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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_LMV
  34 #include <linux/slab.h>
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/slab.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/math64.h>
  40 #include <linux/seq_file.h>
  41 #include <linux/namei.h>
  42 #include <lustre_intent.h>
  43
  44 #include <obd_support.h>
  45 #include <lustre_lib.h>
  46 #include <lustre_net.h>
  47 #include <lustre_dlm.h>
  48 #include <lustre_mdc.h>
  49 #include <obd_class.h>
  50 #include <lprocfs_status.h>
  51 #include "lmv_internal.h"
  52
  53 static int lmv_intent_remote(struct obd_export *exp, struct lookup_intent *it,
  54                              const struct lu_fid *parent_fid,
  55                              struct ptlrpc_request **reqp,
  56                              ldlm_blocking_callback cb_blocking,
  57                              __u64 extra_lock_flags,
  58                              const char *secctx_name, __u32 secctx_name_size)
  59 {
  60         struct obd_device       *obd = exp->exp_obd;
  61         struct lmv_obd          *lmv = &obd->u.lmv;
  62         struct ptlrpc_request   *req = NULL;
  63         struct lustre_handle    plock;
  64         struct md_op_data       *op_data;
  65         struct lmv_tgt_desc     *tgt;
  66         struct mdt_body         *body;
  67         int                     pmode;
  68         int                     rc = 0;
  69         ENTRY;
  70
  71         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
  72         if (body == NULL)
  73                 RETURN(-EPROTO);
  74
  75         LASSERT((body->mbo_valid & OBD_MD_MDS));
  76
  77         /*
  78          * We got LOOKUP lock, but we really need attrs.
  79          */
  80         pmode = it->it_lock_mode;
  81         if (pmode) {
  82                 plock.cookie = it->it_lock_handle;
  83                 it->it_lock_mode = 0;
  84                 it->it_request = NULL;
  85         }
  86
  87         LASSERT(fid_is_sane(&body->mbo_fid1));
  88
  89         tgt = lmv_fid2tgt(lmv, &body->mbo_fid1);
  90         if (IS_ERR(tgt))
  91                 GOTO(out, rc = PTR_ERR(tgt));
  92
  93         OBD_ALLOC_PTR(op_data);
  94         if (op_data == NULL)
  95                 GOTO(out, rc = -ENOMEM);
  96
  97         op_data->op_fid1 = body->mbo_fid1;
  98         /* Sent the parent FID to the remote MDT */
  99         if (parent_fid != NULL) {
 100                 /* The parent fid is only for remote open to
 101                  * check whether the open is from OBF,
 102                  * see mdt_cross_open */
 103                 LASSERT(it->it_op & IT_OPEN);
 104                 op_data->op_fid2 = *parent_fid;
 105         }
 106
 107         op_data->op_bias = MDS_CROSS_REF;
 108         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%u\n",
 109                PFID(&body->mbo_fid1), tgt->ltd_index);
 110
 111         /* ask for security context upon intent */
 112         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_OPEN) &&
 113             secctx_name_size != 0 && secctx_name != NULL) {
 114                 op_data->op_file_secctx_name = secctx_name;
 115                 op_data->op_file_secctx_name_size = secctx_name_size;
 116                 CDEBUG(D_SEC, "'%.*s' is security xattr to fetch for "
 117                        DFID"\n",
 118                        secctx_name_size, secctx_name, PFID(&body->mbo_fid1));
 119         }
 120
 121         rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
 122                             extra_lock_flags);
 123         if (rc)
 124                 GOTO(out_free_op_data, rc);
 125
 126         /*
 127          * LLite needs LOOKUP lock to track dentry revocation in order to
 128          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 129          * and put LOOKUP in request.
 130          */
 131         if (it->it_lock_mode != 0) {
 132                 it->it_remote_lock_handle =
 133                                         it->it_lock_handle;
 134                 it->it_remote_lock_mode = it->it_lock_mode;
 135         }
 136
 137         if (pmode) {
 138                 it->it_lock_handle = plock.cookie;
 139                 it->it_lock_mode = pmode;
 140         }
 141
 142         EXIT;
 143 out_free_op_data:
 144         OBD_FREE_PTR(op_data);
 145 out:
 146         if (rc && pmode)
 147                 ldlm_lock_decref(&plock, pmode);
 148
 149         ptlrpc_req_finished(*reqp);
 150         *reqp = req;
 151         return rc;
 152 }
 153
 154 int lmv_revalidate_slaves(struct obd_export *exp,
 155                           const struct lmv_stripe_md *lsm,
 156                           ldlm_blocking_callback cb_blocking,
 157                           int extra_lock_flags)
 158 {
 159         struct obd_device *obd = exp->exp_obd;
 160         struct lmv_obd *lmv = &obd->u.lmv;
 161         struct ptlrpc_request *req = NULL;
 162         struct mdt_body *body;
 163         struct md_op_data *op_data;
 164         int i;
 165         int valid_stripe_count = 0;
 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                 if (!inode)
 193                         continue;
 194
 195                 /*
 196                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 197                  * defined otherwise it will go to server and take new lock
 198                  * which is not needed here.
