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, 2015, 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,
 156                           const 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         int                     i;
 166         int                     rc = 0;
 167
 168         ENTRY;
 169
 170         /**
 171          * revalidate slaves has some problems, temporarily return,
 172          * we may not need that
 173          */
 174         OBD_ALLOC_PTR(op_data);
 175         if (op_data == NULL)
 176                 RETURN(-ENOMEM);
 177
 178         /**
 179          * Loop over the stripe information, check validity and update them
 180          * from MDS if needed.
 181          */
 182         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 183                 struct lu_fid           fid;
 184                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 185                 struct lustre_handle    *lockh = NULL;
 186                 struct lmv_tgt_desc     *tgt = NULL;
 187                 struct inode            *inode;
 188
 189                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 190                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 191
 192                 /*
 193                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 194                  * defined otherwise it will go to server and take new lock
 195                  * which is not needed here.
 196                  */
 197                 memset(op_data, 0, sizeof(*op_data));
 198                 op_data->op_fid1 = fid;
 199                 op_data->op_fid2 = fid;
 200
 201                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 202                 if (IS_ERR(tgt))
 203                         GOTO(cleanup, rc = PTR_ERR(tgt));
 204
 205                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 206                        PFID(&fid), tgt->ltd_idx);
 207
 208                 if (req != NULL) {
 209                         ptlrpc_req_finished(req);
 210                         req = NULL;
 211                 }
 212
 213                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 214                                     cb_blocking, extra_lock_flags);
 215                 if (rc < 0)
 216                         GOTO(cleanup, rc);
 217
 218                 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
 219                 if (rc > 0 && req == NULL) {
 220                         /* slave inode is still valid */
 221                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 222                                PFID(&fid));
 223                         rc = 0;
 224                 } else {
 225                         /* refresh slave from server */
 226                         body = req_capsule_server_get(&req->rq_pill,
 227                                                       &RMF_MDT_BODY);
 228                         if (body == NULL) {
 229                                 if (it.d.lustre.it_lock_mode && lockh) {
 230                                         ldlm_lock_decref(lockh,
 231                                                  it.d.lustre.it_lock_mode);
 232                                         it.d.lustre.it_lock_mode = 0;
 233                                 }
 234                                 GOTO(cleanup, rc = -ENOENT);
 235                         }
 236
 237                         i_size_write(inode, body->mbo_size);
 238                         inode->i_blocks = body->mbo_blocks;
 239                         set_nlink(inode, body->mbo_nlink);
 240                         LTIME_S(inode->i_atime) = body->mbo_atime;
 241                         LTIME_S(inode->i_ctime) = body->mbo_ctime;
 242                         LTIME_S(inode->i_mtime) = body->mbo_mtime;
 243                 }
 244
 245                 md_set_lock_data(tgt->ltd_exp, &lockh->cookie, inode, NULL);
 246                 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
 247                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 248                         it.d.lustre.it_lock_mode = 0;
 249                 }
 250         }
 251
 252 cleanup:
 253         if (req != NULL)
 254                 ptlrpc_req_finished(req);
 255
 256         OBD_FREE_PTR(op_data);
 257         RETURN(rc);
 258 }
 259
 260
 261 /*
 262  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 263  * may be split dir.
