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 <asm/div64.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 <obd_class.h>
  58 #include <lprocfs_status.h>
  59 #include "lmv_internal.h"
  60
  61 static int lmv_intent_remote(struct obd_export *exp, void *lmm,
  62                              int lmmsize, struct lookup_intent *it,
  63                              const struct lu_fid *parent_fid, int flags,
  64                              struct ptlrpc_request **reqp,
  65                              ldlm_blocking_callback cb_blocking,
  66                              __u64 extra_lock_flags)
  67 {
  68         struct obd_device       *obd = exp->exp_obd;
  69         struct lmv_obd          *lmv = &obd->u.lmv;
  70         struct ptlrpc_request   *req = NULL;
  71         struct lustre_handle    plock;
  72         struct md_op_data       *op_data;
  73         struct lmv_tgt_desc     *tgt;
  74         struct mdt_body         *body;
  75         int                     pmode;
  76         int                     rc = 0;
  77         ENTRY;
  78
  79         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
  80         if (body == NULL)
  81                 RETURN(-EPROTO);
  82
  83         LASSERT((body->valid & OBD_MD_MDS));
  84
  85         /*
  86          * Unfortunately, we have to lie to MDC/MDS to retrieve
  87          * attributes llite needs and provideproper locking.
  88          */
  89         if (it->it_op & IT_LOOKUP)
  90                 it->it_op = IT_GETATTR;
  91
  92         /*
  93          * We got LOOKUP lock, but we really need attrs.
  94          */
  95         pmode = it->d.lustre.it_lock_mode;
  96         if (pmode) {
  97                 plock.cookie = it->d.lustre.it_lock_handle;
  98                 it->d.lustre.it_lock_mode = 0;
  99                 it->d.lustre.it_data = NULL;
 100         }
 101
 102         LASSERT(fid_is_sane(&body->fid1));
 103
 104         tgt = lmv_find_target(lmv, &body->fid1);
 105         if (IS_ERR(tgt))
 106                 GOTO(out, rc = PTR_ERR(tgt));
 107
 108         OBD_ALLOC_PTR(op_data);
 109         if (op_data == NULL)
 110                 GOTO(out, rc = -ENOMEM);
 111
 112         op_data->op_fid1 = body->fid1;
 113         /* Sent the parent FID to the remote MDT */
 114         if (parent_fid != NULL) {
 115                 /* The parent fid is only for remote open to
 116                  * check whether the open is from OBF,
 117                  * see mdt_cross_open */
 118                 LASSERT(it->it_op & IT_OPEN);
 119                 op_data->op_fid2 = *parent_fid;
 120                 /* Add object FID to op_fid3, in case it needs to check stale
 121                  * (M_CHECK_STALE), see mdc_finish_intent_lock */
 122                 op_data->op_fid3 = body->fid1;
 123         }
 124
 125         op_data->op_bias = MDS_CROSS_REF;
 126         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%d\n",
 127                PFID(&body->fid1), tgt->ltd_idx);
 128
 129         rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
 130                             flags, &req, cb_blocking, extra_lock_flags);
 131         if (rc)
 132                 GOTO(out_free_op_data, rc);
 133
 134         /*
 135          * LLite needs LOOKUP lock to track dentry revocation in order to
 136          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 137          * and put LOOKUP in request.
