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 (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
 417                 RETURN(0);
 418         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
 419                      OBD_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                      OBD_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 (OBD_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(OBD_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                              OBD_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                                              lnet_nid_to_nid4(&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         OBD_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_obd_log_cancel(struct tgt_session_info *tsi)
1153 {
1154         return err_serious(-EOPNOTSUPP);
1155 }
1156
1157 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1158 {
1159         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1160         struct obd_export       *exp = req->rq_export;
1161         struct ptlrpc_bulk_desc *desc;
1162         int                      i;
1163         int                      rc;
1164         int                      pages = 0;
1165
1166         ENTRY;
1167
1168         for (i = 0; i < rdbuf->rb_nbufs; i++) {
1169                 unsigned int offset;
1170
1171                 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1172                 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1173                                       PAGE_SIZE);
1174         }
1175
1176         desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1177                                   PTLRPC_BULK_PUT_SOURCE,
1178                                     MDS_BULK_PORTAL,
1179                                     &ptlrpc_bulk_kiov_nopin_ops);
1180         if (desc == NULL)
1181                 RETURN(-ENOMEM);
1182
1183         for (i = 0; i < rdbuf->rb_nbufs; i++)
1184                 desc->bd_frag_ops->add_iov_frag(desc,
1185                                         rdbuf->rb_bufs[i].lb_buf,
1186                                         rdbuf->rb_bufs[i].lb_len);
1187
1188         rc = target_bulk_io(exp, desc);
1189         ptlrpc_free_bulk(desc);
1190         RETURN(rc);
1191 }
1192 EXPORT_SYMBOL(tgt_send_buffer);
1193
1194 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1195 {
1196         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1197         struct obd_export       *exp = req->rq_export;
1198         struct ptlrpc_bulk_desc *desc;
1199         int                      tmpcount;
1200         int                      tmpsize;
1201         int                      i;
1202         int                      rc;
1203
1204         ENTRY;
1205
1206         desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1207                                     PTLRPC_BULK_PUT_SOURCE,
1208                                     MDS_BULK_PORTAL,
1209                                     &ptlrpc_bulk_kiov_pin_ops);
1210         if (desc == NULL)
1211                 RETURN(-ENOMEM);
1212
1213         if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1214                 /* old client requires reply size in it's PAGE_SIZE,
1215                  * which is rdpg->rp_count */
1216                 nob = rdpg->rp_count;
1217
1218         for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1219              i++, tmpcount -= tmpsize) {
1220                 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1221                 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1222                                                  tmpsize);
1223         }
1224
1225         LASSERT(desc->bd_nob == nob);
1226         rc = target_bulk_io(exp, desc);
1227         ptlrpc_free_bulk(desc);
1228         RETURN(rc);
1229 }
1230 EXPORT_SYMBOL(tgt_sendpage);
1231
1232 /*
1233  * OBD_IDX_READ handler
1234  */
1235 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1236 {
1237         struct tgt_thread_info  *tti = tgt_th_info(tsi->tsi_env);
1238         struct lu_rdpg          *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1239         struct idx_info         *req_ii, *rep_ii;
1240         int                      rc, i;
1241
1242         ENTRY;
1243
1244         memset(rdpg, 0, sizeof(*rdpg));
1245         req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1246
1247         /* extract idx_info buffer from request & reply */
1248         req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1249         if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1250                 RETURN(err_serious(-EPROTO));
1251
1252         rc = req_capsule_server_pack(tsi->tsi_pill);
1253         if (rc)
1254                 RETURN(err_serious(rc));
1255
1256         rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1257         if (rep_ii == NULL)
1258                 RETURN(err_serious(-EFAULT));
1259         rep_ii->ii_magic = IDX_INFO_MAGIC;
1260
1261         /* extract hash to start with */
1262         rdpg->rp_hash = req_ii->ii_hash_start;
1263
1264         /* extract requested attributes */
1265         rdpg->rp_attrs = req_ii->ii_attrs;
1266
1267         /* check that fid packed in request is valid and supported */
1268         if (!fid_is_sane(&req_ii->ii_fid))
1269                 RETURN(-EINVAL);
1270         rep_ii->ii_fid = req_ii->ii_fid;
1271
1272         /* copy flags */
1273         rep_ii->ii_flags = req_ii->ii_flags;
1274
1275         /* compute number of pages to allocate, ii_count is the number of 4KB
1276          * containers */
1277         if (req_ii->ii_count <= 0)
1278                 GOTO(out, rc = -EFAULT);
1279         rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1280                                exp_max_brw_size(tsi->tsi_exp));
1281         rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1282
1283         /* allocate pages to store the containers */
1284         OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1285         if (rdpg->rp_pages == NULL)
1286                 GOTO(out, rc = -ENOMEM);
1287         for (i = 0; i < rdpg->rp_npages; i++) {
1288                 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1289                 if (rdpg->rp_pages[i] == NULL)
1290                         GOTO(out, rc = -ENOMEM);
1291         }
1292
1293         /* populate pages with key/record pairs */
1294         rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1295         if (rc < 0)
1296                 GOTO(out, rc);
1297
1298         LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1299                  "asked %d > %d\n", rc, rdpg->rp_count);
1300
1301         /* send pages to client */
1302         rc = tgt_sendpage(tsi, rdpg, rc);
1303         if (rc)
1304                 GOTO(out, rc);
1305         EXIT;
1306 out:
1307         if (rdpg->rp_pages) {
1308                 for (i = 0; i < rdpg->rp_npages; i++)
1309                         if (rdpg->rp_pages[i])
1310                                 __free_page(rdpg->rp_pages[i]);
1311                 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1312         }
1313         return rc;
1314 }
1315
1316 struct tgt_handler tgt_obd_handlers[] = {
1317 TGT_OBD_HDL    (0,      OBD_PING,               tgt_obd_ping),
1318 TGT_OBD_HDL    (0,      OBD_IDX_READ,           tgt_obd_idx_read)
1319 };
1320 EXPORT_SYMBOL(tgt_obd_handlers);
1321
1322 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1323              struct dt_object *obj, __u64 start, __u64 end)
1324 {
1325         int rc = 0;
1326
1327         ENTRY;
1328
1329         /* if no objid is specified, it means "sync whole filesystem" */
1330         if (obj == NULL) {
1331                 rc = dt_sync(env, tgt->lut_bottom);
1332         } else if (dt_version_get(env, obj) >
1333                    tgt->lut_obd->obd_last_committed) {
1334                 rc = dt_object_sync(env, obj, start, end);
1335         }
1336         atomic_inc(&tgt->lut_sync_count);
1337
1338         RETURN(rc);
1339 }
1340 EXPORT_SYMBOL(tgt_sync);
1341 /*
1342  * Unified target DLM handlers.
1343  */
1344
1345 /**
1346  * Unified target BAST
1347  *
1348  * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1349  * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1350  * but if it's cross-MDT lock, because remote object version is not set, a
1351  * filesystem sync is needed.