 199                  */
 200                 memset(op_data, 0, sizeof(*op_data));
 201                 op_data->op_fid1 = fid;
 202                 op_data->op_fid2 = fid;
 203                 /* shard revalidate only needs to fetch attributes and UPDATE
 204                  * lock, which is similar to the bottom half of remote object
 205                  * getattr, set this flag so that MDT skips checking whether
 206                  * it's remote object.
 207                  */
 208                 op_data->op_bias = MDS_CROSS_REF;
 209
 210                 tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[i].lmo_mds);
 211                 if (!tgt)
 212                         GOTO(cleanup, rc = -ENODEV);
 213
 214                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 215                        PFID(&fid), tgt->ltd_index);
 216
 217                 if (req != NULL) {
 218                         ptlrpc_req_finished(req);
 219                         req = NULL;
 220                 }
 221
 222                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 223                                     cb_blocking, extra_lock_flags);
 224                 if (rc == -ENOENT) {
 225                         /* skip stripe is not exists */
 226                         rc = 0;
 227                         continue;
 228                 }
 229
 230                 if (rc < 0)
 231                         GOTO(cleanup, rc);
 232
 233                 lockh = (struct lustre_handle *)&it.it_lock_handle;
 234                 if (rc > 0 && req == NULL) {
 235                         /* slave inode is still valid */
 236                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 237                                PFID(&fid));
 238                         rc = 0;
 239                 } else {
 240                         /* refresh slave from server */
 241                         body = req_capsule_server_get(&req->rq_pill,
 242                                                       &RMF_MDT_BODY);
 243                         if (body == NULL) {
 244                                 if (it.it_lock_mode && lockh) {
 245                                         ldlm_lock_decref(lockh,
 246                                                  it.it_lock_mode);
 247                                         it.it_lock_mode = 0;
 248                                 }
 249                                 GOTO(cleanup, rc = -ENOENT);
 250                         }
 251
 252                         i_size_write(inode, body->mbo_size);
 253                         inode->i_blocks = body->mbo_blocks;
 254                         set_nlink(inode, body->mbo_nlink);
 255                         inode->i_atime.tv_sec = body->mbo_atime;
 256                         inode->i_ctime.tv_sec = body->mbo_ctime;
 257                         inode->i_mtime.tv_sec = body->mbo_mtime;
 258                 }
 259
 260                 md_set_lock_data(tgt->ltd_exp, lockh, inode, NULL);
 261                 if (it.it_lock_mode != 0 && lockh != NULL) {
 262                         ldlm_lock_decref(lockh, it.it_lock_mode);
 263                         it.it_lock_mode = 0;
 264                 }
 265
 266                 valid_stripe_count++;
 267         }
 268
 269 cleanup:
 270         if (req != NULL)
 271                 ptlrpc_req_finished(req);
 272
 273         /* if all stripes are invalid, return -ENOENT to notify user */
 274         if (!rc && !valid_stripe_count)
 275                 rc = -ENOENT;
 276
 277         OBD_FREE_PTR(op_data);
 278         RETURN(rc);
 279 }
 280
 281 /*
 282  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 283  * may be split dir.
 284  */
 285 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 286                            struct lookup_intent *it,
 287                            struct ptlrpc_request **reqp,
 288                            ldlm_blocking_callback cb_blocking,
 289                            __u64 extra_lock_flags)
 290 {
 291         struct obd_device *obd = exp->exp_obd;
 292         struct lmv_obd *lmv = &obd->u.lmv;
 293         struct lmv_tgt_desc *tgt;
 294         struct mdt_body *body;
 295         __u64 flags = it->it_flags;
 296         int rc;
 297
 298         ENTRY;
 299
 300         /* do not allow file creation in foreign dir */
 301         if ((it->it_op & IT_CREAT) && lmv_dir_foreign(op_data->op_mea1))
 302                 RETURN(-ENODATA);
 303
 304         if ((it->it_op & IT_CREAT) && !(flags & MDS_OPEN_BY_FID)) {
 305                 /* don't allow create under dir with bad hash */
 306                 if (lmv_dir_bad_hash(op_data->op_mea1))
 307                         RETURN(-EBADF);
 308
 309                 if (lmv_dir_layout_changing(op_data->op_mea1)) {
 310                         if (flags & O_EXCL) {
 311                                 /*
 312                                  * open(O_CREAT | O_EXCL) needs to check
 313                                  * existing name, which should be done on both
 314                                  * old and new layout, check old layout on
 315                                  * client side.
