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, struct lookup_intent *it,
  63                              const struct lu_fid *parent_fid,
  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, it, &req, cb_blocking,
 130                             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         if (pmode) {
 146                 it->d.lustre.it_lock_handle = plock.cookie;
 147                 it->d.lustre.it_lock_mode = pmode;
 148         }
 149
 150         EXIT;
 151 out_free_op_data:
 152         OBD_FREE_PTR(op_data);
 153 out:
 154         if (rc && pmode)
 155                 ldlm_lock_decref(&plock, pmode);
 156
 157         ptlrpc_req_finished(*reqp);
 158         *reqp = req;
 159         return rc;
 160 }
 161
 162 #ifdef __KERNEL__
 163 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 164                           struct lmv_stripe_md *lsm,
 165                           ldlm_blocking_callback cb_blocking,
 166                           int extra_lock_flags)
 167 {
 168         struct obd_device      *obd = exp->exp_obd;
 169         struct lmv_obd         *lmv = &obd->u.lmv;
 170         struct mdt_body         *body;
 171         struct md_op_data      *op_data;
 172         unsigned long           size = 0;
 173         unsigned long           nlink = 0;
 174         obd_time                atime = 0;
 175         obd_time                ctime = 0;
 176         obd_time                mtime = 0;
 177         int                     i;
 178         int                     rc = 0;
 179
 180         ENTRY;
 181
 182         /**
 183          * revalidate slaves has some problems, temporarily return,
 184          * we may not need that
 185          */
 186         if (lsm->lsm_md_stripe_count <= 1)
 187                 RETURN(0);
 188
 189         OBD_ALLOC_PTR(op_data);
 190         if (op_data == NULL)
 191                 RETURN(-ENOMEM);
 192
 193         /**
 194          * Loop over the stripe information, check validity and update them
 195          * from MDS if needed.
 196          */
 197         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 198                 struct lu_fid           fid;
 199                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 200                 struct ptlrpc_request   *req = NULL;
 201                 struct lustre_handle    *lockh = NULL;
 202                 struct lmv_tgt_desc     *tgt = NULL;
 203                 struct inode            *inode;
 204
 205                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 206                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 207                 if (i == 0) {
 208                         if (mbody != NULL) {
 209                                 body = mbody;
 210                                 goto update;
 211                         } else {
 212                                 goto release_lock;
 213                         }
 214                 }
 215
 216                 /*
 217                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 218                  * defined otherwise it will go to server and take new lock
 219                  * which is not needed here.
 220                  */
 221                 memset(op_data, 0, sizeof(*op_data));
 222                 op_data->op_fid1 = fid;
 223                 op_data->op_fid2 = fid;
 224
 225                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 226                 if (IS_ERR(tgt))
 227                         GOTO(cleanup, rc = PTR_ERR(tgt));
 228
 229                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%d\n",
 230                        PFID(&fid), tgt->ltd_idx);
 231
 232                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 233                                     cb_blocking, extra_lock_flags);
 234                 if (rc < 0)
 235                         GOTO(cleanup, rc);
 236
 237                 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
 238                 if (rc > 0 && req == NULL) {
 239                         /* slave inode is still valid */
 240                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 241                                PFID(&fid));
 242                         rc = 0;
 243                 } else {
 244                         /* refresh slave from server */
 245                         body = req_capsule_server_get(&req->rq_pill,
 246                                                       &RMF_MDT_BODY);
 247                         LASSERT(body != NULL);
 248 update:
 249                         if (unlikely(body->nlink < 2)) {
 250                                 CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID
 251                                        ":" DFID"\n", obd->obd_name, body->nlink,
 252                                        i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
 253                                        PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 254
 255                                 if (req != NULL)
 256                                         ptlrpc_req_finished(req);
 257
 258                                 if (it.d.lustre.it_lock_mode && lockh) {
 259                                         ldlm_lock_decref(lockh,
 260                                                  it.d.lustre.it_lock_mode);
 261                                         it.d.lustre.it_lock_mode = 0;
 262                                 }
 263
 264                                 GOTO(cleanup, rc = -EIO);
 265                         }
 266
 267                         if (i != 0)
 268                                 md_set_lock_data(tgt->ltd_exp, &lockh->cookie,
 269                                                  inode, NULL);
 270
 271                         i_size_write(inode, body->size);
 272                         set_nlink(inode, body->nlink);
 273                         LTIME_S(inode->i_atime) = body->atime;
 274                         LTIME_S(inode->i_ctime) = body->ctime;
 275                         LTIME_S(inode->i_mtime) = body->mtime;
 276
 277                         if (req != NULL)
 278                                 ptlrpc_req_finished(req);
 279                 }
 280 release_lock:
 281                 size += i_size_read(inode);
 282
 283                 if (i != 0)
 284                         nlink += inode->i_nlink - 2;
 285                 else
 286                         nlink += inode->i_nlink;
 287
 288                 atime = LTIME_S(inode->i_atime) > atime ?
