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->mbo_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->mbo_fid1));
 103
 104         tgt = lmv_find_target(lmv, &body->mbo_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->mbo_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         }
 121
 122         op_data->op_bias = MDS_CROSS_REF;
 123         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%d\n",
 124                PFID(&body->mbo_fid1), tgt->ltd_idx);
 125
 126         rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
 127                             extra_lock_flags);
 128         if (rc)
 129                 GOTO(out_free_op_data, rc);
 130
 131         /*
 132          * LLite needs LOOKUP lock to track dentry revocation in order to
 133          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 134          * and put LOOKUP in request.
 135          */
 136         if (it->d.lustre.it_lock_mode != 0) {
 137                 it->d.lustre.it_remote_lock_handle =
 138                                         it->d.lustre.it_lock_handle;
 139                 it->d.lustre.it_remote_lock_mode = it->d.lustre.it_lock_mode;
 140         }
 141
 142         if (pmode) {
 143                 it->d.lustre.it_lock_handle = plock.cookie;
 144                 it->d.lustre.it_lock_mode = pmode;
 145         }
 146
 147         EXIT;
 148 out_free_op_data:
 149         OBD_FREE_PTR(op_data);
 150 out:
 151         if (rc && pmode)
 152                 ldlm_lock_decref(&plock, pmode);
 153
 154         ptlrpc_req_finished(*reqp);
 155         *reqp = req;
 156         return rc;
 157 }
 158
 159 #ifdef __KERNEL__
 160 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 161                           struct lmv_stripe_md *lsm,
 162                           ldlm_blocking_callback cb_blocking,
 163                           int extra_lock_flags)
 164 {
 165         struct obd_device      *obd = exp->exp_obd;
 166         struct lmv_obd         *lmv = &obd->u.lmv;
 167         struct mdt_body         *body;
 168         struct md_op_data      *op_data;
 169         unsigned long           size = 0;
 170         unsigned long           nlink = 0;
 171         obd_time                atime = 0;
 172         obd_time                ctime = 0;
 173         obd_time                mtime = 0;
 174         int                     i;
 175         int                     rc = 0;
 176
 177         ENTRY;
 178
 179         /**
 180          * revalidate slaves has some problems, temporarily return,
 181          * we may not need that
 182          */
 183         OBD_ALLOC_PTR(op_data);
 184         if (op_data == NULL)
 185                 RETURN(-ENOMEM);
 186
 187         /**
 188          * Loop over the stripe information, check validity and update them
 189          * from MDS if needed.
 190          */
 191         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 192                 struct lu_fid           fid;
 193                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 194                 struct ptlrpc_request   *req = NULL;
 195                 struct lustre_handle    *lockh = NULL;
 196                 struct lmv_tgt_desc     *tgt = NULL;
 197                 struct inode            *inode;
 198
 199                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 200                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 201
 202                 /*
 203                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 204                  * defined otherwise it will go to server and take new lock
 205                  * which is not needed here.
 206                  */
 207                 memset(op_data, 0, sizeof(*op_data));
 208                 op_data->op_fid1 = fid;
 209                 op_data->op_fid2 = fid;
 210
 211                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 212                 if (IS_ERR(tgt))
 213                         GOTO(cleanup, rc = PTR_ERR(tgt));
 214
 215                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%d\n",
 216                        PFID(&fid), tgt->ltd_idx);
 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 (req != NULL)
 242                                         ptlrpc_req_finished(req);
 243
 244                                 if (it.d.lustre.it_lock_mode && lockh) {
 245                                         ldlm_lock_decref(lockh,
 246                                                  it.d.lustre.it_lock_mode);
 247                                         it.d.lustre.it_lock_mode = 0;
 248                                 }
 249
 250                                 GOTO(cleanup, rc = -EIO);
 251                         }
 252
 253
 254                         i_size_write(inode, body->mbo_size);
 255                         set_nlink(inode, body->mbo_nlink);
 256                         LTIME_S(inode->i_atime) = body->mbo_atime;
 257                         LTIME_S(inode->i_ctime) = body->mbo_ctime;
 258                         LTIME_S(inode->i_mtime) = body->mbo_mtime;
 259
 260                         if (req != NULL)
 261                                 ptlrpc_req_finished(req);
 262                 }
 263
 264                 md_set_lock_data(tgt->ltd_exp, &lockh->cookie, inode, NULL);
 265
 266                 size += i_size_read(inode);
 267
 268                 if (i != 0)
 269                         nlink += inode->i_nlink - 2;
 270                 else
 271                         nlink += inode->i_nlink;
 272
 273                 atime = LTIME_S(inode->i_atime) > atime ?
