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          * Unfortunately, we have to lie to MDC/MDS to retrieve
  78          * attributes llite needs and provideproper locking.
  79          */
  80         if (it->it_op & IT_LOOKUP)
  81                 it->it_op = IT_GETATTR;
  82
  83         /*
  84          * We got LOOKUP lock, but we really need attrs.
  85          */
  86         pmode = it->it_lock_mode;
  87         if (pmode) {
  88                 plock.cookie = it->it_lock_handle;
  89                 it->it_lock_mode = 0;
  90                 it->it_request = NULL;
  91         }
  92
  93         LASSERT(fid_is_sane(&body->mbo_fid1));
  94
  95         tgt = lmv_find_target(lmv, &body->mbo_fid1);
  96         if (IS_ERR(tgt))
  97                 GOTO(out, rc = PTR_ERR(tgt));
  98
  99         OBD_ALLOC_PTR(op_data);
 100         if (op_data == NULL)
 101                 GOTO(out, rc = -ENOMEM);
 102
 103         op_data->op_fid1 = body->mbo_fid1;
 104         /* Sent the parent FID to the remote MDT */
 105         if (parent_fid != NULL) {
 106                 /* The parent fid is only for remote open to
 107                  * check whether the open is from OBF,
 108                  * see mdt_cross_open */
 109                 LASSERT(it->it_op & IT_OPEN);
 110                 op_data->op_fid2 = *parent_fid;
 111         }
 112
 113         op_data->op_bias = MDS_CROSS_REF;
 114         CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%u\n",
 115                PFID(&body->mbo_fid1), tgt->ltd_idx);
 116
 117         rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
 118                             extra_lock_flags);
 119         if (rc)
 120                 GOTO(out_free_op_data, rc);
 121
 122         /*
 123          * LLite needs LOOKUP lock to track dentry revocation in order to
 124          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 125          * and put LOOKUP in request.
 126          */
 127         if (it->it_lock_mode != 0) {
 128                 it->it_remote_lock_handle =
 129                                         it->it_lock_handle;
 130                 it->it_remote_lock_mode = it->it_lock_mode;
 131         }
 132
 133         if (pmode) {
 134                 it->it_lock_handle = plock.cookie;
 135                 it->it_lock_mode = pmode;
 136         }
 137
 138         EXIT;
 139 out_free_op_data:
 140         OBD_FREE_PTR(op_data);
 141 out:
 142         if (rc && pmode)
 143                 ldlm_lock_decref(&plock, pmode);
 144
 145         ptlrpc_req_finished(*reqp);
 146         *reqp = req;
 147         return rc;
 148 }
 149
 150 int lmv_revalidate_slaves(struct obd_export *exp,
 151                           const struct lmv_stripe_md *lsm,
 152                           ldlm_blocking_callback cb_blocking,
 153                           int extra_lock_flags)
 154 {
 155         struct obd_device      *obd = exp->exp_obd;
 156         struct lmv_obd         *lmv = &obd->u.lmv;
 157         struct ptlrpc_request   *req = NULL;
 158         struct mdt_body         *body;
 159         struct md_op_data      *op_data;
 160         int                     i;
 161         int                     rc = 0;
 162
 163         ENTRY;
 164
 165         /**
 166          * revalidate slaves has some problems, temporarily return,
 167          * we may not need that
 168          */
 169         OBD_ALLOC_PTR(op_data);
 170         if (op_data == NULL)
 171                 RETURN(-ENOMEM);
 172
 173         /**
 174          * Loop over the stripe information, check validity and update them
 175          * from MDS if needed.
 176          */
 177         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 178                 struct lu_fid           fid;
 179                 struct lookup_intent    it = { .it_op = IT_GETATTR };
 180                 struct lustre_handle    *lockh = NULL;
 181                 struct lmv_tgt_desc     *tgt = NULL;
 182                 struct inode            *inode;
 183
 184                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 185                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 186
 187                 /*
 188                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 189                  * defined otherwise it will go to server and take new lock
 190                  * which is not needed here.
