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 lu_fid *pfid,
 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 valid_stripe_count = 0;
 167         int rc = 0;
 168
 169         ENTRY;
 170
 171         /**
 172          * revalidate slaves has some problems, temporarily return,
 173          * we may not need that
 174          */
 175         OBD_ALLOC_PTR(op_data);
 176         if (op_data == NULL)
 177                 RETURN(-ENOMEM);
 178
 179         /**
 180          * Loop over the stripe information, check validity and update them
 181          * from MDS if needed.
 182          */
 183         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 184                 struct lu_fid           fid;
 185                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 186                 struct lustre_handle    *lockh = NULL;
 187                 struct lmv_tgt_desc     *tgt = NULL;
 188                 struct inode            *inode;
 189
 190                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 191                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 192
 193                 if (!inode)
 194                         continue;
 195
 196                 /*
 197                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 198                  * defined otherwise it will go to server and take new lock
 199                  * which is not needed here.
 200                  */
 201                 memset(op_data, 0, sizeof(*op_data));
 202                 op_data->op_fid1 = *pfid;
 203                 op_data->op_fid2 = fid;
 204
 205                 tgt = lmv_tgt(lmv, lsm->lsm_md_oinfo[i].lmo_mds);
 206                 if (!tgt)
 207                         GOTO(cleanup, rc = -ENODEV);
 208
 209                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 210                        PFID(&fid), tgt->ltd_index);
 211
 212                 if (req != NULL) {
 213                         ptlrpc_req_finished(req);
 214                         req = NULL;
 215                 }
 216
 217                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 218                                     cb_blocking, extra_lock_flags);
 219                 if (rc == -ENOENT) {
 220                         /* skip stripe is not exists */
 221                         rc = 0;
 222                         continue;
 223                 }
 224
 225                 if (rc < 0)
 226                         GOTO(cleanup, rc);
 227
 228                 lockh = (struct lustre_handle *)&it.it_lock_handle;
 229                 if (rc > 0 && req == NULL) {
 230                         /* slave inode is still valid */
 231                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 232                                PFID(&fid));
 233                         rc = 0;
 234                 } else {
 235                         /* refresh slave from server */
 236                         body = req_capsule_server_get(&req->rq_pill,
 237                                                       &RMF_MDT_BODY);
 238                         if (body == NULL) {
 239                                 if (it.it_lock_mode && lockh) {
 240                                         ldlm_lock_decref(lockh,
 241                                                  it.it_lock_mode);
 242                                         it.it_lock_mode = 0;
 243                                 }
 244                                 GOTO(cleanup, rc = -ENOENT);
 245                         }
 246
 247                         i_size_write(inode, body->mbo_size);
 248                         inode->i_blocks = body->mbo_blocks;
 249                         set_nlink(inode, body->mbo_nlink);
 250                         inode->i_atime.tv_sec = body->mbo_atime;
 251                         inode->i_ctime.tv_sec = body->mbo_ctime;
 252                         inode->i_mtime.tv_sec = body->mbo_mtime;
 253                 }
 254
 255                 md_set_lock_data(tgt->ltd_exp, lockh, inode, NULL);
 256                 if (it.it_lock_mode != 0 && lockh != NULL) {
 257                         ldlm_lock_decref(lockh, it.it_lock_mode);
 258                         it.it_lock_mode = 0;
 259                 }
 260
 261                 valid_stripe_count++;
 262         }
 263
 264 cleanup:
 265         if (req != NULL)
 266                 ptlrpc_req_finished(req);
 267
 268         /* if all stripes are invalid, return -ENOENT to notify user */
 269         if (!rc && !valid_stripe_count)
 270                 rc = -ENOENT;
 271
 272         OBD_FREE_PTR(op_data);
 273         RETURN(rc);
 274 }
 275
 276 /*
 277  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 278  * may be split dir.
 279  */
 280 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 281                            struct lookup_intent *it,
 282                            struct ptlrpc_request **reqp,
 283                            ldlm_blocking_callback cb_blocking,
 284                            __u64 extra_lock_flags)
 285 {
 286         struct obd_device *obd = exp->exp_obd;
 287         struct lmv_obd *lmv = &obd->u.lmv;
 288         struct lmv_tgt_desc *tgt;
 289         struct mdt_body *body;
 290         __u64 flags = it->it_flags;
 291         int rc;
 292
 293         ENTRY;
 294
 295         /* do not allow file creation in foreign dir */
 296         if ((it->it_op & IT_CREAT) && lmv_dir_foreign(op_data->op_mea1))
 297                 RETURN(-ENODATA);
 298
 299         if ((it->it_op & IT_CREAT) && !(flags & MDS_OPEN_BY_FID)) {
 300                 /* don't allow create under dir with bad hash */
 301                 if (lmv_dir_bad_hash(op_data->op_mea1))
 302                         RETURN(-EBADF);
 303
 304                 if (lmv_dir_layout_changing(op_data->op_mea1)) {
 305                         if (flags & O_EXCL) {
 306                                 /*
 307                                  * open(O_CREAT | O_EXCL) needs to check
 308                                  * existing name, which should be done on both
 309                                  * old and new layout, check old layout on
 310                                  * client side.