 274                                 LTIME_S(inode->i_atime) : atime;
 275                 ctime = LTIME_S(inode->i_ctime) > ctime ?
 276                                 LTIME_S(inode->i_ctime) : ctime;
 277                 mtime = LTIME_S(inode->i_mtime) > mtime ?
 278                                 LTIME_S(inode->i_mtime) : mtime;
 279
 280                 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
 281                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 282                         it.d.lustre.it_lock_mode = 0;
 283                 }
 284
 285                 CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
 286                        "%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
 287                        i_size_read(inode), inode->i_nlink,
 288                        LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
 289                        LTIME_S(inode->i_ctime));
 290         }
 291
 292         /*
 293          * update attr of master request.
 294          */
 295         CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
 296                LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
 297                atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 298
 299         if (mbody != NULL) {
 300                 mbody->mbo_atime = atime;
 301                 mbody->mbo_ctime = ctime;
 302                 mbody->mbo_mtime = mtime;
 303         }
 304 cleanup:
 305         OBD_FREE_PTR(op_data);
 306         RETURN(rc);
 307 }
 308
 309 #else
 310
 311 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
 312                           struct lmv_stripe_md *lsm,
 313                           ldlm_blocking_callback cb_blocking,
 314                           int extra_lock_flags)
 315 {
 316         return 0;
 317 }
 318
 319 #endif
 320
 321 /*
 322  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 323  * may be split dir.
 324  */
 325 int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 326                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 327                     ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
 328 {
 329         struct obd_device       *obd = exp->exp_obd;
 330         struct lmv_obd          *lmv = &obd->u.lmv;
 331         struct lmv_tgt_desc     *tgt;
 332         struct mdt_body         *body;
 333         int                     rc;
 334         ENTRY;
 335
 336         if (it->it_flags & MDS_OPEN_BY_FID) {
 337                 LASSERT(fid_is_sane(&op_data->op_fid2));
 338
 339                 /* for striped directory, we can't know parent stripe fid
 340                  * without name, but we can set it to child fid, and MDT
 341                  * will obtain it from linkea in open in such case. */
 342                 if (op_data->op_mea1 != NULL)
 343                         op_data->op_fid1 = op_data->op_fid2;
 344
 345                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 346                 if (IS_ERR(tgt))
 347                         RETURN(PTR_ERR(tgt));
 348
 349                 op_data->op_mds = tgt->ltd_idx;
 350         } else {
 351                 LASSERT(fid_is_sane(&op_data->op_fid1));
 352                 LASSERT(fid_is_zero(&op_data->op_fid2));
 353                 LASSERT(op_data->op_name != NULL);
 354
 355                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 356                 if (IS_ERR(tgt))
 357                         RETURN(PTR_ERR(tgt));
 358         }
 359
 360         /* If it is ready to open the file by FID, do not need
 361          * allocate FID at all, otherwise it will confuse MDT */
 362         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 363                 /*
 364                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 365                  * for it.
 366                  */
 367                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 368                 if (rc != 0)
 369                         RETURN(rc);
 370         }
 371
 372         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 373                " name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
 374                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 375
 376         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 377                             extra_lock_flags);
 378         if (rc != 0)
 379                 RETURN(rc);
 380         /*
 381          * Nothing is found, do not access body->fid1 as it is zero and thus
 382          * pointless.
 383          */
 384         if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
 385             !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
 386             !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
 387                 RETURN(rc);
 388
 389         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 390         if (body == NULL)
 391                 RETURN(-EPROTO);
 392
 393         /* Not cross-ref case, just get out of here. */
 394         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 395                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 396                                        cb_blocking, extra_lock_flags);
 397                 if (rc != 0)
 398                         RETURN(rc);
 399
 400                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 401                 if (body == NULL)
 402                         RETURN(-EPROTO);
 403         }
 404
 405         RETURN(rc);
 406 }
 407
 408 /*
 409  * Handler for: getattr, lookup and revalidate cases.
