lustre/lmv/lmv_intent.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, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LMV
  38 #ifdef __KERNEL__
  39 #include <linux/slab.h>
  40 #include <linux/module.h>
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/math64.h>
  45 #include <linux/seq_file.h>
  46 #include <linux/namei.h>
  47 #include <linux/lustre_intent.h>
  48 #else
  49 #include <liblustre.h>
  50 #endif
  51
  52 #include <obd_support.h>
  53 #include <lustre/lustre_idl.h>
  54 #include <lustre_lib.h>
  55 #include <lustre_net.h>
  56 #include <lustre_dlm.h>
  57 #include <lustre_mdc.h>
  58 #include <obd_class.h>
  59 #include <lprocfs_status.h>
  60 #include "lmv_internal.h"
  61
  62 static int lmv_intent_remote(struct obd_export *exp, void *lmm,
  63                              int lmmsize, struct lookup_intent *it,
  64                              const struct lu_fid *parent_fid, int flags,
  65                              struct ptlrpc_request **reqp,
  66                              ldlm_blocking_callback cb_blocking,
  67                              __u64 extra_lock_flags)
  68 {
  69         struct obd_device       *obd = exp->exp_obd;
  70         struct lmv_obd          *lmv = &obd->u.lmv;
  71         struct ptlrpc_request   *req = NULL;
  72         struct lustre_handle    plock;
  73         struct md_op_data       *op_data;
  74         struct lmv_tgt_desc     *tgt;
  75         struct mdt_body         *body;
  76         int                     pmode;
  77         int                     rc = 0;
  78         ENTRY;
  79
  80         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
  81         if (body == NULL)
  82                 RETURN(-EPROTO);
  83
  84         LASSERT((body->valid & OBD_MD_MDS));
  85
  86         /*
  87          * Unfortunately, we have to lie to MDC/MDS to retrieve
  88          * attributes llite needs and provideproper locking.
  89          */
  90         if (it->it_op & IT_LOOKUP)
  91                 it->it_op = IT_GETATTR;
  92
  93         /*
  94          * We got LOOKUP lock, but we really need attrs.
  95          */
  96         pmode = it->d.lustre.it_lock_mode;
  97         if (pmode) {
  98                 plock.cookie = it->d.lustre.it_lock_handle;
  99                 it->d.lustre.it_lock_mode = 0;
 100                 it->d.lustre.it_data = NULL;
 101         }
 102
 103         LASSERT(fid_is_sane(&body->fid1));
 104
 105         tgt = lmv_find_target(lmv, &body->fid1);
 106         if (IS_ERR(tgt))
 107                 GOTO(out, rc = PTR_ERR(tgt));
 108
 109         OBD_ALLOC_PTR(op_data);
 110         if (op_data == NULL)
 111                 GOTO(out, rc = -ENOMEM);
 112
 113         op_data->op_fid1 = body->fid1;
 114         /* Sent the parent FID to the remote MDT */
 115         if (parent_fid != NULL) {
 116                 /* The parent fid is only for remote open to
 117                  * check whether the open is from OBF,
 118                  * see mdt_cross_open */
 119                 LASSERT(it->it_op & IT_OPEN);
 120                 op_data->op_fid2 = *parent_fid;
 121                 /* Add object FID to op_fid3, in case it needs to check stale
 122                  * (M_CHECK_STALE), see mdc_finish_intent_lock */
 123                 op_data->op_fid3 = body->fid1;
 124         }
 125
 126         op_data->op_bias = MDS_CROSS_REF;
 127         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%d\n",
 128                PFID(&body->fid1), tgt->ltd_idx);
 129
 130         rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
 131                             flags, &req, cb_blocking, extra_lock_flags);
 132         if (rc)
 133                 GOTO(out_free_op_data, rc);
 134
 135         /*
 136          * LLite needs LOOKUP lock to track dentry revocation in order to
 137          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 138          * and put LOOKUP in request.
