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 {
  59         struct obd_device       *obd = exp->exp_obd;
  60         struct lmv_obd          *lmv = &obd->u.lmv;
  61         struct ptlrpc_request   *req = NULL;
  62         struct lustre_handle    plock;
  63         struct md_op_data       *op_data;
  64         struct lmv_tgt_desc     *tgt;
  65         struct mdt_body         *body;
  66         int                     pmode;
  67         int                     rc = 0;
  68         ENTRY;
  69
  70         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
  71         if (body == NULL)
  72                 RETURN(-EPROTO);
  73
  74         LASSERT((body->mbo_valid & OBD_MD_MDS));
  75
  76         /*
  77          * We got LOOKUP lock, but we really need attrs.
  78          */
  79         pmode = it->it_lock_mode;
  80         if (pmode) {
  81                 plock.cookie = it->it_lock_handle;
  82                 it->it_lock_mode = 0;
  83                 it->it_request = NULL;
  84         }
  85
  86         LASSERT(fid_is_sane(&body->mbo_fid1));
  87
  88         tgt = lmv_find_target(lmv, &body->mbo_fid1);
  89         if (IS_ERR(tgt))
  90                 GOTO(out, rc = PTR_ERR(tgt));
  91
  92         OBD_ALLOC_PTR(op_data);
  93         if (op_data == NULL)
  94                 GOTO(out, rc = -ENOMEM);
  95
  96         op_data->op_fid1 = body->mbo_fid1;
  97         /* Sent the parent FID to the remote MDT */
  98         if (parent_fid != NULL) {
  99                 /* The parent fid is only for remote open to
 100                  * check whether the open is from OBF,
 101                  * see mdt_cross_open */
 102                 LASSERT(it->it_op & IT_OPEN);
 103                 op_data->op_fid2 = *parent_fid;
 104         }
 105
 106         op_data->op_bias = MDS_CROSS_REF;
 107         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%u\n",
 108                PFID(&body->mbo_fid1), tgt->ltd_idx);
 109
 110         rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
 111                             extra_lock_flags);
 112         if (rc)
 113                 GOTO(out_free_op_data, rc);
 114
 115         /*
 116          * LLite needs LOOKUP lock to track dentry revocation in order to
 117          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 118          * and put LOOKUP in request.
 119          */
 120         if (it->it_lock_mode != 0) {
 121                 it->it_remote_lock_handle =
 122                                         it->it_lock_handle;
 123                 it->it_remote_lock_mode = it->it_lock_mode;
 124         }
 125
 126         if (pmode) {
 127                 it->it_lock_handle = plock.cookie;
 128                 it->it_lock_mode = pmode;
 129         }
 130
 131         EXIT;
 132 out_free_op_data:
 133         OBD_FREE_PTR(op_data);
 134 out:
 135         if (rc && pmode)
 136                 ldlm_lock_decref(&plock, pmode);
 137
 138         ptlrpc_req_finished(*reqp);
 139         *reqp = req;
 140         return rc;
 141 }
 142
 143 int lmv_revalidate_slaves(struct obd_export *exp,
 144                           const struct lmv_stripe_md *lsm,
 145                           ldlm_blocking_callback cb_blocking,
 146                           int extra_lock_flags)
 147 {
 148         struct obd_device *obd = exp->exp_obd;
 149         struct lmv_obd *lmv = &obd->u.lmv;
 150         struct ptlrpc_request *req = NULL;
 151         struct mdt_body *body;
 152         struct md_op_data *op_data;
 153         int i;
 154         int valid_stripe_count = 0;
 155         int rc = 0;
 156
 157         ENTRY;
 158
 159         /**
 160          * revalidate slaves has some problems, temporarily return,
 161          * we may not need that
 162          */
 163         OBD_ALLOC_PTR(op_data);
 164         if (op_data == NULL)
 165                 RETURN(-ENOMEM);
 166
 167         /**
 168          * Loop over the stripe information, check validity and update them
 169          * from MDS if needed.
 170          */
 171         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 172                 struct lu_fid           fid;
 173                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 174                 struct lustre_handle    *lockh = NULL;
 175                 struct lmv_tgt_desc     *tgt = NULL;
 176                 struct inode            *inode;
 177
 178                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 179                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 180
 181                 if (!inode)
 182                         continue;
 183
 184                 /*
 185                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 186                  * defined otherwise it will go to server and take new lock
 187                  * which is not needed here.