 264  */
 265 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 266                            struct lookup_intent *it,
 267                            struct ptlrpc_request **reqp,
 268                            ldlm_blocking_callback cb_blocking,
 269                            __u64 extra_lock_flags)
 270 {
 271         struct obd_device       *obd = exp->exp_obd;
 272         struct lmv_obd          *lmv = &obd->u.lmv;
 273         struct lmv_tgt_desc     *tgt;
 274         struct mdt_body         *body;
 275         int                     rc;
 276         ENTRY;
 277
 278         if (it->it_flags & MDS_OPEN_BY_FID) {
 279                 LASSERT(fid_is_sane(&op_data->op_fid2));
 280
 281                 /* for striped directory, we can't know parent stripe fid
 282                  * without name, but we can set it to child fid, and MDT
 283                  * will obtain it from linkea in open in such case. */
 284                 if (op_data->op_mea1 != NULL)
 285                         op_data->op_fid1 = op_data->op_fid2;
 286
 287                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 288                 if (IS_ERR(tgt))
 289                         RETURN(PTR_ERR(tgt));
 290
 291                 op_data->op_mds = tgt->ltd_idx;
 292         } else {
 293                 LASSERT(fid_is_sane(&op_data->op_fid1));
 294                 LASSERT(fid_is_zero(&op_data->op_fid2));
 295                 LASSERT(op_data->op_name != NULL);
 296
 297                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 298                 if (IS_ERR(tgt))
 299                         RETURN(PTR_ERR(tgt));
 300         }
 301
 302         /* If it is ready to open the file by FID, do not need
 303          * allocate FID at all, otherwise it will confuse MDT */
 304         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 305                 /*
 306                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 307                  * for it.
 308                  */
 309                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 310                 if (rc != 0)
 311                         RETURN(rc);
 312         }
 313
 314         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 315                " name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
 316                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 317
 318         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 319                             extra_lock_flags);
 320         if (rc != 0)
 321                 RETURN(rc);
 322         /*
 323          * Nothing is found, do not access body->fid1 as it is zero and thus
 324          * pointless.
 325          */
 326         if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
 327             !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
 328             !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
 329                 RETURN(rc);
 330
 331         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 332         if (body == NULL)
 333                 RETURN(-EPROTO);
 334
 335         /* Not cross-ref case, just get out of here. */
 336         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 337                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 338                                        cb_blocking, extra_lock_flags);
 339                 if (rc != 0)
 340                         RETURN(rc);
 341
 342                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 343                 if (body == NULL)
 344                         RETURN(-EPROTO);
 345         }
 346
 347         RETURN(rc);
 348 }
 349
 350 /*
 351  * Handler for: getattr, lookup and revalidate cases.
 352  */
 353 static int
 354 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 355                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 356                   ldlm_blocking_callback cb_blocking,
 357                   __u64 extra_lock_flags)
 358 {
 359         struct obd_device       *obd = exp->exp_obd;
 360         struct lmv_obd          *lmv = &obd->u.lmv;
 361         struct lmv_tgt_desc     *tgt = NULL;
 362         struct mdt_body         *body;
 363         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 364         int                     rc = 0;
 365         ENTRY;
 366
 367         /* If it returns ERR_PTR(-EBADFD) then it is an unknown hash type
 368          * it will try all stripes to locate the object */
 369         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 370         if (IS_ERR(tgt) && (PTR_ERR(tgt) != -EBADFD))
 371                 RETURN(PTR_ERR(tgt));
 372
 373         /* Both migrating dir and unknown hash dir need to try
 374          * all of sub-stripes */
 375         if (lsm != NULL && !lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
 376                 struct lmv_oinfo *oinfo;
 377
 378                 oinfo = &lsm->lsm_md_oinfo[0];
 379
 380                 op_data->op_fid1 = oinfo->lmo_fid;
 381                 op_data->op_mds = oinfo->lmo_mds;
 382                 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
 383                 if (IS_ERR(tgt))
 384                         RETURN(PTR_ERR(tgt));
 385         }
 386
 387         if (!fid_is_sane(&op_data->op_fid2))
 388                 fid_zero(&op_data->op_fid2);
 389
 390         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 391                ", name='%s' -> mds #%u lsm=%p lsm_magic=%x\n",
 392                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 393                op_data->op_name ? op_data->op_name : "<NULL>",