1352  *
1353  * \param lock server side lock
1354  * \param desc lock desc
1355  * \param data ldlm_cb_set_arg
1356  * \param flag  indicates whether this cancelling or blocking callback
1357  * \retval      0 on success
1358  * \retval      negative number on error
1359  */
1360 int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1361                      void *data, int flag)
1362 {
1363         struct lu_env            env;
1364         struct lu_target        *tgt;
1365         struct dt_object        *obj = NULL;
1366         struct lu_fid            fid;
1367         int                      rc = 0;
1368
1369         ENTRY;
1370
1371         tgt = class_exp2tgt(lock->l_export);
1372
1373         if (unlikely(tgt == NULL)) {
1374                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1375                        class_exp2obd(lock->l_export)->obd_name);
1376                 RETURN(-EINVAL);
1377         }
1378
1379         if (flag == LDLM_CB_CANCELING &&
1380             (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1381             (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1382              (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1383               ldlm_is_cbpending(lock))) &&
1384             ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1385              lock->l_resource->lr_type == LDLM_EXTENT)) {
1386                 __u64 start = 0;
1387                 __u64 end = OBD_OBJECT_EOF;
1388
1389                 rc = lu_env_init(&env, LCT_DT_THREAD);
1390                 if (unlikely(rc != 0))
1391                         GOTO(err, rc);
1392
1393                 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1394                                    tgt->lut_lsd.lsd_osd_index);
1395
1396                 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1397                         obj = dt_locate(&env, tgt->lut_bottom, &fid);
1398                         if (IS_ERR(obj))
1399                                 GOTO(err_env, rc = PTR_ERR(obj));
1400
1401                         if (!dt_object_exists(obj))
1402                                 GOTO(err_put, rc = -ENOENT);
1403
1404                         start = lock->l_policy_data.l_extent.start;
1405                         end = lock->l_policy_data.l_extent.end;
1406                 }
1407
1408                 rc = tgt_sync(&env, tgt, obj, start, end);
1409                 if (rc < 0) {
1410                         CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1411                                "cancel: rc = %d\n",
1412                                tgt_name(tgt), PFID(&fid),
1413                                lock->l_policy_data.l_extent.start,
1414                                lock->l_policy_data.l_extent.end, rc);
1415                 }
1416 err_put:
1417                 if (obj != NULL)
1418                         dt_object_put(&env, obj);
1419 err_env:
1420                 lu_env_fini(&env);
1421         }
1422 err:
1423         rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1424         RETURN(rc);
1425 }
1426 EXPORT_SYMBOL(tgt_blocking_ast);
1427
1428 static struct ldlm_callback_suite tgt_dlm_cbs = {
1429         .lcs_completion = ldlm_server_completion_ast,
1430         .lcs_blocking   = tgt_blocking_ast,
1431         .lcs_glimpse    = ldlm_server_glimpse_ast
1432 };
1433
1434 int tgt_enqueue(struct tgt_session_info *tsi)
1435 {
1436         struct ptlrpc_request *req = tgt_ses_req(tsi);
1437         int rc;
1438
1439         ENTRY;
1440         /*
1441          * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1442          * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1443          */
1444         LASSERT(tsi->tsi_dlm_req != NULL);
1445         rc = ldlm_handle_enqueue(tsi->tsi_exp->exp_obd->obd_namespace,
1446                                  &req->rq_pill, tsi->tsi_dlm_req, &tgt_dlm_cbs);
1447         if (rc)
1448                 RETURN(err_serious(rc));
1449
1450         switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1451         case LUT_FAIL_MDT:
1452                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1453                 break;
1454         case LUT_FAIL_OST:
1455                 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1456                 break;
1457         case LUT_FAIL_MGT:
1458                 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1459                 break;
1460         default:
1461                 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1462                 break;
1463         }
1464         RETURN(req->rq_status);
1465 }
1466 EXPORT_SYMBOL(tgt_enqueue);
1467
1468 int tgt_convert(struct tgt_session_info *tsi)
1469 {
1470         struct ptlrpc_request *req = tgt_ses_req(tsi);
1471         int rc;
1472
1473         ENTRY;
1474         LASSERT(tsi->tsi_dlm_req);
1475         rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1476         if (rc)
1477                 RETURN(err_serious(rc));
1478
1479         RETURN(req->rq_status);
1480 }
1481
1482 int tgt_bl_callback(struct tgt_session_info *tsi)
1483 {
1484         return err_serious(-EOPNOTSUPP);
1485 }
1486
1487 int tgt_cp_callback(struct tgt_session_info *tsi)
1488 {
1489         return err_serious(-EOPNOTSUPP);
1490 }
1491
1492 /* generic LDLM target handler */
1493 struct tgt_handler tgt_dlm_handlers[] = {
1494 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1495 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1496 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1497 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1498 };
1499 EXPORT_SYMBOL(tgt_dlm_handlers);
1500
1501 /*
1502  * Unified target LLOG handlers.
1503  */
1504 int tgt_llog_open(struct tgt_session_info *tsi)
1505 {
1506         int rc;
1507
1508         ENTRY;
1509
1510         rc = llog_origin_handle_open(tgt_ses_req(tsi));
1511
1512         RETURN(rc);
1513 }
1514 EXPORT_SYMBOL(tgt_llog_open);
1515
1516 int tgt_llog_read_header(struct tgt_session_info *tsi)
1517 {
1518         int rc;
1519
1520         ENTRY;
1521
1522         rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1523
1524         RETURN(rc);
1525 }
1526 EXPORT_SYMBOL(tgt_llog_read_header);
1527
1528 int tgt_llog_next_block(struct tgt_session_info *tsi)
1529 {
1530         int rc;
1531
1532         ENTRY;
1533
1534         rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1535
1536         RETURN(rc);
1537 }
1538 EXPORT_SYMBOL(tgt_llog_next_block);
1539
1540 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1541 {
1542         int rc;
1543
1544         ENTRY;
1545
1546         rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1547
1548         RETURN(rc);
1549 }
1550 EXPORT_SYMBOL(tgt_llog_prev_block);
1551
1552 /* generic llog target handler */
1553 struct tgt_handler tgt_llog_handlers[] = {
1554 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_CREATE,      tgt_llog_open),
1555 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_NEXT_BLOCK,  tgt_llog_next_block),
1556 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1557 TGT_LLOG_HDL    (0,     LLOG_ORIGIN_HANDLE_PREV_BLOCK,  tgt_llog_prev_block),
1558 };
1559 EXPORT_SYMBOL(tgt_llog_handlers);
1560
1561 /*
1562  * sec context handlers
1563  */
1564 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1565 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1566 {
1567         return 0;
1568 }
1569
1570 struct tgt_handler tgt_sec_ctx_handlers[] = {
1571 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT,           tgt_sec_ctx_handle),
1572 TGT_SEC_HDL_VAR(0,      SEC_CTX_INIT_CONT,      tgt_sec_ctx_handle),
1573 TGT_SEC_HDL_VAR(0,      SEC_CTX_FINI,           tgt_sec_ctx_handle),
1574 };
1575 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1576
1577 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1578                                  struct dt_device *key,
1579                                  struct lfsck_req_local *lrl,
1580                                  struct thandle *th) = NULL;
1581
1582 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1583                                                       struct dt_device *,
1584                                                       struct lfsck_req_local *,
1585                                                       struct thandle *))
1586 {
1587         tgt_lfsck_in_notify_local = notify;
1588 }
1589 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1590
1591 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1592                            struct dt_device *key,
1593                            struct lfsck_request *lr) = NULL;
1594
1595 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1596                                                 struct dt_device *,
1597                                                 struct lfsck_request *))
1598 {
1599         tgt_lfsck_in_notify = notify;
1600 }
1601 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1602
1603 static int (*tgt_lfsck_query)(const struct lu_env *env,
1604                               struct dt_device *key,
1605                               struct lfsck_request *req,
1606                               struct lfsck_reply *rep,
1607                               struct lfsck_query *que) = NULL;
1608
1609 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1610                                            struct dt_device *,
1611                                            struct lfsck_request *,
1612                                            struct lfsck_reply *,
1613                                            struct lfsck_query *))
1614 {
1615         tgt_lfsck_query = query;
1616 }
1617 EXPORT_SYMBOL(tgt_register_lfsck_query);
1618
1619 /* LFSCK request handlers */
1620 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1621 {
1622         const struct lu_env     *env = tsi->tsi_env;
1623         struct dt_device        *key = tsi->tsi_tgt->lut_bottom;
1624         struct lfsck_request    *lr;
1625         int                      rc;
1626         ENTRY;
1627
1628         lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1629         if (lr == NULL)
1630                 RETURN(-EPROTO);
1631
1632         rc = tgt_lfsck_in_notify(env, key, lr);
1633
1634         RETURN(rc);
1635 }
1636
1637 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1638 {
1639         struct lfsck_request    *request;
1640         struct lfsck_reply      *reply;
1641         int                      rc;
1642         ENTRY;
1643
1644         request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1645         if (request == NULL)
1646                 RETURN(-EPROTO);
1647
1648         reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1649         if (reply == NULL)
1650                 RETURN(-ENOMEM);
1651
1652         rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1653                              request, reply, NULL);
1654
1655         RETURN(rc < 0 ? rc : 0);
1656 }
1657
1658 struct tgt_handler tgt_lfsck_handlers[] = {
1659 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_NOTIFY,   tgt_handle_lfsck_notify),
1660 TGT_LFSCK_HDL(HAS_REPLY,        LFSCK_QUERY,    tgt_handle_lfsck_query),
1661 };
1662 EXPORT_SYMBOL(tgt_lfsck_handlers);
1663
1664 /*
1665  * initialize per-thread page pool (bug 5137).