 188                  */
 189                 memset(op_data, 0, sizeof(*op_data));
 190                 op_data->op_fid1 = fid;
 191                 op_data->op_fid2 = fid;
 192                 /* shard revalidate only needs to fetch attributes and UPDATE
 193                  * lock, which is similar to the bottom half of remote object
 194                  * getattr, set this flag so that MDT skips checking whether
 195                  * it's remote object.
 196                  */
 197                 op_data->op_bias = MDS_CROSS_REF;
 198
 199                 tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[i].lmo_mds, NULL);
 200                 if (IS_ERR(tgt))
 201                         GOTO(cleanup, rc = PTR_ERR(tgt));
 202
 203                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 204                        PFID(&fid), tgt->ltd_idx);
 205
 206                 if (req != NULL) {
 207                         ptlrpc_req_finished(req);
 208                         req = NULL;
 209                 }
 210
 211                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 212                                     cb_blocking, extra_lock_flags);
 213                 if (rc == -ENOENT) {
 214                         /* skip stripe is not exists */
 215                         rc = 0;
 216                         continue;
 217                 }
 218
 219                 if (rc < 0)
 220                         GOTO(cleanup, rc);
 221
 222                 lockh = (struct lustre_handle *)&it.it_lock_handle;
 223                 if (rc > 0 && req == NULL) {
 224                         /* slave inode is still valid */
 225                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 226                                PFID(&fid));
 227                         rc = 0;
 228                 } else {
 229                         /* refresh slave from server */
 230                         body = req_capsule_server_get(&req->rq_pill,
 231                                                       &RMF_MDT_BODY);
 232                         if (body == NULL) {
 233                                 if (it.it_lock_mode && lockh) {
 234                                         ldlm_lock_decref(lockh,
 235                                                  it.it_lock_mode);
 236                                         it.it_lock_mode = 0;
 237                                 }
 238                                 GOTO(cleanup, rc = -ENOENT);
 239                         }
 240
 241                         i_size_write(inode, body->mbo_size);
 242                         inode->i_blocks = body->mbo_blocks;
 243                         set_nlink(inode, body->mbo_nlink);
 244                         inode->i_atime.tv_sec = body->mbo_atime;
 245                         inode->i_ctime.tv_sec = body->mbo_ctime;
 246                         inode->i_mtime.tv_sec = body->mbo_mtime;
 247                 }
 248
 249                 md_set_lock_data(tgt->ltd_exp, lockh, inode, NULL);
 250                 if (it.it_lock_mode != 0 && lockh != NULL) {
 251                         ldlm_lock_decref(lockh, it.it_lock_mode);
 252                         it.it_lock_mode = 0;
 253                 }
 254
 255                 valid_stripe_count++;
 256         }
 257
 258 cleanup:
 259         if (req != NULL)
 260                 ptlrpc_req_finished(req);
 261
 262         /* if all stripes are invalid, return -ENOENT to notify user */
 263         if (!rc && !valid_stripe_count)
 264                 rc = -ENOENT;
 265
 266         OBD_FREE_PTR(op_data);
 267         RETURN(rc);
 268 }
 269
 270 /*
 271  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 272  * may be split dir.
 273  */
 274 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 275                            struct lookup_intent *it,
 276                            struct ptlrpc_request **reqp,
 277                            ldlm_blocking_callback cb_blocking,
 278                            __u64 extra_lock_flags)
 279 {
 280         struct obd_device *obd = exp->exp_obd;
 281         struct lmv_obd *lmv = &obd->u.lmv;
 282         struct lmv_tgt_desc *tgt;
 283         struct mdt_body *body;
 284         __u64 flags = it->it_flags;
 285         int rc;
 286
 287         ENTRY;
 288
 289         if ((it->it_op & IT_CREAT) && !(flags & MDS_OPEN_BY_FID)) {
 290                 /* don't allow create under dir with bad hash */
 291                 if (lmv_is_dir_bad_hash(op_data->op_mea1))
 292                         RETURN(-EBADF);
 293
 294                 if (lmv_is_dir_migrating(op_data->op_mea1)) {
 295                         if (flags & O_EXCL) {
 296                                 /*
 297                                  * open(O_CREAT | O_EXCL) needs to check
 298                                  * existing name, which should be done on both
 299                                  * old and new layout, to avoid creating new
 300                                  * file under old layout, check old layout on
 301                                  * client side.
