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 #include <linux/slab.h>
  39 #include <linux/module.h>
  40 #include <linux/init.h>
  41 #include <linux/slab.h>
  42 #include <linux/pagemap.h>
  43 #include <linux/math64.h>
  44 #include <linux/seq_file.h>
  45 #include <linux/namei.h>
  46 #include <lustre_intent.h>
  47
  48 #include <obd_support.h>
  49 #include <lustre/lustre_idl.h>
  50 #include <lustre_lib.h>
  51 #include <lustre_net.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_mdc.h>
  54 #include <obd_class.h>
  55 #include <lprocfs_status.h>
  56 #include "lmv_internal.h"
  57
  58 static int lmv_intent_remote(struct obd_export *exp, struct lookup_intent *it,
  59                              const struct lu_fid *parent_fid,
  60                              struct ptlrpc_request **reqp,
  61                              ldlm_blocking_callback cb_blocking,
  62                              __u64 extra_lock_flags)
  63 {
  64         struct obd_device       *obd = exp->exp_obd;
  65         struct lmv_obd          *lmv = &obd->u.lmv;
  66         struct ptlrpc_request   *req = NULL;
  67         struct lustre_handle    plock;
  68         struct md_op_data       *op_data;
  69         struct lmv_tgt_desc     *tgt;
  70         struct mdt_body         *body;
  71         int                     pmode;
  72         int                     rc = 0;
  73         ENTRY;
  74
  75         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
  76         if (body == NULL)
  77                 RETURN(-EPROTO);
  78
  79         LASSERT((body->mbo_valid & OBD_MD_MDS));
  80
  81         /*
  82          * Unfortunately, we have to lie to MDC/MDS to retrieve
  83          * attributes llite needs and provideproper locking.
  84          */
  85         if (it->it_op & IT_LOOKUP)
  86                 it->it_op = IT_GETATTR;
  87
  88         /*
  89          * We got LOOKUP lock, but we really need attrs.
  90          */
  91         pmode = it->d.lustre.it_lock_mode;
  92         if (pmode) {
  93                 plock.cookie = it->d.lustre.it_lock_handle;
  94                 it->d.lustre.it_lock_mode = 0;
  95                 it->d.lustre.it_data = NULL;
  96         }
  97
  98         LASSERT(fid_is_sane(&body->mbo_fid1));
  99
 100         tgt = lmv_find_target(lmv, &body->mbo_fid1);
 101         if (IS_ERR(tgt))
 102                 GOTO(out, rc = PTR_ERR(tgt));
 103
 104         OBD_ALLOC_PTR(op_data);
 105         if (op_data == NULL)
 106                 GOTO(out, rc = -ENOMEM);
 107
 108         op_data->op_fid1 = body->mbo_fid1;
 109         /* Sent the parent FID to the remote MDT */
 110         if (parent_fid != NULL) {
 111                 /* The parent fid is only for remote open to
 112                  * check whether the open is from OBF,
 113                  * see mdt_cross_open */
 114                 LASSERT(it->it_op & IT_OPEN);
 115                 op_data->op_fid2 = *parent_fid;
 116         }
 117
 118         op_data->op_bias = MDS_CROSS_REF;
 119         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%u\n",
 120                PFID(&body->mbo_fid1), tgt->ltd_idx);
 121
 122         rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
 123                             extra_lock_flags);
 124         if (rc)
 125                 GOTO(out_free_op_data, rc);
 126
 127         /*
 128          * LLite needs LOOKUP lock to track dentry revocation in order to
 129          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 130          * and put LOOKUP in request.
 131          */
 132         if (it->d.lustre.it_lock_mode != 0) {
 133                 it->d.lustre.it_remote_lock_handle =
 134                                         it->d.lustre.it_lock_handle;
 135                 it->d.lustre.it_remote_lock_mode = it->d.lustre.it_lock_mode;
 136         }
 137
 138         if (pmode) {
 139                 it->d.lustre.it_lock_handle = plock.cookie;
 140                 it->d.lustre.it_lock_mode = pmode;
 141         }
 142
 143         EXIT;
 144 out_free_op_data:
 145         OBD_FREE_PTR(op_data);
 146 out:
 147         if (rc && pmode)
 148                 ldlm_lock_decref(&plock, pmode);
 149
 150         ptlrpc_req_finished(*reqp);
 151         *reqp = req;
 152         return rc;
 153 }
 154
 155 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 156                           struct lmv_stripe_md *lsm,
 157                           ldlm_blocking_callback cb_blocking,
 158                           int extra_lock_flags)
 159 {
 160         struct obd_device      *obd = exp->exp_obd;
 161         struct lmv_obd         *lmv = &obd->u.lmv;
 162         struct ptlrpc_request   *req = NULL;
 163         struct mdt_body         *body;
 164         struct md_op_data      *op_data;
 165         unsigned long           size = 0;
 166         unsigned long           nlink = 0;
 167         obd_time                atime = 0;
 168         obd_time                ctime = 0;
 169         obd_time                mtime = 0;
 170         int                     i;
 171         int                     rc = 0;
 172
 173         ENTRY;
 174
 175         /**
 176          * revalidate slaves has some problems, temporarily return,
 177          * we may not need that
 178          */
 179         OBD_ALLOC_PTR(op_data);
 180         if (op_data == NULL)
 181                 RETURN(-ENOMEM);
 182
 183         /**
 184          * Loop over the stripe information, check validity and update them
 185          * from MDS if needed.