1666  */
1667 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1668 {
1669         struct tgt_thread_big_cache *tbc;
1670
1671         ENTRY;
1672
1673         LASSERT(thread != NULL);
1674         LASSERT(thread->t_data == NULL);
1675
1676         OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1677         if (tbc == NULL)
1678                 RETURN(-ENOMEM);
1679         thread->t_data = tbc;
1680         RETURN(0);
1681 }
1682 EXPORT_SYMBOL(tgt_io_thread_init);
1683
1684 /*
1685  * free per-thread pool created by tgt_thread_init().
1686  */
1687 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1688 {
1689         struct tgt_thread_big_cache *tbc;
1690
1691         ENTRY;
1692
1693         LASSERT(thread != NULL);
1694
1695         /*
1696          * be prepared to handle partially-initialized pools (because this is
1697          * called from ost_io_thread_init() for cleanup.
1698          */
1699         tbc = thread->t_data;
1700         if (tbc != NULL) {
1701                 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1702                 thread->t_data = NULL;
1703         }
1704         EXIT;
1705 }
1706 EXPORT_SYMBOL(tgt_io_thread_done);
1707
1708 /**
1709  * Helper function for getting Data-on-MDT file server DLM lock
1710  * if asked by client.
1711  */
1712 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1713                       struct lustre_handle *lh, int mode, __u64 *flags)
1714 {
1715         union ldlm_policy_data policy = {
1716                 .l_inodebits.bits = MDS_INODELOCK_DOM,
1717         };
1718         int rc;
1719
1720         ENTRY;
1721
1722         LASSERT(lh != NULL);
1723         LASSERT(ns != NULL);
1724         LASSERT(!lustre_handle_is_used(lh));
1725
1726         rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1727                                     flags, ldlm_blocking_ast,
1728                                     ldlm_completion_ast, ldlm_glimpse_ast,
1729                                     NULL, 0, LVB_T_NONE, NULL, lh);
1730
1731         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1732 }
1733 EXPORT_SYMBOL(tgt_mdt_data_lock);
1734
1735 /**
1736  * Helper function for getting server side [start, start+count] DLM lock
1737  * if asked by client.
1738  */
1739 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1740                     struct ldlm_res_id *res_id, __u64 start, __u64 end,
1741                     struct lustre_handle *lh, int mode, __u64 *flags)
1742 {
1743         union ldlm_policy_data policy;
1744         int rc;
1745
1746         ENTRY;
1747
1748         LASSERT(lh != NULL);
1749         LASSERT(ns != NULL);
1750         LASSERT(!lustre_handle_is_used(lh));
1751
1752         policy.l_extent.gid = 0;
1753         policy.l_extent.start = start & PAGE_MASK;
1754
1755         /*
1756          * If ->o_blocks is EOF it means "lock till the end of the file".
1757          * Otherwise, it's size of an extent or hole being punched (in bytes).
1758          */
1759         if (end == OBD_OBJECT_EOF || end < start)
1760                 policy.l_extent.end = OBD_OBJECT_EOF;
1761         else
1762                 policy.l_extent.end = end | ~PAGE_MASK;
1763
1764         rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1765                                     mode, flags, ldlm_blocking_ast,
1766                                     ldlm_completion_ast, ldlm_glimpse_ast,
1767                                     NULL, 0, LVB_T_NONE, NULL, lh);
1768         RETURN(rc == ELDLM_OK ? 0 : -EIO);
1769 }
1770 EXPORT_SYMBOL(tgt_extent_lock);
1771
1772 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1773                          struct ldlm_res_id *res_id, __u64 start, __u64 end,
1774                          struct lustre_handle *lh, enum ldlm_mode mode)
1775 {
1776         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1777         __u64 flags = 0;
1778
1779         /* MDT IO for data-on-mdt */
1780         if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1781                 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1782
1783         return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1784 }
1785
1786 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1787 {
1788         LASSERT(lustre_handle_is_used(lh));
1789         ldlm_lock_decref(lh, mode);
1790 }
1791 EXPORT_SYMBOL(tgt_data_unlock);
1792
1793 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1794                         struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1795                         struct niobuf_remote *nb, struct lustre_handle *lh,
1796                         enum ldlm_mode mode)
1797 {
1798         int nrbufs = obj->ioo_bufcnt;
1799         int i;
1800
1801         ENTRY;
1802
1803         LASSERT(mode == LCK_PR || mode == LCK_PW);
1804         LASSERT(!lustre_handle_is_used(lh));
1805
1806         if (exp->exp_obd->obd_recovering)
1807                 RETURN(0);
1808
1809         if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1810                 RETURN(0);
1811
1812         for (i = 1; i < nrbufs; i++)
1813                 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1814                         RETURN(-EFAULT);
1815
1816         return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1817                              nb[nrbufs - 1].rnb_offset +
1818                              nb[nrbufs - 1].rnb_len - 1, lh, mode);
1819 }
1820
1821 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1822                            struct niobuf_remote *niob,
1823                            struct lustre_handle *lh, enum ldlm_mode mode)
1824 {
1825         ENTRY;
1826
1827         LASSERT(mode == LCK_PR || mode == LCK_PW);
1828         LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1829                  niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1830                 lustre_handle_is_used(lh));
1831
1832         if (lustre_handle_is_used(lh))
1833                 tgt_data_unlock(lh, mode);
1834         EXIT;
1835 }
1836
1837 static int tgt_checksum_niobuf(struct lu_target *tgt,
1838                                  struct niobuf_local *local_nb, int npages,
1839                                  int opc, enum cksum_types cksum_type,
1840                                  __u32 *cksum)
1841 {
1842         struct ahash_request           *req;
1843         unsigned int                    bufsize;
1844         int                             i, err;
1845         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1846
1847         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1848         if (IS_ERR(req)) {
1849                 CERROR("%s: unable to initialize checksum hash %s\n",
1850                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1851                 return PTR_ERR(req);
1852         }
1853
1854         CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1855         for (i = 0; i < npages; i++) {
1856                 /* corrupt the data before we compute the checksum, to
1857                  * simulate a client->OST data error */
1858                 if (i == 0 && opc == OST_WRITE &&
1859                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1860                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1861                         int len = local_nb[i].lnb_len;
1862                         struct page *np = tgt_page_to_corrupt;
1863
1864                         if (np) {
1865                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1866                                 char *ptr2 = page_address(np);
1867
1868                                 memcpy(ptr2 + off, ptr + off, len);
1869                                 memcpy(ptr2 + off, "bad3", min(4, len));
1870                                 kunmap_atomic(ptr);
1871
1872                                 /* LU-8376 to preserve original index for
1873                                  * display in dump_all_bulk_pages() */
1874                                 np->index = i;
1875
1876                                 cfs_crypto_hash_update_page(req, np, off,
1877                                                             len);
1878                                 continue;
1879                         } else {
1880                                 CERROR("%s: can't alloc page for corruption\n",
1881                                        tgt_name(tgt));
1882                         }
1883                 }
1884                 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1885                                   local_nb[i].lnb_page_offset & ~PAGE_MASK,
1886                                   local_nb[i].lnb_len);
1887
1888                  /* corrupt the data after we compute the checksum, to
1889                  * simulate an OST->client data error */
1890                 if (i == 0 && opc == OST_READ &&
1891                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1892                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1893                         int len = local_nb[i].lnb_len;
1894                         struct page *np = tgt_page_to_corrupt;
1895
1896                         if (np) {
1897                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1898                                 char *ptr2 = page_address(np);
1899
1900                                 memcpy(ptr2 + off, ptr + off, len);
1901                                 memcpy(ptr2 + off, "bad4", min(4, len));
1902                                 kunmap_atomic(ptr);
1903
1904                                 /* LU-8376 to preserve original index for
1905                                  * display in dump_all_bulk_pages() */
1906                                 np->index = i;
1907
1908                                 cfs_crypto_hash_update_page(req, np, off,
1909                                                             len);
1910                                 continue;
1911                         } else {
1912                                 CERROR("%s: can't alloc page for corruption\n",
1913                                        tgt_name(tgt));
1914                         }
1915                 }
1916         }
1917
1918         bufsize = sizeof(*cksum);
1919         err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1920
1921         return 0;
1922 }
1923
1924 char dbgcksum_file_name[PATH_MAX];
1925
1926 static void dump_all_bulk_pages(struct obdo *oa, int count,
1927                                 struct niobuf_local *local_nb,
1928                                 __u32 server_cksum, __u32 client_cksum)
1929 {
1930         struct file *filp;
1931         int rc, i;
1932         unsigned int len;
1933         char *buf;
1934
1935         /* will only keep dump of pages on first error for the same range in
1936          * file/fid, not during the resends/retries. */
1937         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1938                  "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1939                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1940                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1941                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1942                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1943                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1944                  local_nb[0].lnb_file_offset,
1945                  local_nb[count-1].lnb_file_offset +
1946                  local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1947         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1948         filp = filp_open(dbgcksum_file_name,
1949                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1950         if (IS_ERR(filp)) {
1951                 rc = PTR_ERR(filp);
1952                 if (rc == -EEXIST)
1953                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1954                                "checksum error: rc = %d\n", dbgcksum_file_name,