 289                                 LTIME_S(inode->i_atime) : atime;
 290                 ctime = LTIME_S(inode->i_ctime) > ctime ?
 291                                 LTIME_S(inode->i_ctime) : ctime;
 292                 mtime = LTIME_S(inode->i_mtime) > mtime ?
 293                                 LTIME_S(inode->i_mtime) : mtime;
 294
 295                 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
 296                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 297                         it.d.lustre.it_lock_mode = 0;
 298                 }
 299
 300                 CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
 301                        "%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
 302                        i_size_read(inode), inode->i_nlink,
 303                        LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
 304                        LTIME_S(inode->i_ctime));
 305         }
 306
 307         /*
 308          * update attr of master request.
 309          */
 310         CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
 311                LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
 312                atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 313
 314         if (mbody != NULL) {
 315                 mbody->atime = atime;
 316                 mbody->ctime = ctime;
 317                 mbody->mtime = mtime;
 318         }
 319 cleanup:
 320         OBD_FREE_PTR(op_data);
 321         RETURN(rc);
 322 }
 323
 324 #else
 325
 326 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 327                           struct lmv_stripe_md *lsm,
 328                           ldlm_blocking_callback cb_blocking,
 329                           int extra_lock_flags)
 330 {
 331         return 0;
 332 }
 333
 334 #endif
 335
 336 /*
 337  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 338  * may be split dir.
 339  */
 340 int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 341                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 342                     ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
 343 {
 344         struct obd_device       *obd = exp->exp_obd;
 345         struct lmv_obd          *lmv = &obd->u.lmv;
 346         struct lmv_tgt_desc     *tgt;
 347         struct mdt_body         *body;
 348         int                     rc;
 349         ENTRY;
 350
 351         /* Note: client might open with some random flags(sanity 33b), so we can
 352          * not make sure op_fid2 is being initialized with BY_FID flag */
 353         if (it->it_flags & MDS_OPEN_BY_FID && fid_is_sane(&op_data->op_fid2)) {
 354                 if (op_data->op_mea1 != NULL) {
 355                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 356                         const struct lmv_oinfo  *oinfo;
 357
 358                         oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
 359                                                         op_data->op_namelen);
 360                         if (IS_ERR(oinfo))
 361                                 RETURN(PTR_ERR(oinfo));
 362                         op_data->op_fid1 = oinfo->lmo_fid;
 363                 }
 364
 365                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 366                 if (IS_ERR(tgt))
 367                         RETURN(PTR_ERR(tgt));
 368
 369                 op_data->op_mds = tgt->ltd_idx;
 370         } else {
 371                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 372                 if (IS_ERR(tgt))
 373                         RETURN(PTR_ERR(tgt));
 374         }
 375
 376         /* If it is ready to open the file by FID, do not need
 377          * allocate FID at all, otherwise it will confuse MDT */
 378         if ((it->it_op & IT_CREAT) &&
 379             !(it->it_flags & MDS_OPEN_BY_FID)) {
 380                 /*
 381                  * For open with IT_CREATE and for IT_CREATE cases allocate new
 382                  * fid and setup FLD for it.
 383                  */
 384                 op_data->op_fid3 = op_data->op_fid2;
 385                 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
 386                 if (rc != 0)
 387                         RETURN(rc);
 388         }
 389
 390         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 391                " name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
 392                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 393
 394         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 395                             extra_lock_flags);
 396         if (rc != 0)
 397                 RETURN(rc);
 398         /*
 399          * Nothing is found, do not access body->fid1 as it is zero and thus
 400          * pointless.
 401          */
 402         if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
 403             !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
 404             !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
 405                 RETURN(rc);
 406
 407         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 408         if (body == NULL)
 409                 RETURN(-EPROTO);
 410
 411         /* Not cross-ref case, just get out of here. */
 412         if (unlikely((body->valid & OBD_MD_MDS))) {
 413                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 414                                        cb_blocking, extra_lock_flags);
 415                 if (rc != 0)
 416                         RETURN(rc);
 417
 418                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 419                 if (body == NULL)
 420                         RETURN(-EPROTO);
 421         }
 422
 423         RETURN(rc);
 424 }
 425
 426 /*
 427  * Handler for: getattr, lookup and revalidate cases.