 186          */
 187         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 188                 struct lu_fid           fid;
 189                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 190                 struct lustre_handle    *lockh = NULL;
 191                 struct lmv_tgt_desc     *tgt = NULL;
 192                 struct inode            *inode;
 193
 194                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 195                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 196
 197                 /*
 198                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 199                  * defined otherwise it will go to server and take new lock
 200                  * which is not needed here.
 201                  */
 202                 memset(op_data, 0, sizeof(*op_data));
 203                 op_data->op_fid1 = fid;
 204                 op_data->op_fid2 = fid;
 205
 206                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 207                 if (IS_ERR(tgt))
 208                         GOTO(cleanup, rc = PTR_ERR(tgt));
 209
 210                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 211                        PFID(&fid), tgt->ltd_idx);
 212
 213                 if (req != NULL) {
 214                         ptlrpc_req_finished(req);
 215                         req = NULL;
 216                 }
 217
 218                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 219                                     cb_blocking, extra_lock_flags);
 220                 if (rc < 0)
 221                         GOTO(cleanup, rc);
 222
 223                 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
 224                 if (rc > 0 && req == NULL) {
 225                         /* slave inode is still valid */
 226                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 227                                PFID(&fid));
 228                         rc = 0;
 229                 } else {
 230                         /* refresh slave from server */
 231                         body = req_capsule_server_get(&req->rq_pill,
 232                                                       &RMF_MDT_BODY);
 233                         LASSERT(body != NULL);
 234                         if (unlikely(body->mbo_nlink < 2)) {
 235                                 CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID
 236                                        ":" DFID"\n",
 237                                        obd->obd_name, body->mbo_nlink, i,
 238                                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
 239                                        PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 240
 241                                 if (it.d.lustre.it_lock_mode && lockh) {
 242                                         ldlm_lock_decref(lockh,
 243                                                  it.d.lustre.it_lock_mode);
 244                                         it.d.lustre.it_lock_mode = 0;
 245                                 }
 246
 247                                 GOTO(cleanup, rc = -EIO);
 248                         }
 249
 250
 251                         i_size_write(inode, body->mbo_size);
 252                         set_nlink(inode, body->mbo_nlink);
 253                         LTIME_S(inode->i_atime) = body->mbo_atime;
 254                         LTIME_S(inode->i_ctime) = body->mbo_ctime;
 255                         LTIME_S(inode->i_mtime) = body->mbo_mtime;
 256                 }
 257
 258                 md_set_lock_data(tgt->ltd_exp, &lockh->cookie, inode, NULL);
 259
 260                 size += i_size_read(inode);
 261
 262                 if (i != 0)
 263                         nlink += inode->i_nlink - 2;
 264                 else
 265                         nlink += inode->i_nlink;
 266
 267                 atime = LTIME_S(inode->i_atime) > atime ?
 268                                 LTIME_S(inode->i_atime) : atime;
 269                 ctime = LTIME_S(inode->i_ctime) > ctime ?
 270                                 LTIME_S(inode->i_ctime) : ctime;
 271                 mtime = LTIME_S(inode->i_mtime) > mtime ?
 272                                 LTIME_S(inode->i_mtime) : mtime;
 273
 274                 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
 275                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 276                         it.d.lustre.it_lock_mode = 0;
 277                 }
 278
 279                 CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
 280                        "%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
 281                        i_size_read(inode), inode->i_nlink,
 282                        LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
 283                        LTIME_S(inode->i_ctime));
 284         }
 285
 286         /*
 287          * update attr of master request.