 138          */
 139         if (it->d.lustre.it_lock_mode != 0) {
 140                 it->d.lustre.it_remote_lock_handle =
 141                                         it->d.lustre.it_lock_handle;
 142                 it->d.lustre.it_remote_lock_mode = it->d.lustre.it_lock_mode;
 143         }
 144
 145         it->d.lustre.it_lock_handle = plock.cookie;
 146         it->d.lustre.it_lock_mode = pmode;
 147
 148         EXIT;
 149 out_free_op_data:
 150         OBD_FREE_PTR(op_data);
 151 out:
 152         if (rc && pmode)
 153                 ldlm_lock_decref(&plock, pmode);
 154
 155         ptlrpc_req_finished(*reqp);
 156         *reqp = req;
 157         return rc;
 158 }
 159
 160 #ifdef __KERNEL__
 161 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 162                           struct lmv_stripe_md *lsm,
 163                           ldlm_blocking_callback cb_blocking,
 164                           int extra_lock_flags)
 165 {
 166         struct obd_device      *obd = exp->exp_obd;
 167         struct lmv_obd         *lmv = &obd->u.lmv;
 168         struct mdt_body         *body;
 169         struct md_op_data      *op_data;
 170         unsigned long           size = 0;
 171         unsigned long           nlink = 0;
 172         obd_time                atime = 0;
 173         obd_time                ctime = 0;
 174         obd_time                mtime = 0;
 175         int                     i;
 176         int                     rc = 0;
 177
 178         ENTRY;
 179
 180         /**
 181          * revalidate slaves has some problems, temporarily return,
 182          * we may not need that
 183          */
 184         if (lsm->lsm_md_stripe_count <= 1)
 185                 RETURN(0);
 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                 if (i == 0) {
 206                         if (mbody != NULL) {
 207                                 body = mbody;
 208                                 goto update;
 209                         } else {
 210                                 goto release_lock;
 211                         }
 212                 }
 213
 214                 /*
 215                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 216                  * defined otherwise it will go to server and take new lock
 217                  * which is not needed here.
 218                  */
 219                 memset(op_data, 0, sizeof(*op_data));
 220                 op_data->op_fid1 = fid;
 221                 op_data->op_fid2 = fid;
 222
 223                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 224                 if (IS_ERR(tgt))
 225                         GOTO(cleanup, rc = PTR_ERR(tgt));
 226
 227                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%d\n",
 228                        PFID(&fid), tgt->ltd_idx);
 229
 230                 rc = md_intent_lock(tgt->ltd_exp, op_data, NULL, 0, &it, 0,
 231                                     &req, cb_blocking, extra_lock_flags);
 232                 if (rc < 0)
 233                         GOTO(cleanup, rc);
 234
 235                 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
 236                 if (rc > 0 && req == NULL) {
 237                         /* slave inode is still valid */
 238                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 239                                PFID(&fid));
 240                         rc = 0;
 241                 } else {
 242                         /* refresh slave from server */
 243                         body = req_capsule_server_get(&req->rq_pill,
 244                                                       &RMF_MDT_BODY);
 245                         LASSERT(body != NULL);
 246 update:
 247                         if (unlikely(body->nlink < 2)) {
 248                                 CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID
 249                                        ":" DFID"\n", obd->obd_name, body->nlink,
 250                                        i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
 251                                        PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 252
 253                                 if (req != NULL)
 254                                         ptlrpc_req_finished(req);
 255
 256                                 GOTO(cleanup, rc = -EIO);
 257                         }
 258
 259                         if (i != 0)
 260                                 md_set_lock_data(tgt->ltd_exp, &lockh->cookie,
 261                                                  inode, NULL);
 262
 263                         i_size_write(inode, body->size);
 264                         set_nlink(inode, body->nlink);
 265                         LTIME_S(inode->i_atime) = body->atime;
 266                         LTIME_S(inode->i_ctime) = body->ctime;
 267                         LTIME_S(inode->i_mtime) = body->mtime;
 268
 269                         if (req != NULL)
 270                                 ptlrpc_req_finished(req);
 271                 }
 272 release_lock:
 273                 size += i_size_read(inode);
 274
 275                 if (i != 0)
 276                         nlink += inode->i_nlink - 2;
 277                 else
 278                         nlink += inode->i_nlink;
 279
 280                 atime = LTIME_S(inode->i_atime) > atime ?
 281                                 LTIME_S(inode->i_atime) : atime;
 282                 ctime = LTIME_S(inode->i_ctime) > ctime ?
 283                                 LTIME_S(inode->i_ctime) : ctime;
 284                 mtime = LTIME_S(inode->i_mtime) > mtime ?