1955                                rc);
1956                 else
1957                         CERROR("%s: can't open to dump pages with checksum "
1958                                "error: rc = %d\n", dbgcksum_file_name, rc);
1959                 return;
1960         }
1961
1962         for (i = 0; i < count; i++) {
1963                 len = local_nb[i].lnb_len;
1964                 buf = kmap(local_nb[i].lnb_page);
1965                 while (len != 0) {
1966                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1967                         if (rc < 0) {
1968                                 CERROR("%s: wanted to write %u but got %d "
1969                                        "error\n", dbgcksum_file_name, len, rc);
1970                                 break;
1971                         }
1972                         len -= rc;
1973                         buf += rc;
1974                 }
1975                 kunmap(local_nb[i].lnb_page);
1976         }
1977
1978         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1979         if (rc)
1980                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1981         filp_close(filp, NULL);
1982
1983         libcfs_debug_dumplog();
1984 }
1985
1986 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1987                                struct obd_export *exp, struct obdo *oa,
1988                                const struct lnet_processid *peer,
1989                                __u32 client_cksum, __u32 server_cksum,
1990                                enum cksum_types server_cksum_type)
1991 {
1992         char *msg;
1993         enum cksum_types cksum_type;
1994         loff_t start, end;
1995
1996         if (unlikely(npages <= 0))
1997                 return 0;
1998
1999         /* unlikely to happen and only if resend does not occur due to cksum
2000          * control failure on Client */
2001         if (unlikely(server_cksum == client_cksum)) {
2002                 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
2003                        client_cksum);
2004                 return 0;
2005         }
2006
2007         if (exp->exp_obd->obd_checksum_dump)
2008                 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
2009                                     client_cksum);
2010
2011         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
2012                                            oa->o_flags : 0);
2013
2014         if (cksum_type != server_cksum_type)
2015                 msg = "the server may have not used the checksum type specified"
2016                       " in the original request - likely a protocol problem";
2017         else
2018                 msg = "should have changed on the client or in transit";
2019
2020         start = local_nb[0].lnb_file_offset;
2021         end = local_nb[npages-1].lnb_file_offset +
2022                                         local_nb[npages-1].lnb_len - 1;
2023
2024         LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
2025                 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
2026                 " %x (type %x), server csum %x (type %x)\n",
2027                 exp->exp_obd->obd_name,
2028                 msg, libcfs_nidstr(&peer->nid),
2029                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
2030                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
2031                 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
2032                 POSTID(&oa->o_oi),
2033                 start, end, client_cksum, cksum_type, server_cksum,
2034                 server_cksum_type);
2035
2036         return 1;
2037 }
2038
2039 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
2040                              unsigned char *buf, int size)
2041 {
2042         int     i, off, len, copied = size;
2043         char    *ptr;
2044
2045         for (i = 0; i < npages; i++) {
2046                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2047                 len = local[i].lnb_len;
2048
2049                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2050                        i, off, len, size);
2051                 if (len > size)
2052                         return -EINVAL;
2053
2054                 ptr = kmap_atomic(local[i].lnb_page);
2055                 memcpy(buf, ptr + off, len);
2056                 kunmap_atomic(ptr);
2057                 buf += len;
2058                 size -= len;
2059         }
2060         return copied - size;
2061 }
2062
2063 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2064                                      enum cksum_types cksum_type,
2065                                      struct niobuf_local *local_nb, int npages,
2066                                      int opc, obd_dif_csum_fn *fn,
2067                                      int sector_size, u32 *check_sum,
2068                                      bool resend)
2069 {
2070         enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2071         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2072         const char *obd_name = tgt->lut_obd->obd_name;
2073         struct ahash_request *req;
2074         unsigned int bufsize;
2075         unsigned char *buffer;
2076         struct page *__page;
2077         __be16 *guard_start;
2078         int guard_number;
2079         int used_number = 0;
2080         __u32 cksum;
2081         int rc = 0;
2082         int used;
2083         int i;
2084
2085         __page = alloc_page(GFP_KERNEL);
2086         if (__page == NULL)
2087                 return -ENOMEM;
2088
2089         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2090         if (IS_ERR(req)) {
2091                 rc = PTR_ERR(req);
2092                 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
2093                        tgt_name(tgt), cfs_crypto_hash_name(cfs_alg), rc);
2094                 goto out;
2095         }
2096
2097         buffer = kmap(__page);
2098         guard_start = (__be16 *)buffer;
2099         guard_number = PAGE_SIZE / sizeof(*guard_start);
2100         if (unlikely(resend))
2101                 CDEBUG(D_PAGE | D_HA, "GRD tags per page = %u\n", guard_number);
2102         for (i = 0; i < npages; i++) {
2103                 bool use_t10_grd;
2104
2105                 /* corrupt the data before we compute the checksum, to
2106                  * simulate a client->OST data error */
2107                 if (i == 0 && opc == OST_WRITE &&
2108                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2109                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2110                         int len = local_nb[i].lnb_len;
2111                         struct page *np = tgt_page_to_corrupt;
2112
2113                         if (np) {
2114                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2115                                 char *ptr2 = page_address(np);
2116
2117                                 memcpy(ptr2 + off, ptr + off, len);
2118                                 memcpy(ptr2 + off, "bad3", min(4, len));
2119                                 kunmap_atomic(ptr);
2120
2121                                 /* LU-8376 to preserve original index for
2122                                  * display in dump_all_bulk_pages() */
2123                                 np->index = i;
2124
2125                                 cfs_crypto_hash_update_page(req, np, off,
2126                                                             len);
2127                                 continue;
2128                         } else {
2129                                 CERROR("%s: can't alloc page for corruption\n",
2130                                        tgt_name(tgt));
2131                         }
2132                 }
2133
2134                 /*
2135                  * The left guard number should be able to hold checksums of a
2136                  * whole page
2137                  */
2138                 use_t10_grd = t10_cksum_type && t10_cksum_type == cksum_type &&
2139                               opc == OST_READ &&
2140                               local_nb[i].lnb_len == PAGE_SIZE &&
2141                               local_nb[i].lnb_guard_disk;
2142                 if (use_t10_grd) {
2143                         used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2144                         if (used > (guard_number - used_number)) {
2145                                 rc = -E2BIG;
2146                                 break;
2147                         }
2148                         memcpy(guard_start + used_number,
2149                                local_nb[i].lnb_guards,
2150                                used * sizeof(*guard_start));
2151                         if (unlikely(resend))
2152                                 CDEBUG(D_PAGE | D_HA,
2153                                        "lnb[%u]: used %u off %u+%u lnb checksum: %*phN\n",
2154                                        i, used,
2155                                        local_nb[i].lnb_page_offset,
2156                                        local_nb[i].lnb_len,
2157                                        (int)(used * sizeof(*guard_start)),
2158                                        guard_start + used_number);
2159                 }
2160                 if (!use_t10_grd || unlikely(resend)) {
2161                         __be16 guard_tmp[MAX_GUARD_NUMBER];
2162                         __be16 *guards = guard_start + used_number;
2163                         int used_tmp = -1, *usedp = &used;
2164
2165                         if (unlikely(use_t10_grd)) {
2166                                 guards = guard_tmp;
2167                                 usedp = &used_tmp;
2168                         }
2169                         rc = obd_page_dif_generate_buffer(obd_name,
2170                                 local_nb[i].lnb_page,
2171                                 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2172                                 local_nb[i].lnb_len, guards,
2173                                 guard_number - used_number, usedp, sector_size,
2174                                 fn);
2175                         if (unlikely(resend)) {
2176                                 bool bad = use_t10_grd &&
2177                                         memcmp(guard_tmp,
2178                                                local_nb[i].lnb_guards,
2179                                                used_tmp * sizeof(*guard_tmp));
2180
2181                                 if (bad)
2182                                         CERROR("lnb[%u]: used %u/%u off %u+%u tmp checksum: %*phN\n",
2183                                                i, used, used_tmp,
2184                                                local_nb[i].lnb_page_offset,
2185                                                local_nb[i].lnb_len,
2186                                                (int)(used_tmp * sizeof(*guard_start)),
2187                                                guard_tmp);
2188                                 CDEBUG_LIMIT(D_PAGE | D_HA | (bad ? D_ERROR : 0),
2189                                        "lnb[%u]: used %u/%u off %u+%u gen checksum: %*phN\n",
2190                                        i, used, used_tmp,
2191                                        local_nb[i].lnb_page_offset,
2192                                        local_nb[i].lnb_len,
2193                                        (int)(used * sizeof(*guard_start)),
2194                                        guard_start + used_number);
2195                         }
2196                         if (rc)
2197                                 break;
2198                 }
2199
2200                 LASSERT(used <= MAX_GUARD_NUMBER);
2201                 /*
2202                  * If disk support T10PI checksum, copy guards to local_nb.
