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_find_target(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_idx);
 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                     rc = 0;
 166
 167         ENTRY;
 168
 169         /**
 170          * revalidate slaves has some problems, temporarily return,
 171          * we may not need that
 172          */
 173         OBD_ALLOC_PTR(op_data);
 174         if (op_data == NULL)
 175                 RETURN(-ENOMEM);
 176
 177         /**
 178          * Loop over the stripe information, check validity and update them
 179          * from MDS if needed.
 180          */
 181         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 182                 struct lu_fid           fid;
 183                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 184                 struct lustre_handle    *lockh = NULL;
 185                 struct lmv_tgt_desc     *tgt = NULL;
 186                 struct inode            *inode;
 187
 188                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 189                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 190
 191                 /*
 192                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 193                  * defined otherwise it will go to server and take new lock
 194                  * which is not needed here.
 195                  */
 196                 memset(op_data, 0, sizeof(*op_data));
 197                 op_data->op_fid1 = fid;
 198                 op_data->op_fid2 = fid;
 199
 200                 tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[i].lmo_mds, NULL);
 201                 if (IS_ERR(tgt))
 202                         GOTO(cleanup, rc = PTR_ERR(tgt));
 203
 204                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 205                        PFID(&fid), tgt->ltd_idx);
 206
 207                 if (req != NULL) {
 208                         ptlrpc_req_finished(req);
 209                         req = NULL;
 210                 }
 211
 212                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 213                                     cb_blocking, extra_lock_flags);
 214                 if (rc < 0)
 215                         GOTO(cleanup, rc);
 216
 217                 lockh = (struct lustre_handle *)&it.it_lock_handle;
 218                 if (rc > 0 && req == NULL) {
 219                         /* slave inode is still valid */
 220                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 221                                PFID(&fid));
 222                         rc = 0;
 223                 } else {
 224                         /* refresh slave from server */
 225                         body = req_capsule_server_get(&req->rq_pill,
 226                                                       &RMF_MDT_BODY);
 227                         if (body == NULL) {
 228                                 if (it.it_lock_mode && lockh) {
 229                                         ldlm_lock_decref(lockh,
 230                                                  it.it_lock_mode);
 231                                         it.it_lock_mode = 0;
 232                                 }
 233                                 GOTO(cleanup, rc = -ENOENT);
 234                         }
 235
 236                         i_size_write(inode, body->mbo_size);
 237                         inode->i_blocks = body->mbo_blocks;
 238                         set_nlink(inode, body->mbo_nlink);
 239                         inode->i_atime.tv_sec = body->mbo_atime;
 240                         inode->i_ctime.tv_sec = body->mbo_ctime;
 241                         inode->i_mtime.tv_sec = body->mbo_mtime;
 242                 }
 243
 244                 md_set_lock_data(tgt->ltd_exp, lockh, inode, NULL);
 245                 if (it.it_lock_mode != 0 && lockh != NULL) {
 246                         ldlm_lock_decref(lockh, it.it_lock_mode);
 247                         it.it_lock_mode = 0;
 248                 }
 249         }
 250
 251 cleanup:
 252         if (req != NULL)
 253                 ptlrpc_req_finished(req);
 254
 255         OBD_FREE_PTR(op_data);
 256         RETURN(rc);
 257 }
 258
 259
 260 /*
 261  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 262  * may be split dir.
 263  */
 264 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 265                            struct lookup_intent *it,
 266                            struct ptlrpc_request **reqp,
 267                            ldlm_blocking_callback cb_blocking,
 268                            __u64 extra_lock_flags)
 269 {
 270         struct obd_device *obd = exp->exp_obd;
 271         struct lmv_obd *lmv = &obd->u.lmv;
 272         struct lmv_tgt_desc *tgt;
 273         struct mdt_body *body;
 274         __u64 flags = it->it_flags;
 275         int rc;
 276
 277         ENTRY;
 278
 279         if ((it->it_op & IT_CREAT) && !(flags & MDS_OPEN_BY_FID)) {
 280                 /* don't allow create under dir with bad hash */
 281                 if (lmv_is_dir_bad_hash(op_data->op_mea1))
 282                         RETURN(-EBADF);
 283
 284                 if (lmv_is_dir_migrating(op_data->op_mea1)) {
 285                         if (flags & O_EXCL) {
 286                                 /*
 287                                  * open(O_CREAT | O_EXCL) needs to check
 288                                  * existing name, which should be done on both
 289                                  * old and new layout, to avoid creating new
 290                                  * file under old layout, check old layout on
 291                                  * client side.