 139          */
 140         if (it->d.lustre.it_lock_mode != 0) {
 141                 it->d.lustre.it_remote_lock_handle =
 142                                         it->d.lustre.it_lock_handle;
 143                 it->d.lustre.it_remote_lock_mode = it->d.lustre.it_lock_mode;
 144         }
 145
 146         if (pmode) {
 147                 it->d.lustre.it_lock_handle = plock.cookie;
 148                 it->d.lustre.it_lock_mode = pmode;
 149         }
 150
 151         EXIT;
 152 out_free_op_data:
 153         OBD_FREE_PTR(op_data);
 154 out:
 155         if (rc && pmode)
 156                 ldlm_lock_decref(&plock, pmode);
 157
 158         ptlrpc_req_finished(*reqp);
 159         *reqp = req;
 160         return rc;
 161 }
 162
 163 #ifdef __KERNEL__
 164 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 165                           struct lmv_stripe_md *lsm,
 166                           ldlm_blocking_callback cb_blocking,
 167                           int extra_lock_flags)
 168 {
 169         struct obd_device      *obd = exp->exp_obd;
 170         struct lmv_obd         *lmv = &obd->u.lmv;
 171         struct mdt_body         *body;
 172         struct md_op_data      *op_data;
 173         unsigned long           size = 0;
 174         unsigned long           nlink = 0;
 175         obd_time                atime = 0;
 176         obd_time                ctime = 0;
 177         obd_time                mtime = 0;
 178         int                     i;
 179         int                     rc = 0;
 180
 181         ENTRY;
 182
 183         /**
 184          * revalidate slaves has some problems, temporarily return,
 185          * we may not need that
 186          */
 187         OBD_ALLOC_PTR(op_data);
 188         if (op_data == NULL)
 189                 RETURN(-ENOMEM);
 190
 191         /**
 192          * Loop over the stripe information, check validity and update them
 193          * from MDS if needed.
 194          */
 195         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 196                 struct lu_fid           fid;
 197                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 198                 struct ptlrpc_request   *req = NULL;
 199                 struct lustre_handle    *lockh = NULL;
 200                 struct lmv_tgt_desc     *tgt = NULL;
 201                 struct inode            *inode;
 202
 203                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 204                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 205
 206                 /*
 207                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 208                  * defined otherwise it will go to server and take new lock
 209                  * which is not needed here.
 210                  */
 211                 memset(op_data, 0, sizeof(*op_data));
 212                 op_data->op_fid1 = fid;
 213                 op_data->op_fid2 = fid;
 214
 215                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 216                 if (IS_ERR(tgt))
 217                         GOTO(cleanup, rc = PTR_ERR(tgt));
 218
 219                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%d\n",
 220                        PFID(&fid), tgt->ltd_idx);
 221
 222                 rc = md_intent_lock(tgt->ltd_exp, op_data, NULL, 0, &it, 0,
 223                                     &req, cb_blocking, extra_lock_flags);
 224                 if (rc < 0)
 225                         GOTO(cleanup, rc);
 226
 227                 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
 228                 if (rc > 0 && req == NULL) {
 229                         /* slave inode is still valid */
 230                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 231                                PFID(&fid));
 232                         rc = 0;
 233                 } else {
 234                         /* refresh slave from server */
 235                         body = req_capsule_server_get(&req->rq_pill,
 236                                                       &RMF_MDT_BODY);
 237                         LASSERT(body != NULL);
 238                         if (unlikely(body->nlink < 2)) {
 239                                 CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID
 240                                        ":" DFID"\n", obd->obd_name, body->nlink,
 241                                        i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
 242                                        PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 243
 244                                 if (req != NULL)
 245                                         ptlrpc_req_finished(req);
 246
 247                                 if (it.d.lustre.it_lock_mode && lockh) {
 248                                         ldlm_lock_decref(lockh,
 249                                                  it.d.lustre.it_lock_mode);
 250                                         it.d.lustre.it_lock_mode = 0;
 251                                 }
 252
 253                                 GOTO(cleanup, rc = -EIO);
 254                         }
 255
 256
 257                         i_size_write(inode, body->size);
 258                         set_nlink(inode, body->nlink);
 259                         LTIME_S(inode->i_atime) = body->atime;
 260                         LTIME_S(inode->i_ctime) = body->ctime;
 261                         LTIME_S(inode->i_mtime) = body->mtime;
 262
 263                         if (req != NULL)
 264                                 ptlrpc_req_finished(req);
 265                 }
 266
 267                 md_set_lock_data(tgt->ltd_exp, &lockh->cookie, inode, NULL);
 268
 269                 size += i_size_read(inode);
 270
 271                 if (i != 0)
 272                         nlink += inode->i_nlink - 2;
 273                 else
 274                         nlink += inode->i_nlink;
 275
 276                 atime = LTIME_S(inode->i_atime) > atime ?
 277                                 LTIME_S(inode->i_atime) : atime;
 278                 ctime = LTIME_S(inode->i_ctime) > ctime ?
 279                                 LTIME_S(inode->i_ctime) : ctime;
 280                 mtime = LTIME_S(inode->i_mtime) > mtime ?