 285                                 LTIME_S(inode->i_mtime) : mtime;
 286
 287                 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
 288                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 289                         it.d.lustre.it_lock_mode = 0;
 290                 }
 291
 292                 CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
 293                        "%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
 294                        i_size_read(inode), inode->i_nlink,
 295                        LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
 296                        LTIME_S(inode->i_ctime));
 297         }
 298
 299         /*
 300          * update attr of master request.
 301          */
 302         CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
 303                LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
 304                atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 305
 306         if (mbody != NULL) {
 307                 mbody->atime = atime;
 308                 mbody->ctime = ctime;
 309                 mbody->mtime = mtime;
 310         }
 311 cleanup:
 312         OBD_FREE_PTR(op_data);
 313         RETURN(rc);
 314 }
 315
 316 #else
 317
 318 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 319                           struct lmv_stripe_md *lsm,
 320                           ldlm_blocking_callback cb_blocking,
 321                           int extra_lock_flags)
 322 {
 323         return 0;
 324 }
 325
 326 #endif
 327
 328 /*
 329  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 330  * may be split dir.
 331  */
 332 int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 333                     void *lmm, int lmmsize, struct lookup_intent *it,
 334                     int flags, struct ptlrpc_request **reqp,
 335                     ldlm_blocking_callback cb_blocking,
 336                     __u64 extra_lock_flags)
 337 {
 338         struct obd_device       *obd = exp->exp_obd;
 339         struct lmv_obd          *lmv = &obd->u.lmv;
 340         struct lmv_tgt_desc     *tgt;
 341         struct mdt_body         *body;
 342         int                     rc;
 343         ENTRY;
 344
 345         /* Note: client might open with some random flags(sanity 33b), so we can
 346          * not make sure op_fid2 is being initialized with BY_FID flag */
 347         if (it->it_flags & MDS_OPEN_BY_FID && fid_is_sane(&op_data->op_fid2)) {
 348                 if (op_data->op_mea1 != NULL) {
 349                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 350                         const struct lmv_oinfo  *oinfo;
 351
 352                         oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
 353                                                         op_data->op_namelen);
 354                         op_data->op_fid1 = oinfo->lmo_fid;
 355                 }
 356
 357                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 358                 if (IS_ERR(tgt))
 359                         RETURN(PTR_ERR(tgt));
 360
 361                 op_data->op_mds = tgt->ltd_idx;
 362         } else {
 363                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 364                 if (IS_ERR(tgt))
 365                         RETURN(PTR_ERR(tgt));
 366         }
 367
 368         /* If it is ready to open the file by FID, do not need
 369          * allocate FID at all, otherwise it will confuse MDT */
 370         if ((it->it_op & IT_CREAT) &&
 371             !(it->it_flags & MDS_OPEN_BY_FID)) {
 372                 /*
 373                  * For open with IT_CREATE and for IT_CREATE cases allocate new
 374                  * fid and setup FLD for it.
 375                  */
 376                 op_data->op_fid3 = op_data->op_fid2;
 377                 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
 378                 if (rc != 0)
 379                         RETURN(rc);
 380         }
 381
 382         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 383                " name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
 384                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 385
 386         rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it, flags,
 387                             reqp, cb_blocking, extra_lock_flags);
 388         if (rc != 0)
 389                 RETURN(rc);
 390         /*
 391          * Nothing is found, do not access body->fid1 as it is zero and thus
 392          * pointless.
 393          */
 394         if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
 395             !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
 396             !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
 397                 RETURN(rc);
 398
 399         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 400         if (body == NULL)
 401                 RETURN(-EPROTO);
 402
 403         /* Not cross-ref case, just get out of here. */
 404         if (unlikely((body->valid & OBD_MD_MDS))) {
 405                 rc = lmv_intent_remote(exp, lmm, lmmsize, it, &op_data->op_fid1,
 406                                        flags, reqp, cb_blocking,
 407                                        extra_lock_flags);
 408                 if (rc != 0)
 409                         RETURN(rc);
 410
 411                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 412                 if (body == NULL)
 413                         RETURN(-EPROTO);
 414         }
 415
 416         RETURN(rc);
 417 }
 418
 419 /*
 420  * Handler for: getattr, lookup and revalidate cases.