 428  */
 429 static int
 430 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 431                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 432                   ldlm_blocking_callback cb_blocking,
 433                   __u64 extra_lock_flags)
 434 {
 435         struct obd_device       *obd = exp->exp_obd;
 436         struct lmv_obd          *lmv = &obd->u.lmv;
 437         struct lmv_tgt_desc     *tgt = NULL;
 438         struct mdt_body         *body;
 439         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 440         int                     rc = 0;
 441         ENTRY;
 442
 443         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 444         if (IS_ERR(tgt))
 445                 RETURN(PTR_ERR(tgt));
 446
 447         if (!fid_is_sane(&op_data->op_fid2))
 448                 fid_zero(&op_data->op_fid2);
 449
 450         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 451                ", name='%s' -> mds #%d lsm=%p lsm_magic=%x\n",
 452                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 453                op_data->op_name ? op_data->op_name : "<NULL>",
 454                tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
 455
 456         op_data->op_bias &= ~MDS_CROSS_REF;
 457
 458         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 459                             extra_lock_flags);
 460         if (rc < 0)
 461                 RETURN(rc);
 462
 463         if (*reqp == NULL) {
 464                 /* If RPC happens, lsm information will be revalidated
 465                  * during update_inode process (see ll_update_lsm_md) */
 466                 if (op_data->op_mea2 != NULL) {
 467                         rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
 468                                                    cb_blocking,
 469                                                    extra_lock_flags);
 470                         if (rc != 0)
 471                                 RETURN(rc);
 472                 }
 473                 RETURN(rc);
 474         } else if (it_disposition(it, DISP_LOOKUP_NEG) &&
 475                    lsm != NULL && lsm->lsm_md_magic == LMV_MAGIC_MIGRATE) {
 476                 /* For migrating directory, if it can not find the child in
 477                  * the source directory(master stripe), try the targeting
 478                  * directory(stripe 1) */
 479                 tgt = lmv_find_target(lmv, &lsm->lsm_md_oinfo[1].lmo_fid);
 480                 if (IS_ERR(tgt))
 481                         RETURN(PTR_ERR(tgt));
 482
 483                 ptlrpc_req_finished(*reqp);
 484                 it->d.lustre.it_data = NULL;
 485                 *reqp = NULL;
 486
 487                 CDEBUG(D_INODE, "For migrating dir, try target dir "DFID"\n",
 488                        PFID(&lsm->lsm_md_oinfo[1].lmo_fid));
 489
 490                 op_data->op_fid1 = lsm->lsm_md_oinfo[1].lmo_fid;
 491                 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
 492                 rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
 493                                     cb_blocking, extra_lock_flags);
 494                 RETURN(rc);
 495         }
 496         /*
 497          * MDS has returned success. Probably name has been resolved in
 498          * remote inode. Let's check this.
 499          */
 500         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 501         if (body == NULL)
 502                 RETURN(-EPROTO);
 503
 504         /* Not cross-ref case, just get out of here. */
 505         if (unlikely((body->valid & OBD_MD_MDS))) {
 506                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 507                                        extra_lock_flags);
 508                 if (rc != 0)
 509                         RETURN(rc);
 510                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 511                 if (body == NULL)
 512                         RETURN(-EPROTO);
 513         }
 514
 515         RETURN(rc);
 516 }
 517
 518 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 519                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 520                     ldlm_blocking_callback cb_blocking,
 521                     __u64 extra_lock_flags)
 522 {
 523         struct obd_device *obd = exp->exp_obd;
 524         int                rc;
 525         ENTRY;
 526
 527         LASSERT(it != NULL);
 528         LASSERT(fid_is_sane(&op_data->op_fid1));
 529
 530         CDEBUG(D_INODE, "INTENT LOCK '%s' for '%*s' on "DFID"\n",
 531                LL_IT2STR(it), op_data->op_namelen, op_data->op_name,
 532                PFID(&op_data->op_fid1));
 533
 534         rc = lmv_check_connect(obd);
 535         if (rc)
 536                 RETURN(rc);
 537
 538         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 539                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 540                                        extra_lock_flags);
 541         else if (it->it_op & IT_OPEN)
 542                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 543                                      extra_lock_flags);
 544         else
 545                 LBUG();
 546
 547         RETURN(rc);
 548 }