 311                                  */
 312                                 rc = lmv_old_layout_lookup(lmv, op_data);
 313                                 if (rc != -ENOENT)
 314                                         RETURN(rc);
 315
 316                                 op_data->op_new_layout = true;
 317                         } else {
 318                                 /*
 319                                  * open(O_CREAT) will be sent to MDT in old
 320                                  * layout first, to avoid creating new file
 321                                  * under old layout, clear O_CREAT.
 322                                  */
 323                                 it->it_flags &= ~O_CREAT;
 324                         }
 325                 }
 326         }
 327
 328 retry:
 329         if (it->it_flags & MDS_OPEN_BY_FID) {
 330                 LASSERT(fid_is_sane(&op_data->op_fid2));
 331
 332                 /* for striped directory, we can't know parent stripe fid
 333                  * without name, but we can set it to child fid, and MDT
 334                  * will obtain it from linkea in open in such case. */
 335                 if (lmv_dir_striped(op_data->op_mea1))
 336                         op_data->op_fid1 = op_data->op_fid2;
 337
 338                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
 339                 if (IS_ERR(tgt))
 340                         RETURN(PTR_ERR(tgt));
 341
 342                 op_data->op_mds = tgt->ltd_index;
 343         } else {
 344                 LASSERT(fid_is_sane(&op_data->op_fid1));
 345                 LASSERT(it->it_flags & MDS_OPEN_PCC ||
 346                         fid_is_zero(&op_data->op_fid2));
 347                 LASSERT(op_data->op_name != NULL);
 348
 349                 tgt = lmv_locate_tgt(lmv, op_data);
 350                 if (IS_ERR(tgt))
 351                         RETURN(PTR_ERR(tgt));
 352         }
 353
 354         /* If it is ready to open the file by FID, do not need
 355          * allocate FID at all, otherwise it will confuse MDT */
 356         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID ||
 357                                         it->it_flags & MDS_OPEN_PCC)) {
 358                 /*
 359                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 360                  * for it.
 361                  */
 362                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 363                 if (rc != 0)
 364                         RETURN(rc);
 365         }
 366
 367         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 368                " name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
 369                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_index);
 370
 371         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 372                             extra_lock_flags);
 373         if (rc != 0)
 374                 RETURN(rc);
 375         /*
 376          * Nothing is found, do not access body->fid1 as it is zero and thus
 377          * pointless.
 378          */
 379         if ((it->it_disposition & DISP_LOOKUP_NEG) &&
 380             !(it->it_disposition & DISP_OPEN_CREATE) &&
 381             !(it->it_disposition & DISP_OPEN_OPEN)) {
 382                 if (!(it->it_flags & MDS_OPEN_BY_FID) &&
 383                     lmv_dir_retry_check_update(op_data)) {
 384                         ptlrpc_req_finished(*reqp);
 385                         it->it_request = NULL;
 386                         it->it_disposition = 0;
 387                         *reqp = NULL;
 388
 389                         it->it_flags = flags;
 390                         fid_zero(&op_data->op_fid2);
 391                         goto retry;
 392                 }
 393
 394                 RETURN(rc);
 395         }
 396
 397         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 398         if (body == NULL)
 399                 RETURN(-EPROTO);
 400
 401         /* Not cross-ref case, just get out of here. */
 402         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 403                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 404                                        cb_blocking, extra_lock_flags,
 405                                        op_data->op_file_secctx_name,
 406                                        op_data->op_file_secctx_name_size);
 407                 if (rc != 0)
 408                         RETURN(rc);
 409
 410                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 411                 if (body == NULL)
 412                         RETURN(-EPROTO);
 413         }
 414
 415         RETURN(rc);
 416 }
 417
 418 /*
 419  * Handler for: getattr, lookup and revalidate cases.