 281                                 LTIME_S(inode->i_mtime) : mtime;
 282
 283                 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
 284                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 285                         it.d.lustre.it_lock_mode = 0;
 286                 }
 287
 288                 CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
 289                        "%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
 290                        i_size_read(inode), inode->i_nlink,
 291                        LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
 292                        LTIME_S(inode->i_ctime));
 293         }
 294
 295         /*
 296          * update attr of master request.
 297          */
 298         CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
 299                LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
 300                atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 301
 302         if (mbody != NULL) {
 303                 mbody->atime = atime;
 304                 mbody->ctime = ctime;
 305                 mbody->mtime = mtime;
 306         }
 307 cleanup:
 308         OBD_FREE_PTR(op_data);
 309         RETURN(rc);
 310 }
 311
 312 #else
 313
 314 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 315                           struct lmv_stripe_md *lsm,
 316                           ldlm_blocking_callback cb_blocking,
 317                           int extra_lock_flags)
 318 {
 319         return 0;
 320 }
 321
 322 #endif
 323
 324 /*
 325  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 326  * may be split dir.
 327  */
 328 int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 329                     void *lmm, int lmmsize, struct lookup_intent *it,
 330                     int flags, struct ptlrpc_request **reqp,
 331                     ldlm_blocking_callback cb_blocking,
 332                     __u64 extra_lock_flags)
 333 {
 334         struct obd_device       *obd = exp->exp_obd;
 335         struct lmv_obd          *lmv = &obd->u.lmv;
 336         struct lmv_tgt_desc     *tgt;
 337         struct mdt_body         *body;
 338         int                     rc;
 339         ENTRY;
 340
 341         /* Note: client might open with some random flags(sanity 33b), so we can
 342          * not make sure op_fid2 is being initialized with BY_FID flag */
 343         if (it->it_flags & MDS_OPEN_BY_FID && fid_is_sane(&op_data->op_fid2)) {
 344                 if (op_data->op_mea1 != NULL) {
 345                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 346                         const struct lmv_oinfo  *oinfo;
 347
 348                         oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
 349                                                         op_data->op_namelen);
 350                         if (IS_ERR(oinfo))
 351                                 RETURN(PTR_ERR(oinfo));
 352                         op_data->op_fid1 = oinfo->lmo_fid;
 353                 }
 354
 355                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 356                 if (IS_ERR(tgt))
 357                         RETURN(PTR_ERR(tgt));
 358
 359                 op_data->op_mds = tgt->ltd_idx;
 360         } else {
 361                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 362                 if (IS_ERR(tgt))
 363                         RETURN(PTR_ERR(tgt));
 364         }
 365
 366         /* If it is ready to open the file by FID, do not need
 367          * allocate FID at all, otherwise it will confuse MDT */
 368         if ((it->it_op & IT_CREAT) &&
 369             !(it->it_flags & MDS_OPEN_BY_FID)) {
 370                 /*
 371                  * For open with IT_CREATE and for IT_CREATE cases allocate new
 372                  * fid and setup FLD for it.
 373                  */
 374                 op_data->op_fid3 = op_data->op_fid2;
 375                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 376                 if (rc != 0)
 377                         RETURN(rc);
 378         }
 379
 380         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 381                " name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
 382                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 383
 384         rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it, flags,
 385                             reqp, cb_blocking, extra_lock_flags);
 386         if (rc != 0)
 387                 RETURN(rc);
 388         /*
 389          * Nothing is found, do not access body->fid1 as it is zero and thus
 390          * pointless.
 391          */
 392         if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
 393             !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
 394             !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
 395                 RETURN(rc);
 396
 397         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 398         if (body == NULL)
 399                 RETURN(-EPROTO);
 400
 401         /* Not cross-ref case, just get out of here. */
 402         if (unlikely((body->valid & OBD_MD_MDS))) {
 403                 rc = lmv_intent_remote(exp, lmm, lmmsize, it, &op_data->op_fid1,
 404                                        flags, reqp, cb_blocking,
 405                                        extra_lock_flags);
 406                 if (rc != 0)
 407                         RETURN(rc);
 408
 409                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 410                 if (body == NULL)
 411                         RETURN(-EPROTO);
 412         }
 413
 414         RETURN(rc);
 415 }
 416
 417 /*
 418  * Handler for: getattr, lookup and revalidate cases.