 410  */
 411 static int
 412 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 413                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 414                   ldlm_blocking_callback cb_blocking,
 415                   __u64 extra_lock_flags)
 416 {
 417         struct obd_device       *obd = exp->exp_obd;
 418         struct lmv_obd          *lmv = &obd->u.lmv;
 419         struct lmv_tgt_desc     *tgt = NULL;
 420         struct mdt_body         *body;
 421         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 422         int                     rc = 0;
 423         ENTRY;
 424
 425         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 426         if (IS_ERR(tgt))
 427                 RETURN(PTR_ERR(tgt));
 428
 429         if (!fid_is_sane(&op_data->op_fid2))
 430                 fid_zero(&op_data->op_fid2);
 431
 432         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 433                ", name='%s' -> mds #%d lsm=%p lsm_magic=%x\n",
 434                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 435                op_data->op_name ? op_data->op_name : "<NULL>",
 436                tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
 437
 438         op_data->op_bias &= ~MDS_CROSS_REF;
 439
 440         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 441                             extra_lock_flags);
 442         if (rc < 0)
 443                 RETURN(rc);
 444
 445         if (*reqp == NULL) {
 446                 /* If RPC happens, lsm information will be revalidated
 447                  * during update_inode process (see ll_update_lsm_md) */
 448                 if (op_data->op_mea2 != NULL) {
 449                         rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
 450                                                    cb_blocking,
 451                                                    extra_lock_flags);
 452                         if (rc != 0)
 453                                 RETURN(rc);
 454                 }
 455                 RETURN(rc);
 456         } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
 457                    lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION) {
 458                 /* For migrating directory, if it can not find the child in
 459                  * the source directory(master stripe), try the targeting
 460                  * directory(stripe 1) */
 461                 tgt = lmv_find_target(lmv, &lsm->lsm_md_oinfo[1].lmo_fid);
 462                 if (IS_ERR(tgt))
 463                         RETURN(PTR_ERR(tgt));
 464
 465                 ptlrpc_req_finished(*reqp);
 466                 it->d.lustre.it_data = NULL;
 467                 *reqp = NULL;
 468
 469                 CDEBUG(D_INODE, "For migrating dir, try target dir "DFID"\n",
 470                        PFID(&lsm->lsm_md_oinfo[1].lmo_fid));
 471
 472                 op_data->op_fid1 = lsm->lsm_md_oinfo[1].lmo_fid;
 473                 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
 474                 rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
 475                                     cb_blocking, extra_lock_flags);
 476         }
 477         /*
 478          * MDS has returned success. Probably name has been resolved in
 479          * remote inode. Let's check this.
 480          */
 481         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 482         if (body == NULL)
 483                 RETURN(-EPROTO);
 484
 485         /* Not cross-ref case, just get out of here. */
 486         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 487                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 488                                        extra_lock_flags);
 489                 if (rc != 0)
 490                         RETURN(rc);
 491                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 492                 if (body == NULL)
 493                         RETURN(-EPROTO);
 494         }
 495
 496         RETURN(rc);
 497 }
 498
 499 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 500                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 501                     ldlm_blocking_callback cb_blocking,
 502                     __u64 extra_lock_flags)
 503 {
 504         struct obd_device *obd = exp->exp_obd;
 505         int                rc;
 506         ENTRY;
 507
 508         LASSERT(it != NULL);
 509         LASSERT(fid_is_sane(&op_data->op_fid1));
 510
 511         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 512                 LL_IT2STR(it), PFID(&op_data->op_fid2), op_data->op_namelen,
 513                 op_data->op_name, PFID(&op_data->op_fid1));
 514
 515         rc = lmv_check_connect(obd);
 516         if (rc)
 517                 RETURN(rc);
 518
 519         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 520                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 521                                        extra_lock_flags);
 522         else if (it->it_op & IT_OPEN)
 523                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 524                                      extra_lock_flags);
 525         else
 526                 LBUG();
 527
 528         RETURN(rc);
 529 }