 420  */
 421 static int
 422 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 423                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 424                   ldlm_blocking_callback cb_blocking,
 425                   __u64 extra_lock_flags)
 426 {
 427         struct obd_device *obd = exp->exp_obd;
 428         struct lmv_obd *lmv = &obd->u.lmv;
 429         struct lmv_tgt_desc *tgt = NULL;
 430         struct mdt_body *body;
 431         int rc;
 432         ENTRY;
 433
 434         /* foreign dir is not striped */
 435         if (lmv_dir_foreign(op_data->op_mea1)) {
 436                 /* only allow getattr/lookup for itself */
 437                 if (op_data->op_name != NULL)
 438                         RETURN(-ENODATA);
 439                 RETURN(0);
 440         }
 441
 442 retry:
 443         if (op_data->op_name) {
 444                 tgt = lmv_locate_tgt(lmv, op_data);
 445                 if (!fid_is_sane(&op_data->op_fid2))
 446                         fid_zero(&op_data->op_fid2);
 447         } else if (fid_is_sane(&op_data->op_fid2)) {
 448                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
 449         } else {
 450                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 451         }
 452         if (IS_ERR(tgt))
 453                 RETURN(PTR_ERR(tgt));
 454
 455         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 456                ", name='%s' -> mds #%u\n",
 457                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 458                op_data->op_name ? op_data->op_name : "<NULL>",
 459                tgt->ltd_index);
 460
 461         op_data->op_bias &= ~MDS_CROSS_REF;
 462
 463         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 464                             extra_lock_flags);
 465         if (rc < 0)
 466                 RETURN(rc);
 467
 468         if (*reqp == NULL) {
 469                 /* If RPC happens, lsm information will be revalidated
 470                  * during update_inode process (see ll_update_lsm_md) */
 471                 if (lmv_dir_striped(op_data->op_mea2)) {
 472                         rc = lmv_revalidate_slaves(exp, &op_data->op_fid2,
 473                                                    op_data->op_mea2,
 474                                                    cb_blocking,
 475                                                    extra_lock_flags);
 476                         if (rc != 0)
 477                                 RETURN(rc);
 478                 }
 479                 RETURN(rc);
 480         } else if (it_disposition(it, DISP_LOOKUP_NEG) &&
 481                    lmv_dir_retry_check_update(op_data)) {
 482                 ptlrpc_req_finished(*reqp);
 483                 it->it_request = NULL;
 484                 it->it_disposition = 0;
 485                 *reqp = NULL;
 486
 487                 goto retry;
 488         }
 489
 490         if (!it_has_reply_body(it))
 491                 RETURN(0);
 492
 493         /*
 494          * MDS has returned success. Probably name has been resolved in
 495          * remote inode. Let's check this.
 496          */
 497         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 498         if (body == NULL)
 499                 RETURN(-EPROTO);
 500
 501         /* Not cross-ref case, just get out of here. */
 502         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 503                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 504                                        extra_lock_flags,
 505                                        op_data->op_file_secctx_name,
 506                                        op_data->op_file_secctx_name_size);
 507                 if (rc != 0)
 508                         RETURN(rc);
 509                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 510                 if (body == NULL)
 511                         RETURN(-EPROTO);
 512         }
 513
 514         RETURN(rc);
 515 }
 516
 517 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 518                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 519                     ldlm_blocking_callback cb_blocking,
 520                     __u64 extra_lock_flags)
 521 {
 522         int rc;
 523         ENTRY;
 524
 525         LASSERT(it != NULL);
 526         LASSERT(fid_is_sane(&op_data->op_fid1));
 527
 528         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 529                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 530                 (int)op_data->op_namelen, op_data->op_name,
 531                 PFID(&op_data->op_fid1));
 532
 533         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT | IT_GETXATTR))
 534                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 535                                        extra_lock_flags);
 536         else if (it->it_op & IT_OPEN)
 537                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 538                                      extra_lock_flags);
 539         else
 540                 LBUG();
 541
 542         if (rc < 0) {
 543                 struct lustre_handle lock_handle;
 544
 545                 if (it->it_lock_mode != 0) {
 546                         lock_handle.cookie = it->it_lock_handle;
 547                         ldlm_lock_decref_and_cancel(&lock_handle,
 548                                                     it->it_lock_mode);
 549                 }
 550
 551                 it->it_lock_handle = 0;
 552                 it->it_lock_mode = 0;
 553
 554                 if (it->it_remote_lock_mode != 0) {
 555                         lock_handle.cookie = it->it_remote_lock_handle;
 556                         ldlm_lock_decref_and_cancel(&lock_handle,
 557                                                     it->it_remote_lock_mode);
 558                 }
 559
 560                 it->it_remote_lock_handle = 0;
 561                 it->it_remote_lock_mode = 0;
 562         }
 563
 564         RETURN(rc);
 565 }