2203                  * If the write is partial page, do not use the guards for bio
2204                  * submission since the data might not be full-sector. The bio
2205                  * guards will be generated later based on the full sectors. If
2206                  * the sector size is 512B rather than 4 KB, or the page size
2207                  * is larger than 4KB, this might drop some useful guards for
2208                  * partial page write, but it will only add minimal extra time
2209                  * of checksum calculation.
2210                  */
2211                 if (t10_cksum_type && t10_cksum_type == cksum_type &&
2212                     opc == OST_WRITE &&
2213                     local_nb[i].lnb_len == PAGE_SIZE) {
2214                         local_nb[i].lnb_guard_rpc = 1;
2215                         memcpy(local_nb[i].lnb_guards,
2216                                guard_start + used_number,
2217                                used * sizeof(*local_nb[i].lnb_guards));
2218                 }
2219
2220                 used_number += used;
2221                 if (used_number == guard_number) {
2222                         cfs_crypto_hash_update_page(req, __page, 0,
2223                                 used_number * sizeof(*guard_start));
2224                         used_number = 0;
2225                 }
2226
2227                  /* corrupt the data after we compute the checksum, to
2228                  * simulate an OST->client data error */
2229                 if (unlikely(i == 0 && opc == OST_READ &&
2230                              OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2231                         int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2232                         int len = local_nb[i].lnb_len;
2233                         struct page *np = tgt_page_to_corrupt;
2234
2235                         if (np) {
2236                                 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2237                                 char *ptr2 = page_address(np);
2238
2239                                 memcpy(ptr2 + off, ptr + off, len);
2240                                 memcpy(ptr2 + off, "bad4", min(4, len));
2241                                 kunmap_atomic(ptr);
2242
2243                                 /* LU-8376 to preserve original index for
2244                                  * display in dump_all_bulk_pages() */
2245                                 np->index = i;
2246
2247                                 cfs_crypto_hash_update_page(req, np, off,
2248                                                             len);
2249                                 continue;
2250                         } else {
2251                                 CERROR("%s: can't alloc page for corruption\n",
2252                                        tgt_name(tgt));
2253                         }
2254                 }
2255         }
2256         kunmap(__page);
2257         if (rc)
2258                 GOTO(out, rc);
2259
2260         if (used_number != 0)
2261                 cfs_crypto_hash_update_page(req, __page, 0,
2262                         used_number * sizeof(*guard_start));
2263
2264         bufsize = sizeof(cksum);
2265         rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2266
2267         if (rc == 0)
2268                 *check_sum = cksum;
2269 out:
2270         __free_page(__page);
2271         return rc;
2272 }
2273
2274 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2275                                   enum cksum_types cksum_type,
2276                                   struct niobuf_local *local_nb,
2277                                   int npages, int opc, u32 *check_sum,
2278                                   bool resend)
2279 {
2280         obd_dif_csum_fn *fn = NULL;
2281         int sector_size = 0;
2282         int rc;
2283
2284         ENTRY;
2285         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
2286
2287         if (fn)
2288                 rc = tgt_checksum_niobuf_t10pi(tgt, cksum_type,
2289                                                local_nb, npages,
2290                                                opc, fn, sector_size,
2291                                                check_sum, resend);
2292         else
2293                 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2294                                          cksum_type, check_sum);
2295
2296         RETURN(rc);
2297 }
2298
2299 int tgt_brw_read(struct tgt_session_info *tsi)
2300 {
2301         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2302         struct ptlrpc_bulk_desc *desc = NULL;
2303         struct obd_export       *exp = tsi->tsi_exp;
2304         struct niobuf_remote    *remote_nb;
2305         struct niobuf_local     *local_nb;
2306         struct obd_ioobj        *ioo;
2307         struct ost_body         *body, *repbody;
2308         struct lustre_handle     lockh = { 0 };
2309         int                      npages, nob = 0, rc, i, no_reply = 0,
2310                                  npages_read;
2311         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2312         const char *obd_name = exp->exp_obd->obd_name;
2313         ktime_t kstart;
2314
2315         ENTRY;
2316
2317         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2318             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2319                 CERROR("%s: deny read request from %s to portal %u\n",
2320                        tgt_name(tsi->tsi_tgt),
2321                        obd_export_nid2str(req->rq_export),
2322                        ptlrpc_req2svc(req)->srv_req_portal);
2323                 RETURN(-EPROTO);
2324         }
2325
2326         req->rq_bulk_read = 1;
2327
2328         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) {
2329                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2330                  * this server to fail requests.