 302                                  */
 303                                 tgt = lmv_locate_tgt(lmv, op_data,
 304                                                      &op_data->op_fid1);
 305                                 if (IS_ERR(tgt))
 306                                         RETURN(PTR_ERR(tgt));
 307
 308                                 rc = md_getattr_name(tgt->ltd_exp, op_data,
 309                                                      reqp);
 310                                 if (!rc) {
 311                                         ptlrpc_req_finished(*reqp);
 312                                         *reqp = NULL;
 313                                         RETURN(-EEXIST);
 314                                 }
 315
 316                                 if (rc != -ENOENT)
 317                                         RETURN(rc);
 318
 319                                 op_data->op_post_migrate = true;
 320                         } else {
 321                                 /*
 322                                  * open(O_CREAT) will be sent to MDT in old
 323                                  * layout first, to avoid creating new file
 324                                  * under old layout, clear O_CREAT.
 325                                  */
 326                                 it->it_flags &= ~O_CREAT;
 327                         }
 328                 }
 329         }
 330
 331 retry:
 332         if (it->it_flags & MDS_OPEN_BY_FID) {
 333                 LASSERT(fid_is_sane(&op_data->op_fid2));
 334
 335                 /* for striped directory, we can't know parent stripe fid
 336                  * without name, but we can set it to child fid, and MDT
 337                  * will obtain it from linkea in open in such case. */
 338                 if (op_data->op_mea1 != NULL)
 339                         op_data->op_fid1 = op_data->op_fid2;
 340
 341                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 342                 if (IS_ERR(tgt))
 343                         RETURN(PTR_ERR(tgt));
 344
 345                 op_data->op_mds = tgt->ltd_idx;
 346         } else {
 347                 LASSERT(fid_is_sane(&op_data->op_fid1));
 348                 LASSERT(fid_is_zero(&op_data->op_fid2));
 349                 LASSERT(op_data->op_name != NULL);
 350
 351                 tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
 352                 if (IS_ERR(tgt))
 353                         RETURN(PTR_ERR(tgt));
 354         }
 355
 356         /* If it is ready to open the file by FID, do not need
 357          * allocate FID at all, otherwise it will confuse MDT */
 358         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 359                 /*
 360                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 361                  * for it.
 362                  */
 363                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 364                 if (rc != 0)
 365                         RETURN(rc);
 366         }
 367
 368         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 369                " name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
 370                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 371
 372         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 373                             extra_lock_flags);
 374         if (rc != 0)
 375                 RETURN(rc);
 376         /*
 377          * Nothing is found, do not access body->fid1 as it is zero and thus
 378          * pointless.
 379          */
 380         if ((it->it_disposition & DISP_LOOKUP_NEG) &&
 381             !(it->it_disposition & DISP_OPEN_CREATE) &&
 382             !(it->it_disposition & DISP_OPEN_OPEN)) {
 383                 if (!(it->it_flags & MDS_OPEN_BY_FID) &&
 384                     lmv_dir_retry_check_update(op_data)) {
 385                         ptlrpc_req_finished(*reqp);
 386                         it->it_request = NULL;
 387                         it->it_disposition = 0;
 388                         *reqp = NULL;
 389
 390                         it->it_flags = flags;
 391                         fid_zero(&op_data->op_fid2);
 392                         goto retry;
 393                 }
 394
 395                 RETURN(rc);
 396         }
 397
 398         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 399         if (body == NULL)
 400                 RETURN(-EPROTO);
 401
 402         /* Not cross-ref case, just get out of here. */
 403         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 404                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 405                                        cb_blocking, extra_lock_flags);
 406                 if (rc != 0)
 407                         RETURN(rc);
 408
 409                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 410                 if (body == NULL)
 411                         RETURN(-EPROTO);
 412         }
 413
 414         RETURN(rc);
 415 }
 416
 417 /*
 418  * Handler for: getattr, lookup and revalidate cases.
 419  */
 420 static int
 421 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 422                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 423                   ldlm_blocking_callback cb_blocking,
 424                   __u64 extra_lock_flags)
 425 {
 426         struct obd_device *obd = exp->exp_obd;
 427         struct lmv_obd *lmv = &obd->u.lmv;
 428         struct lmv_tgt_desc *tgt = NULL;
 429         struct mdt_body *body;
 430         int rc;
 431         ENTRY;
 432
 433 retry:
 434         if (op_data->op_flags & MF_GETATTR_BY_FID) {
 435                 /* getattr by FID, replace fid1 with stripe FID */
 436                 LASSERT(op_data->op_name);
 437                 tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
 438                 if (IS_ERR(tgt))
 439                         RETURN(PTR_ERR(tgt));
 440
 441                 /* name is used to locate stripe target, clear it here
 442                  * to avoid packing name in request, so that MDS knows
 443                  * it's getattr by FID.