 394                tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
 395
 396         op_data->op_bias &= ~MDS_CROSS_REF;
 397
 398         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 399                             extra_lock_flags);
 400         if (rc < 0)
 401                 RETURN(rc);
 402
 403         if (*reqp == NULL) {
 404                 /* If RPC happens, lsm information will be revalidated
 405                  * during update_inode process (see ll_update_lsm_md) */
 406                 if (op_data->op_mea2 != NULL) {
 407                         rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
 408                                                    cb_blocking,
 409                                                    extra_lock_flags);
 410                         if (rc != 0)
 411                                 RETURN(rc);
 412                 }
 413                 RETURN(rc);
 414         } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
 415                    lmv_need_try_all_stripes(lsm)) {
 416                 /* For migrating and unknown hash type directory, it will
 417                  * try to target the entry on other stripes */
 418                 int stripe_index;
 419
 420                 for (stripe_index = 1;
 421                      stripe_index < lsm->lsm_md_stripe_count &&
 422                      it_disposition(it, DISP_LOOKUP_NEG); stripe_index++) {
 423                         struct lmv_oinfo *oinfo;
 424
 425                         /* release the previous request */
 426                         ptlrpc_req_finished(*reqp);
 427                         it->d.lustre.it_data = NULL;
 428                         *reqp = NULL;
 429
 430                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
 431                         tgt = lmv_find_target(lmv, &oinfo->lmo_fid);
 432                         if (IS_ERR(tgt))
 433                                 RETURN(PTR_ERR(tgt));
 434
 435                         CDEBUG(D_INODE, "Try other stripes " DFID"\n",
 436                                PFID(&oinfo->lmo_fid));
 437
 438                         op_data->op_fid1 = oinfo->lmo_fid;
 439                         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
 440                         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
 441                                             cb_blocking, extra_lock_flags);
 442                         if (rc != 0)
 443                                 RETURN(rc);
 444                 }
 445         }
 446
 447         /*
 448          * MDS has returned success. Probably name has been resolved in
 449          * remote inode. Let's check this.
 450          */
 451         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 452         if (body == NULL)
 453                 RETURN(-EPROTO);
 454
 455         /* Not cross-ref case, just get out of here. */
 456         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 457                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 458                                        extra_lock_flags);
 459                 if (rc != 0)
 460                         RETURN(rc);
 461                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 462                 if (body == NULL)
 463                         RETURN(-EPROTO);
 464         }
 465
 466         RETURN(rc);
 467 }
 468
 469 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 470                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 471                     ldlm_blocking_callback cb_blocking,
 472                     __u64 extra_lock_flags)
 473 {
 474         struct obd_device *obd = exp->exp_obd;
 475         int                rc;
 476         ENTRY;
 477
 478         LASSERT(it != NULL);
 479         LASSERT(fid_is_sane(&op_data->op_fid1));
 480
 481         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 482                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 483                 (int)op_data->op_namelen, op_data->op_name,
 484                 PFID(&op_data->op_fid1));
 485
 486         rc = lmv_check_connect(obd);
 487         if (rc)
 488                 RETURN(rc);
 489
 490         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 491                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 492                                        extra_lock_flags);
 493         else if (it->it_op & IT_OPEN)
 494                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 495                                      extra_lock_flags);
 496         else
 497                 LBUG();
 498
 499         if (rc < 0) {
 500                 struct lustre_handle lock_handle;
 501
 502                 if (it->d.lustre.it_lock_mode != 0) {
 503                         lock_handle.cookie = it->d.lustre.it_lock_handle;
 504                         ldlm_lock_decref(&lock_handle,
 505                                          it->d.lustre.it_lock_mode);
 506                 }
 507
 508                 it->d.lustre.it_lock_handle = 0;
 509                 it->d.lustre.it_lock_mode = 0;
 510
 511                 if (it->d.lustre.it_remote_lock_mode != 0) {
 512                         lock_handle.cookie = it->d.lustre.it_remote_lock_handle;
 513                         ldlm_lock_decref(&lock_handle,
 514                                          it->d.lustre.it_remote_lock_mode);
 515                 }
 516
 517                 it->d.lustre.it_remote_lock_handle = 0;
 518                 it->d.lustre.it_remote_lock_mode = 0;
 519         }
 520
 521         RETURN(rc);
 522 }