2331                  */
2332                 char fail_ost_name[MAX_OBD_NAME];
2333
2334                 if (cfs_fail_val > 0) {
2335                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2336                                  cfs_fail_val - 1);
2337
2338                         if (strstr(obd_name, fail_ost_name))
2339                                 RETURN(err_serious(-EIO));
2340                 } else {
2341                         RETURN(err_serious(-EIO));
2342                 }
2343         }
2344
2345         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2346                          cfs_fail_val : (obd_timeout + 1) / 4);
2347
2348         /* Check if there is eviction in progress, and if so, wait for it to
2349          * finish */
2350         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2351                 /* We do not care how long it takes */
2352                 wait_event_idle(
2353                         exp->exp_obd->obd_evict_inprogress_waitq,
2354                         !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2355         }
2356
2357         /* There must be big cache in current thread to process this request
2358          * if it is NULL then something went wrong and it wasn't allocated,
2359          * report -ENOMEM in that case */
2360         if (tbc == NULL)
2361                 RETURN(-ENOMEM);
2362
2363         body = tsi->tsi_ost_body;
2364         LASSERT(body != NULL);
2365
2366         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2367             body->oa.o_flags & OBD_FL_NORPC)
2368                 RETURN(0);
2369
2370         ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2371         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2372
2373         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2374         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2375
2376         local_nb = tbc->local;
2377
2378         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2379                           &lockh, LCK_PR);
2380         if (rc != 0)
2381                 RETURN(rc);
2382
2383         /*
2384          * If getting the lock took more time than
2385          * client was willing to wait, drop it. b=11330
2386          */
2387         if (ktime_get_real_seconds() > req->rq_deadline ||
2388             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2389                 no_reply = 1;
2390                 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2391                        libcfs_idstr(&req->rq_peer), POSTID(&ioo->ioo_oid),
2392                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2393                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2394                 GOTO(out_lock, rc = -ETIMEDOUT);
2395         }
2396
2397         /*
2398          * Because we already sync grant info with client when
2399          * reconnect, grant info will be cleared for resent req,
2400          * otherwise, outdated grant count in the rpc would de-sync
2401          * grant counters in case of shrink
2402          */
2403         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2404                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2405                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2406         }
2407
2408         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2409         repbody->oa = body->oa;
2410
2411         npages = PTLRPC_MAX_BRW_PAGES;
2412         kstart = ktime_get();
2413         rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2414                         ioo, remote_nb, &npages, local_nb);
2415         if (rc != 0)
2416                 GOTO(out_lock, rc);
2417
2418         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2419             body->oa.o_flags & OBD_FL_SHORT_IO) {
2420                 desc = NULL;
2421         } else {
2422                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2423                                             PTLRPC_BULK_PUT_SOURCE,
2424                                             OST_BULK_PORTAL,
2425                                             &ptlrpc_bulk_kiov_nopin_ops);
2426                 if (desc == NULL)
2427                         GOTO(out_commitrw, rc = -ENOMEM);
2428         }
2429
2430         npages_read = npages;
2431         for (i = 0; i < npages; i++) {
2432                 int page_rc = local_nb[i].lnb_rc;
2433
2434                 if (page_rc < 0) {
2435                         rc = page_rc;
2436                         npages_read = i;
2437                         break;
2438                 }
2439
2440                 nob += page_rc;
2441                 if (page_rc != 0 && desc != NULL) { /* some data! */
2442                         LASSERT(local_nb[i].lnb_page != NULL);
2443                         desc->bd_frag_ops->add_kiov_frag
2444                           (desc, local_nb[i].lnb_page,
2445                            local_nb[i].lnb_page_offset & ~PAGE_MASK,
2446                            page_rc);
2447                 }
2448
2449                 if (page_rc != local_nb[i].lnb_len) { /* short read */
2450                         local_nb[i].lnb_len = page_rc;
2451                         npages_read = i + (page_rc != 0 ? 1 : 0);
2452                         /* All subsequent pages should be 0 */
2453                         while (++i < npages)
2454                                 LASSERT(local_nb[i].lnb_rc == 0);
2455                         break;
2456                 }
2457         }
2458
2459         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2460                 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2461                            body->oa.o_flags : 0;
2462                 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2463                 bool resend = (body->oa.o_valid & OBD_MD_FLFLAGS) &&
2464                         (body->oa.o_flags & OBD_FL_RECOV_RESEND);
2465
2466                 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2467                                                           cksum_type);
2468                 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2469
2470                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2471                                             local_nb, npages_read, OST_READ,
2472                                             &repbody->oa.o_cksum, resend);
2473                 if (rc < 0)
2474                         GOTO(out_commitrw, rc);
2475                 CDEBUG(D_PAGE | (resend ? D_HA : 0),
2476                        "checksum at read origin: %x (%x)\n",
2477                        repbody->oa.o_cksum, cksum_type);
2478
2479                 /* if a resend it could be for a cksum error, so check Server
2480                  * cksum with returned Client cksum (this should even cover
2481                  * zero-cksum case) */
2482                 if (resend)
2483                         check_read_checksum(local_nb, npages_read, exp,
2484                                             &body->oa, &req->rq_peer,
2485                                             body->oa.o_cksum,
2486                                             repbody->oa.o_cksum, cksum_type);
2487         } else {
2488                 repbody->oa.o_valid = 0;
2489         }
2490         if (body->oa.o_valid & OBD_MD_FLGRANT)
2491                 repbody->oa.o_valid |= OBD_MD_FLGRANT;
2492         /* We're finishing using body->oa as an input variable */
2493
2494         /* Check if client was evicted while we were doing i/o before touching
2495          * network */
2496         if (rc == 0) {
2497                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2498                     body->oa.o_flags & OBD_FL_SHORT_IO) {
2499                         unsigned char *short_io_buf;
2500                         int short_io_size;
2501
2502                         short_io_buf = req_capsule_server_get(&req->rq_pill,
2503                                                               &RMF_SHORT_IO);
2504                         short_io_size = req_capsule_get_size(&req->rq_pill,
2505                                                              &RMF_SHORT_IO,
2506                                                              RCL_SERVER);
2507                         rc = tgt_pages2shortio(local_nb, npages_read,
2508                                                short_io_buf, short_io_size);
2509                         if (rc >= 0)
2510                                 req_capsule_shrink(&req->rq_pill,
2511                                                    &RMF_SHORT_IO, rc,
2512                                                    RCL_SERVER);
2513                         rc = rc > 0 ? 0 : rc;
2514                 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2515                         rc = target_bulk_io(exp, desc);
2516                 }
2517                 no_reply = rc != 0;
2518         } else {
2519                 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2520                     body->oa.o_flags & OBD_FL_SHORT_IO)
2521                         req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2522                                            RCL_SERVER);
2523         }
2524
2525 out_commitrw:
2526         /* Must commit after prep above in all cases */
2527         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2528                           remote_nb, npages, local_nb, rc, nob, kstart);
2529 out_lock:
2530         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2531
2532         if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2533                 ptlrpc_free_bulk(desc);
2534
2535         LASSERT(rc <= 0);
2536         if (rc == 0) {
2537                 rc = nob;
2538                 ptlrpc_lprocfs_brw(req, nob);
2539         } else if (no_reply) {
2540                 req->rq_no_reply = 1;
2541                 /* reply out callback would free */
2542                 ptlrpc_req_drop_rs(req);
2543                 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2544                               "client will retry: rc %d\n",
2545                               obd_name,
2546                               obd_uuid2str(&exp->exp_client_uuid),
2547                               obd_export_nid2str(exp), rc);
2548         }
2549         /* send a bulk after reply to simulate a network delay or reordering
2550          * by a router - Note that !desc implies short io, so there is no bulk
2551          * to reorder. */
2552         if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2553             desc) {
2554                 /* Calculate checksum before request transfer, original
2555                  * it is done by target_bulk_io() */
2556                 rc = sptlrpc_svc_wrap_bulk(req, desc);
2557                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
2558                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2559                 else /* old version, bulk matchbits is rq_xid */
2560                         req->rq_mbits = req->rq_xid;
2561
2562                 req->rq_status = rc;
2563                 target_committed_to_req(req);
2564                 target_send_reply(req, 0, 0);
2565
2566                 CDEBUG(D_INFO, "reorder BULK\n");
2567                 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
2568                                  cfs_fail_val ? : 3);
2569
2570                 target_bulk_io(exp, desc);
2571                 ptlrpc_free_bulk(desc);
2572         }
2573
2574         RETURN(rc);
2575 }
2576 EXPORT_SYMBOL(tgt_brw_read);
2577
2578 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2579                              unsigned char *buf, unsigned int size)
2580 {
2581         int     i, off, len;
2582         char    *ptr;
2583
2584         for (i = 0; i < npages; i++) {
2585                 off = local[i].lnb_page_offset & ~PAGE_MASK;
2586                 len = local[i].lnb_len;
2587
2588                 if (len == 0)
2589                         continue;
2590
2591                 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2592                        i, off, len, size);
2593                 ptr = kmap_atomic(local[i].lnb_page);
2594                 if (ptr == NULL)
2595                         return -EINVAL;
2596                 memcpy(ptr + off, buf, len < size ? len : size);
2597                 kunmap_atomic(ptr);
2598                 buf += len;
2599                 size -= len;
2600         }
2601         return 0;
2602 }
2603
2604 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2605                               struct ptlrpc_bulk_desc *desc,
2606                               struct niobuf_local *local_nb, int npages,
2607                               u32 client_cksum, u32 server_cksum,
2608                               bool mmap)
2609 {
2610         struct obd_export *exp = req->rq_export;
2611         struct ost_body *body;
2612         char *router = "";
2613         char *via = "";
2614
2615         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2616         LASSERT(body != NULL);
2617
2618         if (desc && !nid_same(&req->rq_peer.nid, &desc->bd_sender)) {
2619                 via = " via ";
2620                 router = libcfs_nidstr(&desc->bd_sender);
2621         }
2622
2623         if (exp->exp_obd->obd_checksum_dump)
2624                 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2625                                     client_cksum);
2626
2627         if (mmap) {
2628                 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2629                              client_cksum, server_cksum);
2630                 return;
2631         }
2632
2633         LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2634                            DFID" object "DOSTID" extent [%llu-%llu"
2635                            "]: client csum %x, server csum %x\n",
2636                            exp->exp_obd->obd_name, libcfs_idstr(&req->rq_peer),
2637                            via, router,
2638                            body->oa.o_valid & OBD_MD_FLFID ?