 292                                  */
 293                                 tgt = lmv_locate_tgt(lmv, op_data,
 294                                                      &op_data->op_fid1);
 295                                 if (IS_ERR(tgt))
 296                                         RETURN(PTR_ERR(tgt));
 297
 298                                 rc = md_getattr_name(tgt->ltd_exp, op_data,
 299                                                      reqp);
 300                                 if (!rc) {
 301                                         ptlrpc_req_finished(*reqp);
 302                                         *reqp = NULL;
 303                                         RETURN(-EEXIST);
 304                                 }
 305
 306                                 if (rc != -ENOENT)
 307                                         RETURN(rc);
 308
 309                                 op_data->op_post_migrate = true;
 310                         } else {
 311                                 /*
 312                                  * open(O_CREAT) will be sent to MDT in old
 313                                  * layout first, to avoid creating new file
 314                                  * under old layout, clear O_CREAT.
 315                                  */
 316                                 it->it_flags &= ~O_CREAT;
 317                         }
 318                 }
 319         }
 320
 321 retry:
 322         if (it->it_flags & MDS_OPEN_BY_FID) {
 323                 LASSERT(fid_is_sane(&op_data->op_fid2));
 324
 325                 /* for striped directory, we can't know parent stripe fid
 326                  * without name, but we can set it to child fid, and MDT
 327                  * will obtain it from linkea in open in such case. */
 328                 if (op_data->op_mea1 != NULL)
 329                         op_data->op_fid1 = op_data->op_fid2;
 330
 331                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 332                 if (IS_ERR(tgt))
 333                         RETURN(PTR_ERR(tgt));
 334
 335                 op_data->op_mds = tgt->ltd_idx;
 336         } else {
 337                 LASSERT(fid_is_sane(&op_data->op_fid1));
 338                 LASSERT(fid_is_zero(&op_data->op_fid2));
 339                 LASSERT(op_data->op_name != NULL);
 340
 341                 tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
 342                 if (IS_ERR(tgt))
 343                         RETURN(PTR_ERR(tgt));
 344         }
 345
 346         /* If it is ready to open the file by FID, do not need
 347          * allocate FID at all, otherwise it will confuse MDT */
 348         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 349                 /*
 350                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 351                  * for it.
 352                  */
 353                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 354                 if (rc != 0)
 355                         RETURN(rc);
 356         }
 357
 358         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 359                " name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
 360                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 361
 362         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 363                             extra_lock_flags);
 364         if (rc != 0)
 365                 RETURN(rc);
 366         /*
 367          * Nothing is found, do not access body->fid1 as it is zero and thus
 368          * pointless.
 369          */
 370         if ((it->it_disposition & DISP_LOOKUP_NEG) &&
 371             !(it->it_disposition & DISP_OPEN_CREATE) &&
 372             !(it->it_disposition & DISP_OPEN_OPEN)) {
 373                 if (!(it->it_flags & MDS_OPEN_BY_FID) &&
 374                     lmv_dir_retry_check_update(op_data)) {
 375                         ptlrpc_req_finished(*reqp);
 376                         it->it_request = NULL;
 377                         it->it_disposition = 0;
 378                         *reqp = NULL;
 379
 380                         it->it_flags = flags;
 381                         fid_zero(&op_data->op_fid2);
 382                         goto retry;
 383                 }
 384
 385                 RETURN(rc);
 386         }
 387
 388         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 389         if (body == NULL)
 390                 RETURN(-EPROTO);
 391
 392         /* Not cross-ref case, just get out of here. */
 393         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 394                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 395                                        cb_blocking, extra_lock_flags,
 396                                        op_data->op_file_secctx_name,
 397                                        op_data->op_file_secctx_name_size);
 398                 if (rc != 0)
 399                         RETURN(rc);
 400
 401                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 402                 if (body == NULL)
 403                         RETURN(-EPROTO);
 404         }
 405
 406         RETURN(rc);
 407 }
 408
 409 /*
 410  * Handler for: getattr, lookup and revalidate cases.
