply;
3074                 req->rq_transno = lrd->lrd_transno;
3075                 req->rq_status = lrd->lrd_result;
3076                 if (req->rq_status != 0)
3077                         req->rq_transno = 0;
3078                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
3079                 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
3080
3081                 DEBUG_REQ(D_HA, req, "reconstruct resent RPC");
3082                 need_reconstruct = true;
3083         } else {
3084                 DEBUG_REQ(D_HA, req, "no reply for RESENT req");
3085         }
3086
3087         return need_reconstruct;
3088 }
3089 EXPORT_SYMBOL(tgt_check_resent);