2639                            body->oa.o_parent_seq : (__u64)0,
2640                            body->oa.o_valid & OBD_MD_FLFID ?
2641                            body->oa.o_parent_oid : 0,
2642                            body->oa.o_valid & OBD_MD_FLFID ?
2643                            body->oa.o_parent_ver : 0,
2644                            POSTID(&body->oa.o_oi),
2645                            local_nb[0].lnb_file_offset,
2646                            local_nb[npages-1].lnb_file_offset +
2647                            local_nb[npages - 1].lnb_len - 1,
2648                            client_cksum, server_cksum);
2649 }
2650
2651 int tgt_brw_write(struct tgt_session_info *tsi)
2652 {
2653         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2654         struct ptlrpc_bulk_desc *desc = NULL;
2655         struct obd_export       *exp = req->rq_export;
2656         struct niobuf_remote    *remote_nb;
2657         struct niobuf_local     *local_nb;
2658         struct obd_ioobj        *ioo;
2659         struct ost_body         *body, *repbody;
2660         struct lustre_handle     lockh = {0};
2661         __u32                   *rcs;
2662         int                      objcount, niocount, npages;
2663         int                      rc = 0;
2664         int                      i, j;
2665         enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2666         bool                     no_reply = false, mmap;
2667         struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2668         bool wait_sync = false;
2669         const char *obd_name = exp->exp_obd->obd_name;
2670         /* '1' for consistency with code that checks !mpflag to restore */
2671         unsigned int mpflags = 1;
2672         ktime_t kstart;
2673         int nob = 0;
2674
2675         ENTRY;
2676
2677         if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2678             ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2679                 CERROR("%s: deny write request from %s to portal %u\n",
2680                        tgt_name(tsi->tsi_tgt),
2681                        obd_export_nid2str(req->rq_export),
2682                        ptlrpc_req2svc(req)->srv_req_portal);
2683                 RETURN(err_serious(-EPROTO));
2684         }
2685
2686         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2687                 RETURN(err_serious(-ENOSPC));
2688         if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2689                 RETURN(err_serious(-EROFS));
2690
2691         req->rq_bulk_write = 1;
2692
2693         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2694                 rc = -EIO;
2695         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2696                 rc = -EFAULT;
2697         if (rc < 0) {
2698                 /* optionally use cfs_fail_val - 1 to select a specific OST on
2699                  * this server to fail requests.
2700                  */
2701                 char fail_ost_name[MAX_OBD_NAME];
2702
2703                 if (cfs_fail_val > 0) {
2704                         snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2705                                  cfs_fail_val - 1);
2706
2707                         if (strstr(obd_name, fail_ost_name))
2708                                 RETURN(err_serious(rc));
2709                 } else {
2710                         RETURN(err_serious(rc));
2711                 }
2712         }
2713
2714         /* pause before transaction has been started */
2715         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2716                          cfs_fail_val : (obd_timeout + 1) / 4);
2717
2718         /* Delay write commit to show stale size information */
2719         CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2720
2721         /* There must be big cache in current thread to process this request
2722          * if it is NULL then something went wrong and it wasn't allocated,
2723          * report -ENOMEM in that case */
2724         if (tbc == NULL)
2725                 RETURN(-ENOMEM);
2726
2727         body = tsi->tsi_ost_body;
2728         LASSERT(body != NULL);
2729
2730         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2731             body->oa.o_flags & OBD_FL_NORPC)
2732                 RETURN(0);
2733
2734
2735         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2736         LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2737
2738         objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2739                                         RCL_CLIENT) / sizeof(*ioo);
2740
2741         for (niocount = i = 0; i < objcount; i++)
2742                 niocount += ioo[i].ioo_bufcnt;
2743
2744         remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2745         LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2746         if (niocount != req_capsule_get_size(&req->rq_pill,
2747                                              &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2748                         sizeof(*remote_nb))
2749                 RETURN(err_serious(-EPROTO));
2750
2751         if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2752             ptlrpc_connection_is_local(exp->exp_connection))
2753                 mpflags = memalloc_noreclaim_save();
2754
2755         req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2756                              niocount * sizeof(*rcs));
2757         rc = req_capsule_server_pack(&req->rq_pill);
2758         if (rc != 0)
2759                 GOTO(out, rc = err_serious(rc));
2760
2761         CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2762         rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2763
2764         local_nb = tbc->local;
2765
2766         rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2767                           &lockh, LCK_PW);
2768         if (rc != 0)
2769                 GOTO(out, rc);
2770
2771         /*
2772          * If getting the lock took more time than
2773          * client was willing to wait, drop it. b=11330
2774          */
2775         if (ktime_get_real_seconds() > req->rq_deadline ||
2776             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2777                 no_reply = true;
2778                 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2779                        tgt_name(tsi->tsi_tgt), libcfs_idstr(&req->rq_peer),
2780                        POSTID(&ioo->ioo_oid),
2781                        ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2782                        req->rq_deadline - req->rq_arrival_time.tv_sec);
2783                 GOTO(out_lock, rc = -ETIMEDOUT);
2784         }
2785
2786         /* Because we already sync grant info with client when reconnect,
2787          * grant info will be cleared for resent req, then fed_grant and
2788          * total_grant will not be modified in following preprw_write */
2789         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2790                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2791                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2792         }
2793
2794         repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2795         if (repbody == NULL)
2796                 GOTO(out_lock, rc = -ENOMEM);
2797         repbody->oa = body->oa;
2798
2799         npages = PTLRPC_MAX_BRW_PAGES;
2800         kstart = ktime_get();
2801         rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2802                         objcount, ioo, remote_nb, &npages, local_nb);
2803         if (rc < 0)
2804                 GOTO(out_lock, rc);
2805         if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2806             body->oa.o_flags & OBD_FL_SHORT_IO) {
2807                 unsigned int short_io_size;
2808                 unsigned char *short_io_buf;
2809
2810                 short_io_size = req_capsule_get_size(&req->rq_pill,
2811                                                      &RMF_SHORT_IO,
2812                                                      RCL_CLIENT);
2813                 short_io_buf = req_capsule_client_get(&req->rq_pill,
2814                                                       &RMF_SHORT_IO);
2815                 CDEBUG(D_INFO, "Client use short io for data transfer,"
2816                                " size = %d\n", short_io_size);
2817
2818                 /* Copy short io buf to pages */
2819                 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2820                                        short_io_size);
2821                 desc = NULL;
2822         } else {
2823                 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2824                                             PTLRPC_BULK_GET_SINK,
2825                                             OST_BULK_PORTAL,
2826                                             &ptlrpc_bulk_kiov_nopin_ops);
2827                 if (desc == NULL)
2828                         GOTO(skip_transfer, rc = -ENOMEM);
2829
2830                 /* NB Having prepped, we must commit... */
2831                 for (i = 0; i < npages; i++)
2832                         desc->bd_frag_ops->add_kiov_frag(desc,
2833                                         local_nb[i].lnb_page,
2834                                         local_nb[i].lnb_page_offset & ~PAGE_MASK,
2835                                         local_nb[i].lnb_len);
2836
2837                 rc = sptlrpc_svc_prep_bulk(req, desc);
2838                 if (rc != 0)
2839                         GOTO(skip_transfer, rc);
2840
2841                 rc = target_bulk_io(exp, desc);
2842         }
2843
2844         no_reply = rc != 0;
2845
2846 skip_transfer:
2847         if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2848                 static int cksum_counter;
2849
2850                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2851                         cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2852
2853                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2854                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2855                 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2856                                                            cksum_type);
2857
2858                 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2859                                             local_nb, npages, OST_WRITE,
2860                                             &repbody->oa.o_cksum, false);
2861                 if (rc < 0)
2862                         GOTO(out_commitrw, rc);
2863
2864                 cksum_counter++;
2865
2866                 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2867                         mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2868                                 body->oa.o_flags & OBD_FL_MMAP);
2869
2870                         tgt_warn_on_cksum(req, desc, local_nb, npages,
2871                                           body->oa.o_cksum,
2872                                           repbody->oa.o_cksum, mmap);
2873                         cksum_counter = 0;
2874                 } else if ((cksum_counter & (-cksum_counter)) ==
2875                            cksum_counter) {
2876                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2877                                cksum_counter, libcfs_idstr(&req->rq_peer),
2878                                repbody->oa.o_cksum);
2879                 }
2880         }
2881
2882         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2883
2884 out_commitrw:
2885         /* calculate the expected actual write bytes (nob) for OFD stats.