 411  */
 412 static int
 413 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 414                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 415                   ldlm_blocking_callback cb_blocking,
 416                   __u64 extra_lock_flags)
 417 {
 418         struct obd_device *obd = exp->exp_obd;
 419         struct lmv_obd *lmv = &obd->u.lmv;
 420         struct lmv_tgt_desc *tgt = NULL;
 421         struct mdt_body *body;
 422         int rc;
 423         ENTRY;
 424
 425 retry:
 426         tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
 427         if (IS_ERR(tgt))
 428                 RETURN(PTR_ERR(tgt));
 429
 430         if (!fid_is_sane(&op_data->op_fid2))
 431                 fid_zero(&op_data->op_fid2);
 432
 433         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 434                ", name='%s' -> mds #%u\n",
 435                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 436                op_data->op_name ? op_data->op_name : "<NULL>",
 437                tgt->ltd_idx);
 438
 439         op_data->op_bias &= ~MDS_CROSS_REF;
 440
 441         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 442                             extra_lock_flags);
 443         if (rc < 0)
 444                 RETURN(rc);
 445
 446         if (*reqp == NULL) {
 447                 /* If RPC happens, lsm information will be revalidated
 448                  * during update_inode process (see ll_update_lsm_md) */
 449                 if (op_data->op_mea2 != NULL) {
 450                         rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
 451                                                    cb_blocking,
 452                                                    extra_lock_flags);
 453                         if (rc != 0)
 454                                 RETURN(rc);
 455                 }
 456                 RETURN(rc);
 457         } else if (it_disposition(it, DISP_LOOKUP_NEG) &&
 458                    lmv_dir_retry_check_update(op_data)) {
 459                 ptlrpc_req_finished(*reqp);
 460                 it->it_request = NULL;
 461                 it->it_disposition = 0;
 462                 *reqp = NULL;
 463
 464                 goto retry;
 465         }
 466
 467         if (!it_has_reply_body(it))
 468                 RETURN(0);
 469
 470         /*
 471          * MDS has returned success. Probably name has been resolved in
 472          * remote inode. Let's check this.
 473          */
 474         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 475         if (body == NULL)
 476                 RETURN(-EPROTO);
 477
 478         /* Not cross-ref case, just get out of here. */
 479         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 480                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 481                                        extra_lock_flags,
 482                                        op_data->op_file_secctx_name,
 483                                        op_data->op_file_secctx_name_size);
 484                 if (rc != 0)
 485                         RETURN(rc);
 486                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 487                 if (body == NULL)
 488                         RETURN(-EPROTO);
 489         }
 490
 491         RETURN(rc);
 492 }
 493
 494 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 495                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 496                     ldlm_blocking_callback cb_blocking,
 497                     __u64 extra_lock_flags)
 498 {
 499         int rc;
 500         ENTRY;
 501
 502         LASSERT(it != NULL);
 503         LASSERT(fid_is_sane(&op_data->op_fid1));
 504
 505         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 506                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 507                 (int)op_data->op_namelen, op_data->op_name,
 508                 PFID(&op_data->op_fid1));
 509
 510         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT | IT_GETXATTR))
 511                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 512                                        extra_lock_flags);
 513         else if (it->it_op & IT_OPEN)
 514                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 515                                      extra_lock_flags);
 516         else
 517                 LBUG();
 518
 519         if (rc < 0) {
 520                 struct lustre_handle lock_handle;
 521
 522                 if (it->it_lock_mode != 0) {
 523                         lock_handle.cookie = it->it_lock_handle;
 524                         ldlm_lock_decref_and_cancel(&lock_handle,
 525                                                     it->it_lock_mode);
 526                 }
 527
 528                 it->it_lock_handle = 0;
 529                 it->it_lock_mode = 0;
 530
 531                 if (it->it_remote_lock_mode != 0) {
 532                         lock_handle.cookie = it->it_remote_lock_handle;
 533                         ldlm_lock_decref_and_cancel(&lock_handle,
 534                                                     it->it_remote_lock_mode);
 535                 }
 536
 537                 it->it_remote_lock_handle = 0;
 538                 it->it_remote_lock_mode = 0;
 539         }
 540
 541         RETURN(rc);
 542 }