 316                                  */
 317                                 rc = lmv_old_layout_lookup(lmv, op_data);
 318                                 if (rc != -ENOENT)
 319                                         RETURN(rc);
 320
 321                                 op_data->op_new_layout = true;
 322                         } else {
 323                                 /*
 324                                  * open(O_CREAT) will be sent to MDT in old
 325                                  * layout first, to avoid creating new file
 326                                  * under old layout, clear O_CREAT.
 327                                  */
 328                                 it->it_flags &= ~O_CREAT;
 329                         }
 330                 }
 331         }
 332
 333 retry:
 334         if (it->it_flags & MDS_OPEN_BY_FID) {
 335                 LASSERT(fid_is_sane(&op_data->op_fid2));
 336
 337                 /* for striped directory, we can't know parent stripe fid
 338                  * without name, but we can set it to child fid, and MDT
 339                  * will obtain it from linkea in open in such case. */
 340                 if (lmv_dir_striped(op_data->op_mea1))
 341                         op_data->op_fid1 = op_data->op_fid2;
 342
 343                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
 344                 if (IS_ERR(tgt))
 345                         RETURN(PTR_ERR(tgt));
 346
 347                 op_data->op_mds = tgt->ltd_index;
 348         } else {
 349                 LASSERT(fid_is_sane(&op_data->op_fid1));
 350                 LASSERT(it->it_flags & MDS_OPEN_PCC ||
 351                         fid_is_zero(&op_data->op_fid2));
 352                 LASSERT(op_data->op_name != NULL);
 353
 354                 tgt = lmv_locate_tgt(lmv, op_data);
 355                 if (IS_ERR(tgt))
 356                         RETURN(PTR_ERR(tgt));
 357         }
 358
 359         /* If it is ready to open the file by FID, do not need
 360          * allocate FID at all, otherwise it will confuse MDT */
 361         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID ||
 362                                         it->it_flags & MDS_OPEN_PCC)) {
 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 #%u\n", PFID(&op_data->op_fid1),
 374                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_index);
 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->it_disposition & DISP_LOOKUP_NEG) &&
 385             !(it->it_disposition & DISP_OPEN_CREATE) &&
 386             !(it->it_disposition & DISP_OPEN_OPEN)) {
 387                 if (!(it->it_flags & MDS_OPEN_BY_FID) &&
 388                     lmv_dir_retry_check_update(op_data)) {
 389                         ptlrpc_req_finished(*reqp);
 390                         it->it_request = NULL;
 391                         it->it_disposition = 0;
 392                         *reqp = NULL;
 393
 394                         it->it_flags = flags;
 395                         fid_zero(&op_data->op_fid2);
 396                         goto retry;
 397                 }
 398
 399                 RETURN(rc);
 400         }
 401
 402         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 403         if (body == NULL)
 404                 RETURN(-EPROTO);
 405
 406         /* Not cross-ref case, just get out of here. */
 407         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 408                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 409                                        cb_blocking, extra_lock_flags,
 410                                        op_data->op_file_secctx_name,
 411                                        op_data->op_file_secctx_name_size);
 412                 if (rc != 0)
 413                         RETURN(rc);
 414
 415                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 416                 if (body == NULL)
 417                         RETURN(-EPROTO);
 418         }
 419
 420         RETURN(rc);
 421 }
 422
 423 /*
 424  * Handler for: getattr, lookup and revalidate cases.
 425  */
 426 static int
 427 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 428                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 429                   ldlm_blocking_callback cb_blocking,
 430                   __u64 extra_lock_flags)
 431 {
 432         struct obd_device *obd = exp->exp_obd;
 433         struct lmv_obd *lmv = &obd->u.lmv;
 434         struct lmv_tgt_desc *tgt = NULL;
 435         struct mdt_body *body;
 436         int rc;
 437         ENTRY;
 438
 439         /* foreign dir is not striped */
 440         if (lmv_dir_foreign(op_data->op_mea1)) {
 441                 /* only allow getattr/lookup for itself */
 442                 if (op_data->op_name != NULL)
 443                         RETURN(-ENODATA);
 444                 RETURN(0);
 445         }
 446
 447 retry:
 448         if (op_data->op_flags & MF_GETATTR_BY_FID) {
 449                 /* getattr by FID, replace fid1 with stripe FID,
 450                  * NB, don't replace if name is "/", because it may be a subtree
 451                  * mount, and if it's a striped directory, fid1 will be replaced
 452                  * to stripe FID by hash, while fid2 is master object FID, which
 453                  * will be treated as a remote object if the two FIDs are
 454                  * located on different MDTs, and LOOKUP lock can't be fetched.