2886          * Technically, if commit fails this would be wrong, but that should be
2887          * very rare
2888          */
2889         for (i = 0; i < niocount; i++) {
2890                 int len = remote_nb[i].rnb_len;
2891
2892                 nob += len;
2893         }
2894
2895         /* multiple transactions can be assigned during write commit */
2896         tsi->tsi_mult_trans = 1;
2897
2898         /* Must commit after prep above in all cases */
2899         rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2900                           objcount, ioo, remote_nb, npages, local_nb, rc, nob,
2901                           kstart);
2902         if (rc == -ENOTCONN)
2903                 /* quota acquire process has been given up because
2904                  * either the client has been evicted or the client
2905                  * has timed out the request already
2906                  */
2907                 no_reply = true;
2908
2909         for (i = 0; i < niocount; i++) {
2910                 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2911                         wait_sync = true;
2912                         break;
2913                 }
2914         }
2915         /*
2916          * Disable sending mtime back to the client. If the client locked the
2917          * whole object, then it has already updated the mtime on its side,
2918          * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2919          */
2920         repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2921
2922         if (rc == 0) {
2923                 /* set per-requested niobuf return codes */
2924                 for (i = j = 0; i < niocount; i++) {
2925                         int len = remote_nb[i].rnb_len;
2926
2927                         rcs[i] = 0;
2928                         do {
2929                                 LASSERT(j < npages);
2930                                 if (local_nb[j].lnb_rc < 0)
2931                                         rcs[i] = local_nb[j].lnb_rc;
2932                                 len -= local_nb[j].lnb_len;
2933                                 j++;
2934                         } while (len > 0);
2935                         LASSERT(len == 0);
2936                 }
2937                 LASSERT(j == npages);
2938                 ptlrpc_lprocfs_brw(req, nob);
2939         }
2940 out_lock:
2941         tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2942         if (desc)
2943                 ptlrpc_free_bulk(desc);
2944 out:
2945         if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2946                 req->rq_no_reply = 1;
2947                 /* reply out callback would free */
2948                 ptlrpc_req_drop_rs(req);
2949                 if (!exp->exp_obd->obd_no_transno)
2950                         LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2951                                       " client will retry: rc = %d\n",
2952                                       obd_name,
2953                                       obd_uuid2str(&exp->exp_client_uuid),
2954                                       obd_export_nid2str(exp), rc);
2955         }
2956
2957         if (mpflags)
2958                 memalloc_noreclaim_restore(mpflags);
2959
2960         RETURN(rc);
2961 }
2962 EXPORT_SYMBOL(tgt_brw_write);
2963
2964 /**
2965  * Common request handler for OST_SEEK RPC.
2966  *
2967  * Unified request handling for OST_SEEK RPC.
2968  * It takes object by its FID, does needed lseek and packs result
2969  * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2970  *
2971  * \param[in] tsi       target session environment for this request
2972  *
2973  * \retval              0 if successful
2974  * \retval              negative value on error
2975  */
2976 int tgt_lseek(struct tgt_session_info *tsi)
2977 {
2978         struct lustre_handle lh = { 0 };
2979         struct dt_object *dob;
2980         struct ost_body *repbody;
2981         loff_t offset = tsi->tsi_ost_body->oa.o_size;
2982         int whence = tsi->tsi_ost_body->oa.o_mode;
2983         bool srvlock;
2984         int rc = 0;
2985
2986         ENTRY;
2987
2988         if (whence != SEEK_HOLE && whence != SEEK_DATA)
2989                 RETURN(-EPROTO);
2990
2991         /* Negative offset is prohibited on wire and must be handled on client
2992          * prior sending RPC.
2993          */
2994         if (offset < 0)
2995                 RETURN(-EPROTO);
2996
2997         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2998         if (repbody == NULL)
2999                 RETURN(-ENOMEM);
3000         repbody->oa = tsi->tsi_ost_body->oa;
3001
3002         srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
3003                   tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
3004         if (srvlock) {
3005                 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
3006                                    offset, OBD_OBJECT_EOF, &lh, LCK_PR);
3007                 if (rc)
3008                         RETURN(rc);
3009         }
3010
3011         dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
3012         if (IS_ERR(dob))
3013                 GOTO(out, rc = PTR_ERR(dob));
3014
3015         if (!dt_object_exists(dob))
3016                 GOTO(obj_put, rc = -ENOENT);
3017
3018         repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
3019         rc = 0;
3020 obj_put:
3021         dt_object_put(tsi->tsi_env, dob);
3022 out:
3023         if (srvlock)
3024                 tgt_data_unlock(&lh, LCK_PR);
3025
3026         RETURN(rc);
3027 }
3028 EXPORT_SYMBOL(tgt_lseek);
3029
3030 /* Check if request can be reconstructed from saved reply data
3031  * A copy of the reply data is returned in @trd if the pointer is not NULL
3032  */
3033 int req_can_reconstruct(struct ptlrpc_request *req,
3034                          struct tg_reply_data *trd)
3035 {
3036         struct tg_export_data *ted = &req->rq_export->exp_target_data;
3037         struct lsd_client_data *lcd = ted->ted_lcd;
3038         int found;
3039
3040         if (tgt_is_multimodrpcs_client(req->rq_export))
3041                 return tgt_lookup_reply(req, trd);
3042
3043         mutex_lock(&ted->ted_lcd_lock);
3044         found = req->rq_xid == lcd->lcd_last_xid ||
3045                 req->rq_xid == lcd->lcd_last_close_xid;
3046
3047         if (found && trd != NULL) {
3048                 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
3049                         trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
3050                         trd->trd_reply.lrd_transno =
3051                                                 lcd->lcd_last_close_transno;
3052                         trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
3053                 } else {
3054                         trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
3055                         trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
3056                         trd->trd_reply.lrd_result = lcd->lcd_last_result;
3057                         trd->trd_reply.lrd_data = lcd->lcd_last_data;
3058                         trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
3059                         trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
3060                         trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
3061                         trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
3062                 }
3063         }
3064         mutex_unlock(&ted->ted_lcd_lock);
3065
3066         return found;
3067 }
3068 EXPORT_SYMBOL(req_can_reconstruct);
3069
3070 bool tgt_check_resent(struct ptlrpc_request *req, struct tg_reply_data *trd)
3071 {
3072         struct lsd_reply_data *lrd;
3073         bool need_reconstruct = false;
3074
3075         ENTRY;
3076
3077         if (likely(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
3078                 return false;
3079
3080         if (req_can_reconstruct(req, trd)) {
3081                 lrd = &trd->trd_reply;
3082                 req->rq_transno = lrd->lrd_transno;
3083                 req->rq_status = lrd->lrd_result;
3084                 if (req->rq_status != 0)
3085                         req->rq_transno = 0;
3086                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
3087                 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
3088
3089                 DEBUG_REQ(D_HA, req, "reconstruct resent RPC");
3090                 need_reconstruct = true;
3091         } else {
3092                 DEBUG_REQ(D_HA, req, "no reply for RESENT req");
3093         }
3094
3095         return need_reconstruct;
3096 }
3097 EXPORT_SYMBOL(tgt_check_resent);