 421  */
 422 int lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 423                       void *lmm, int lmmsize, struct lookup_intent *it,
 424                       int flags, struct ptlrpc_request **reqp,
 425                       ldlm_blocking_callback cb_blocking,
 426                       __u64 extra_lock_flags)
 427 {
 428         struct obd_device      *obd = exp->exp_obd;
 429         struct lmv_obd         *lmv = &obd->u.lmv;
 430         struct lmv_tgt_desc    *tgt = NULL;
 431         struct mdt_body        *body;
 432         int                     rc = 0;
 433         ENTRY;
 434
 435         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 436         if (IS_ERR(tgt))
 437                 RETURN(PTR_ERR(tgt));
 438
 439         if (!fid_is_sane(&op_data->op_fid2))
 440                 fid_zero(&op_data->op_fid2);
 441
 442         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 443                ", name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
 444                PFID(&op_data->op_fid2),
 445                op_data->op_name ? op_data->op_name : "<NULL>",
 446                tgt->ltd_idx);
 447
 448         op_data->op_bias &= ~MDS_CROSS_REF;
 449
 450         rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
 451                              flags, reqp, cb_blocking, extra_lock_flags);
 452
 453         if (rc < 0)
 454                 RETURN(rc);
 455
 456         if (*reqp == NULL) {
 457                 /* If RPC happens, lsm information will be revalidated
 458                  * during update_inode process (see ll_update_lsm_md) */
 459                 if (op_data->op_mea2 != NULL) {
 460                         rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
 461                                                 cb_blocking, extra_lock_flags);
 462                         if (rc != 0)
 463                                 RETURN(rc);
 464                 }
 465                 RETURN(rc);
 466         }
 467
 468         /*
 469          * MDS has returned success. Probably name has been resolved in
 470          * remote inode. Let's check this.
 471          */
 472         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 473         if (body == NULL)
 474                 RETURN(-EPROTO);
 475
 476         /* Not cross-ref case, just get out of here. */
 477         if (unlikely((body->valid & OBD_MD_MDS))) {
 478                 rc = lmv_intent_remote(exp, lmm, lmmsize, it, NULL, flags,
 479                                        reqp, cb_blocking, extra_lock_flags);
 480                 if (rc != 0)
 481                         RETURN(rc);
 482                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 483                 if (body == NULL)
 484                         RETURN(-EPROTO);
 485         }
 486
 487         RETURN(rc);
 488 }
 489
 490 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 491                     void *lmm, int lmmsize, struct lookup_intent *it,
 492                     int flags, struct ptlrpc_request **reqp,
 493                     ldlm_blocking_callback cb_blocking,
 494                     __u64 extra_lock_flags)
 495 {
 496         struct obd_device *obd = exp->exp_obd;
 497         int                rc;
 498         ENTRY;
 499
 500         LASSERT(it != NULL);
 501         LASSERT(fid_is_sane(&op_data->op_fid1));
 502
 503         CDEBUG(D_INODE, "INTENT LOCK '%s' for '%*s' on "DFID"\n",
 504                LL_IT2STR(it), op_data->op_namelen, op_data->op_name,
 505                PFID(&op_data->op_fid1));
 506
 507         rc = lmv_check_connect(obd);
 508         if (rc)
 509                 RETURN(rc);
 510
 511         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 512                 rc = lmv_intent_lookup(exp, op_data, lmm, lmmsize, it,
 513                                        flags, reqp, cb_blocking,
 514                                        extra_lock_flags);
 515         else if (it->it_op & IT_OPEN)
 516                 rc = lmv_intent_open(exp, op_data, lmm, lmmsize, it,
 517                                      flags, reqp, cb_blocking,
 518                                      extra_lock_flags);
 519         else
 520                 LBUG();
 521         RETURN(rc);
 522 }