 419  */
 420 int lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 421                       void *lmm, int lmmsize, struct lookup_intent *it,
 422                       int flags, struct ptlrpc_request **reqp,
 423                       ldlm_blocking_callback cb_blocking,
 424                       __u64 extra_lock_flags)
 425 {
 426         struct obd_device       *obd = exp->exp_obd;
 427         struct lmv_obd          *lmv = &obd->u.lmv;
 428         struct lmv_tgt_desc     *tgt = NULL;
 429         struct mdt_body         *body;
 430         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 431         int                     rc = 0;
 432         ENTRY;
 433
 434         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 435         if (IS_ERR(tgt))
 436                 RETURN(PTR_ERR(tgt));
 437
 438         if (!fid_is_sane(&op_data->op_fid2))
 439                 fid_zero(&op_data->op_fid2);
 440
 441         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 442                ", name='%s' -> mds #%d lsm=%p lsm_magic=%x\n",
 443                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 444                op_data->op_name ? op_data->op_name : "<NULL>",
 445                tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
 446
 447         op_data->op_bias &= ~MDS_CROSS_REF;
 448
 449         rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
 450                              flags, reqp, cb_blocking, extra_lock_flags);
 451         if (rc < 0)
 452                 RETURN(rc);
 453
 454         if (*reqp == NULL) {
 455                 /* If RPC happens, lsm information will be revalidated
 456                  * during update_inode process (see ll_update_lsm_md) */
 457                 if (op_data->op_mea2 != NULL) {
 458                         rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
 459                                                    cb_blocking,
 460                                                    extra_lock_flags);
 461                         if (rc != 0)
 462                                 RETURN(rc);
 463                 }
 464                 RETURN(rc);
 465         } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
 466                    lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION) {
 467                 /* For migrating directory, if it can not find the child in
 468                  * the source directory(master stripe), try the targeting
 469                  * directory(stripe 1) */
 470                 tgt = lmv_find_target(lmv, &lsm->lsm_md_oinfo[1].lmo_fid);
 471                 if (IS_ERR(tgt))
 472                         RETURN(PTR_ERR(tgt));
 473
 474                 ptlrpc_req_finished(*reqp);
 475                 it->d.lustre.it_data = NULL;
 476                 *reqp = NULL;
 477
 478                 CDEBUG(D_INODE, "For migrating dir, try target dir "DFID"\n",
 479                        PFID(&lsm->lsm_md_oinfo[1].lmo_fid));
 480
 481                 op_data->op_fid1 = lsm->lsm_md_oinfo[1].lmo_fid;
 482                 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
 483                 rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
 484                                     flags, reqp, cb_blocking, extra_lock_flags);
 485         }
 486         /*
 487          * MDS has returned success. Probably name has been resolved in
 488          * remote inode. Let's check this.
 489          */
 490         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 491         if (body == NULL)
 492                 RETURN(-EPROTO);
 493
 494         /* Not cross-ref case, just get out of here. */
 495         if (unlikely((body->valid & OBD_MD_MDS))) {
 496                 rc = lmv_intent_remote(exp, lmm, lmmsize, it, NULL, flags,
 497                                        reqp, cb_blocking, extra_lock_flags);
 498                 if (rc != 0)
 499                         RETURN(rc);
 500                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 501                 if (body == NULL)
 502                         RETURN(-EPROTO);
 503         }
 504
 505         RETURN(rc);
 506 }
 507
 508 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 509                     void *lmm, int lmmsize, struct lookup_intent *it,
 510                     int flags, struct ptlrpc_request **reqp,
 511                     ldlm_blocking_callback cb_blocking,
 512                     __u64 extra_lock_flags)
 513 {
 514         struct obd_device *obd = exp->exp_obd;
 515         int                rc;
 516         ENTRY;
 517
 518         LASSERT(it != NULL);
 519         LASSERT(fid_is_sane(&op_data->op_fid1));
 520
 521         CDEBUG(D_INODE, "INTENT LOCK '%s' for '%*s' on "DFID"\n",
 522                LL_IT2STR(it), op_data->op_namelen, op_data->op_name,
 523                PFID(&op_data->op_fid1));
 524
 525         rc = lmv_check_connect(obd);
 526         if (rc)
 527                 RETURN(rc);
 528
 529         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 530                 rc = lmv_intent_lookup(exp, op_data, lmm, lmmsize, it,
 531                                        flags, reqp, cb_blocking,
 532                                        extra_lock_flags);
 533         else if (it->it_op & IT_OPEN)
 534                 rc = lmv_intent_open(exp, op_data, lmm, lmmsize, it,
 535                                      flags, reqp, cb_blocking,
 536                                      extra_lock_flags);
 537         else
 538                 LBUG();
 539         RETURN(rc);
 540 }