 288          */
 289         CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
 290                LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
 291                atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 292
 293         if (mbody != NULL) {
 294                 mbody->mbo_atime = atime;
 295                 mbody->mbo_ctime = ctime;
 296                 mbody->mbo_mtime = mtime;
 297         }
 298 cleanup:
 299         if (req != NULL)
 300                 ptlrpc_req_finished(req);
 301
 302         OBD_FREE_PTR(op_data);
 303         RETURN(rc);
 304 }
 305
 306
 307 /*
 308  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 309  * may be split dir.
 310  */
 311 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 312                            struct lookup_intent *it,
 313                            struct ptlrpc_request **reqp,
 314                            ldlm_blocking_callback cb_blocking,
 315                            __u64 extra_lock_flags)
 316 {
 317         struct obd_device       *obd = exp->exp_obd;
 318         struct lmv_obd          *lmv = &obd->u.lmv;
 319         struct lmv_tgt_desc     *tgt;
 320         struct mdt_body         *body;
 321         int                     rc;
 322         ENTRY;
 323
 324         if (it->it_flags & MDS_OPEN_BY_FID) {
 325                 LASSERT(fid_is_sane(&op_data->op_fid2));
 326
 327                 /* for striped directory, we can't know parent stripe fid
 328                  * without name, but we can set it to child fid, and MDT
 329                  * will obtain it from linkea in open in such case. */
 330                 if (op_data->op_mea1 != NULL)
 331                         op_data->op_fid1 = op_data->op_fid2;
 332
 333                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 334                 if (IS_ERR(tgt))
 335                         RETURN(PTR_ERR(tgt));
 336
 337                 op_data->op_mds = tgt->ltd_idx;
 338         } else {
 339                 LASSERT(fid_is_sane(&op_data->op_fid1));
 340                 LASSERT(fid_is_zero(&op_data->op_fid2));
 341                 LASSERT(op_data->op_name != NULL);
 342
 343                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 344                 if (IS_ERR(tgt))
 345                         RETURN(PTR_ERR(tgt));
 346         }
 347
 348         /* If it is ready to open the file by FID, do not need
 349          * allocate FID at all, otherwise it will confuse MDT */
 350         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 351                 /*
 352                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 353                  * for it.
 354                  */
 355                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 356                 if (rc != 0)
 357                         RETURN(rc);
 358         }
 359
 360         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 361                " name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
 362                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 363
 364         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 365                             extra_lock_flags);
 366         if (rc != 0)
 367                 RETURN(rc);
 368         /*
 369          * Nothing is found, do not access body->fid1 as it is zero and thus
 370          * pointless.
 371          */
 372         if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
 373             !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
 374             !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
 375                 RETURN(rc);
 376
 377         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 378         if (body == NULL)
 379                 RETURN(-EPROTO);
 380
 381         /* Not cross-ref case, just get out of here. */
 382         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 383                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 384                                        cb_blocking, extra_lock_flags);
 385                 if (rc != 0)
 386                         RETURN(rc);
 387
 388                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 389                 if (body == NULL)
 390                         RETURN(-EPROTO);
 391         }
 392
 393         RETURN(rc);
 394 }
 395
 396 /*
 397  * Handler for: getattr, lookup and revalidate cases.
 398  */
 399 static int
 400 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 401                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 402                   ldlm_blocking_callback cb_blocking,
 403                   __u64 extra_lock_flags)
 404 {
 405         struct obd_device       *obd = exp->exp_obd;
 406         struct lmv_obd          *lmv = &obd->u.lmv;
 407         struct lmv_tgt_desc     *tgt = NULL;
 408         struct mdt_body         *body;
 409         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 410         int                     rc = 0;
 411         ENTRY;
 412
 413         /* If it returns ERR_PTR(-EBADFD) then it is an unknown hash type
 414          * it will try all stripes to locate the object */
 415         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 416         if (IS_ERR(tgt) && (PTR_ERR(tgt) != -EBADFD))
 417                 RETURN(PTR_ERR(tgt));
 418
 419         /* Both migrating dir and unknown hash dir need to try
 420          * all of sub-stripes */
 421         if (lsm != NULL && !lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
 422                 struct lmv_oinfo *oinfo;
 423
 424                 oinfo = &lsm->lsm_md_oinfo[0];
 425
 426                 op_data->op_fid1 = oinfo->lmo_fid;
 427                 op_data->op_mds = oinfo->lmo_mds;
 428                 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
 429                 if (IS_ERR(tgt))
 430                         RETURN(PTR_ERR(tgt));
 431         }
 432