 455                  */
 456                 LASSERT(op_data->op_name);
 457                 if (op_data->op_namelen != 1 ||
 458                     strncmp(op_data->op_name, "/", 1) != 0) {
 459                         tgt = lmv_locate_tgt(lmv, op_data);
 460                         if (IS_ERR(tgt))
 461                                 RETURN(PTR_ERR(tgt));
 462                 }
 463
 464                 /* name is used to locate stripe target, clear it here
 465                  * to avoid packing name in request, so that MDS knows
 466                  * it's getattr by FID.
 467                  */
 468                 op_data->op_name = NULL;
 469                 op_data->op_namelen = 0;
 470
 471                 /* getattr request is sent to MDT where fid2 inode is */
 472                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
 473         } else if (op_data->op_name) {
 474                 /* getattr by name */
 475                 tgt = lmv_locate_tgt(lmv, op_data);
 476                 if (!fid_is_sane(&op_data->op_fid2))
 477                         fid_zero(&op_data->op_fid2);
 478         } else {
 479                 /* old way to getattr by FID, parent FID not packed */
 480                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 481         }
 482         if (IS_ERR(tgt))
 483                 RETURN(PTR_ERR(tgt));
 484
 485         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 486                ", name='%s' -> mds #%u\n",
 487                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 488                op_data->op_name ? op_data->op_name : "<NULL>",
 489                tgt->ltd_index);
 490
 491         op_data->op_bias &= ~MDS_CROSS_REF;
 492
 493         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 494                             extra_lock_flags);
 495         if (rc < 0)
 496                 RETURN(rc);
 497
 498         if (*reqp == NULL) {
 499                 /* If RPC happens, lsm information will be revalidated
 500                  * during update_inode process (see ll_update_lsm_md) */
 501                 if (lmv_dir_striped(op_data->op_mea2)) {
 502                         rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
 503                                                    cb_blocking,
 504                                                    extra_lock_flags);
 505                         if (rc != 0)
 506                                 RETURN(rc);
 507                 }
 508                 RETURN(rc);
 509         } else if (it_disposition(it, DISP_LOOKUP_NEG) &&
 510                    lmv_dir_retry_check_update(op_data)) {
 511                 ptlrpc_req_finished(*reqp);
 512                 it->it_request = NULL;
 513                 it->it_disposition = 0;
 514                 *reqp = NULL;
 515
 516                 goto retry;
 517         }
 518
 519         if (!it_has_reply_body(it))
 520                 RETURN(0);
 521
 522         /*
 523          * MDS has returned success. Probably name has been resolved in
 524          * remote inode. Let's check this.
 525          */
 526         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 527         if (body == NULL)
 528                 RETURN(-EPROTO);
 529
 530         /* Not cross-ref case, just get out of here. */
 531         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 532                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 533                                        extra_lock_flags,
 534                                        op_data->op_file_secctx_name,
 535                                        op_data->op_file_secctx_name_size);
 536                 if (rc != 0)
 537                         RETURN(rc);
 538                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 539                 if (body == NULL)
 540                         RETURN(-EPROTO);
 541         }
 542
 543         RETURN(rc);
 544 }
 545
 546 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 547                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 548                     ldlm_blocking_callback cb_blocking,
 549                     __u64 extra_lock_flags)
 550 {
 551         int rc;
 552         ENTRY;
 553
 554         LASSERT(it != NULL);
 555         LASSERT(fid_is_sane(&op_data->op_fid1));
 556
 557         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 558                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 559                 (int)op_data->op_namelen, op_data->op_name,
 560                 PFID(&op_data->op_fid1));
 561
 562         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT | IT_GETXATTR))
 563                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 564                                        extra_lock_flags);
 565         else if (it->it_op & IT_OPEN)
 566                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 567                                      extra_lock_flags);
 568         else
 569                 LBUG();
 570
 571         if (rc < 0) {
 572                 struct lustre_handle lock_handle;
 573
 574                 if (it->it_lock_mode != 0) {
 575                         lock_handle.cookie = it->it_lock_handle;
 576                         ldlm_lock_decref_and_cancel(&lock_handle,
 577                                                     it->it_lock_mode);
 578                 }
 579
 580                 it->it_lock_handle = 0;
 581                 it->it_lock_mode = 0;
 582
 583                 if (it->it_remote_lock_mode != 0) {
 584                         lock_handle.cookie = it->it_remote_lock_handle;
 585                         ldlm_lock_decref_and_cancel(&lock_handle,
 586                                                     it->it_remote_lock_mode);
 587                 }
 588
 589                 it->it_remote_lock_handle = 0;
 590                 it->it_remote_lock_mode = 0;
 591         }
 592
 593         RETURN(rc);
 594 }