 444                  */
 445                 op_data->op_name = NULL;
 446                 op_data->op_namelen = 0;
 447
 448                 /* getattr request is sent to MDT where fid2 inode is */
 449                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 450         } else if (op_data->op_name) {
 451                 /* getattr by name */
 452                 tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
 453                 if (!fid_is_sane(&op_data->op_fid2))
 454                         fid_zero(&op_data->op_fid2);
 455         } else {
 456                 /* old way to getattr by FID, parent FID not packed */
 457                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
 458         }
 459         if (IS_ERR(tgt))
 460                 RETURN(PTR_ERR(tgt));
 461
 462         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 463                ", name='%s' -> mds #%u\n",
 464                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 465                op_data->op_name ? op_data->op_name : "<NULL>",
 466                tgt->ltd_idx);
 467
 468         op_data->op_bias &= ~MDS_CROSS_REF;
 469
 470         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 471                             extra_lock_flags);
 472         if (rc < 0)
 473                 RETURN(rc);
 474
 475         if (*reqp == NULL) {
 476                 /* If RPC happens, lsm information will be revalidated
 477                  * during update_inode process (see ll_update_lsm_md) */
 478                 if (op_data->op_mea2 != NULL) {
 479                         rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
 480                                                    cb_blocking,
 481                                                    extra_lock_flags);
 482                         if (rc != 0)
 483                                 RETURN(rc);
 484                 }
 485                 RETURN(rc);
 486         } else if (it_disposition(it, DISP_LOOKUP_NEG) &&
 487                    lmv_dir_retry_check_update(op_data)) {
 488                 ptlrpc_req_finished(*reqp);
 489                 it->it_request = NULL;
 490                 it->it_disposition = 0;
 491                 *reqp = NULL;
 492
 493                 goto retry;
 494         }
 495
 496         if (!it_has_reply_body(it))
 497                 RETURN(0);
 498
 499         /*
 500          * MDS has returned success. Probably name has been resolved in
 501          * remote inode. Let's check this.
 502          */
 503         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 504         if (body == NULL)
 505                 RETURN(-EPROTO);
 506
 507         /* Not cross-ref case, just get out of here. */
 508         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 509                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 510                                        extra_lock_flags);
 511                 if (rc != 0)
 512                         RETURN(rc);
 513                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 514                 if (body == NULL)
 515                         RETURN(-EPROTO);
 516         }
 517
 518         RETURN(rc);
 519 }
 520
 521 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 522                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 523                     ldlm_blocking_callback cb_blocking,
 524                     __u64 extra_lock_flags)
 525 {
 526         int rc;
 527         ENTRY;
 528
 529         LASSERT(it != NULL);
 530         LASSERT(fid_is_sane(&op_data->op_fid1));
 531
 532         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 533                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 534                 (int)op_data->op_namelen, op_data->op_name,
 535                 PFID(&op_data->op_fid1));
 536
 537         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT | IT_GETXATTR))
 538                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 539                                        extra_lock_flags);
 540         else if (it->it_op & IT_OPEN)
 541                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 542                                      extra_lock_flags);
 543         else
 544                 LBUG();
 545
 546         if (rc < 0) {
 547                 struct lustre_handle lock_handle;
 548
 549                 if (it->it_lock_mode != 0) {
 550                         lock_handle.cookie = it->it_lock_handle;
 551                         ldlm_lock_decref_and_cancel(&lock_handle,
 552                                                     it->it_lock_mode);
 553                 }
 554
 555                 it->it_lock_handle = 0;
 556                 it->it_lock_mode = 0;
 557
 558                 if (it->it_remote_lock_mode != 0) {
 559                         lock_handle.cookie = it->it_remote_lock_handle;
 560                         ldlm_lock_decref_and_cancel(&lock_handle,
 561                                                     it->it_remote_lock_mode);
 562                 }
 563
 564                 it->it_remote_lock_handle = 0;
 565                 it->it_remote_lock_mode = 0;
 566         }
 567
 568         RETURN(rc);
 569 }