 433         if (!fid_is_sane(&op_data->op_fid2))
 434                 fid_zero(&op_data->op_fid2);
 435
 436         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 437                ", name='%s' -> mds #%u lsm=%p lsm_magic=%x\n",
 438                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 439                op_data->op_name ? op_data->op_name : "<NULL>",
 440                tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
 441
 442         op_data->op_bias &= ~MDS_CROSS_REF;
 443
 444         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 445                             extra_lock_flags);
 446         if (rc < 0)
 447                 RETURN(rc);
 448
 449         if (*reqp == NULL) {
 450                 /* If RPC happens, lsm information will be revalidated
 451                  * during update_inode process (see ll_update_lsm_md) */
 452                 if (op_data->op_mea2 != NULL) {
 453                         rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
 454                                                    cb_blocking,
 455                                                    extra_lock_flags);
 456                         if (rc != 0)
 457                                 RETURN(rc);
 458                 }
 459                 RETURN(rc);
 460         } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
 461                    lmv_need_try_all_stripes(lsm)) {
 462                 /* For migrating and unknown hash type directory, it will
 463                  * try to target the entry on other stripes */
 464                 int stripe_index;
 465
 466                 for (stripe_index = 1;
 467                      stripe_index < lsm->lsm_md_stripe_count &&
 468                      it_disposition(it, DISP_LOOKUP_NEG); stripe_index++) {
 469                         struct lmv_oinfo *oinfo;
 470
 471                         /* release the previous request */
 472                         ptlrpc_req_finished(*reqp);
 473                         it->d.lustre.it_data = NULL;
 474                         *reqp = NULL;
 475
 476                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
 477                         tgt = lmv_find_target(lmv, &oinfo->lmo_fid);
 478                         if (IS_ERR(tgt))
 479                                 RETURN(PTR_ERR(tgt));
 480
 481                         CDEBUG(D_INODE, "Try other stripes " DFID"\n",
 482                                PFID(&oinfo->lmo_fid));
 483
 484                         op_data->op_fid1 = oinfo->lmo_fid;
 485                         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
 486                         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
 487                                             cb_blocking, extra_lock_flags);
 488                         if (rc != 0)
 489                                 RETURN(rc);
 490                 }
 491         }
 492
 493         /*
 494          * MDS has returned success. Probably name has been resolved in
 495          * remote inode. Let's check this.
 496          */
 497         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 498         if (body == NULL)
 499                 RETURN(-EPROTO);
 500
 501         /* Not cross-ref case, just get out of here. */
 502         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 503                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 504                                        extra_lock_flags);
 505                 if (rc != 0)
 506                         RETURN(rc);
 507                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 508                 if (body == NULL)
 509                         RETURN(-EPROTO);
 510         }
 511
 512         RETURN(rc);
 513 }
 514
 515 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 516                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 517                     ldlm_blocking_callback cb_blocking,
 518                     __u64 extra_lock_flags)
 519 {
 520         struct obd_device *obd = exp->exp_obd;
 521         int                rc;
 522         ENTRY;
 523
 524         LASSERT(it != NULL);
 525         LASSERT(fid_is_sane(&op_data->op_fid1));
 526
 527         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 528                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 529                 (int)op_data->op_namelen, op_data->op_name,
 530                 PFID(&op_data->op_fid1));
 531
 532         rc = lmv_check_connect(obd);
 533         if (rc)
 534                 RETURN(rc);
 535
 536         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 537                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 538                                        extra_lock_flags);
 539         else if (it->it_op & IT_OPEN)
 540                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 541                                      extra_lock_flags);
 542         else
 543                 LBUG();
 544
 545         if (rc < 0) {
 546                 struct lustre_handle lock_handle;
 547
 548                 if (it->d.lustre.it_lock_mode != 0) {
 549                         lock_handle.cookie = it->d.lustre.it_lock_handle;
 550                         ldlm_lock_decref(&lock_handle,
 551                                          it->d.lustre.it_lock_mode);
 552                 }
 553
 554                 it->d.lustre.it_lock_handle = 0;
 555                 it->d.lustre.it_lock_mode = 0;
 556
 557                 if (it->d.lustre.it_remote_lock_mode != 0) {
 558                         lock_handle.cookie = it->d.lustre.it_remote_lock_handle;
 559                         ldlm_lock_decref(&lock_handle,
 560                                          it->d.lustre.it_remote_lock_mode);
 561                 }
 562
 563                 it->d.lustre.it_remote_lock_handle = 0;
 564                 it->d.lustre.it_remote_lock_mode = 0;
 565         }
 566
 567         RETURN(rc);
 568 }