 191                  */
 192                 memset(op_data, 0, sizeof(*op_data));
 193                 op_data->op_fid1 = fid;
 194                 op_data->op_fid2 = fid;
 195
 196                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 197                 if (IS_ERR(tgt))
 198                         GOTO(cleanup, rc = PTR_ERR(tgt));
 199
 200                 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
 201                        PFID(&fid), tgt->ltd_idx);
 202
 203                 if (req != NULL) {
 204                         ptlrpc_req_finished(req);
 205                         req = NULL;
 206                 }
 207
 208                 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
 209                                     cb_blocking, extra_lock_flags);
 210                 if (rc < 0)
 211                         GOTO(cleanup, rc);
 212
 213                 lockh = (struct lustre_handle *)&it.it_lock_handle;
 214                 if (rc > 0 && req == NULL) {
 215                         /* slave inode is still valid */
 216                         CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
 217                                PFID(&fid));
 218                         rc = 0;
 219                 } else {
 220                         /* refresh slave from server */
 221                         body = req_capsule_server_get(&req->rq_pill,
 222                                                       &RMF_MDT_BODY);
 223                         if (body == NULL) {
 224                                 if (it.it_lock_mode && lockh) {
 225                                         ldlm_lock_decref(lockh,
 226                                                  it.it_lock_mode);
 227                                         it.it_lock_mode = 0;
 228                                 }
 229                                 GOTO(cleanup, rc = -ENOENT);
 230                         }
 231
 232                         i_size_write(inode, body->mbo_size);
 233                         inode->i_blocks = body->mbo_blocks;
 234                         set_nlink(inode, body->mbo_nlink);
 235                         LTIME_S(inode->i_atime) = body->mbo_atime;
 236                         LTIME_S(inode->i_ctime) = body->mbo_ctime;
 237                         LTIME_S(inode->i_mtime) = body->mbo_mtime;
 238                 }
 239
 240                 md_set_lock_data(tgt->ltd_exp, lockh, inode, NULL);
 241                 if (it.it_lock_mode != 0 && lockh != NULL) {
 242                         ldlm_lock_decref(lockh, it.it_lock_mode);
 243                         it.it_lock_mode = 0;
 244                 }
 245         }
 246
 247 cleanup:
 248         if (req != NULL)
 249                 ptlrpc_req_finished(req);
 250
 251         OBD_FREE_PTR(op_data);
 252         RETURN(rc);
 253 }
 254
 255
 256 /*
 257  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 258  * may be split dir.
 259  */
 260 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 261                            struct lookup_intent *it,
 262                            struct ptlrpc_request **reqp,
 263                            ldlm_blocking_callback cb_blocking,
 264                            __u64 extra_lock_flags)
 265 {
 266         struct obd_device       *obd = exp->exp_obd;
 267         struct lmv_obd          *lmv = &obd->u.lmv;
 268         struct lmv_tgt_desc     *tgt;
 269         struct mdt_body         *body;
 270         int                     rc;
 271         ENTRY;
 272
 273         if (it->it_flags & MDS_OPEN_BY_FID) {
 274                 LASSERT(fid_is_sane(&op_data->op_fid2));
 275
 276                 /* for striped directory, we can't know parent stripe fid
 277                  * without name, but we can set it to child fid, and MDT
 278                  * will obtain it from linkea in open in such case. */
 279                 if (op_data->op_mea1 != NULL)
 280                         op_data->op_fid1 = op_data->op_fid2;
 281
 282                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 283                 if (IS_ERR(tgt))
 284                         RETURN(PTR_ERR(tgt));
 285
 286                 op_data->op_mds = tgt->ltd_idx;
 287         } else {
 288                 LASSERT(fid_is_sane(&op_data->op_fid1));
 289                 LASSERT(fid_is_zero(&op_data->op_fid2));
 290                 LASSERT(op_data->op_name != NULL);
 291
 292                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 293                 if (IS_ERR(tgt))
 294                         RETURN(PTR_ERR(tgt));
 295         }
 296
 297         /* If it is ready to open the file by FID, do not need
 298          * allocate FID at all, otherwise it will confuse MDT */
 299         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 300                 /*
 301                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 302                  * for it.
 303                  */
 304                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 305                 if (rc != 0)
 306                         RETURN(rc);
 307         }
 308
 309         CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
 310                " name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
 311                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 312
 313         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 314                             extra_lock_flags);
 315         if (rc != 0)
 316                 RETURN(rc);
 317         /*
 318          * Nothing is found, do not access body->fid1 as it is zero and thus
 319          * pointless.
 320          */
 321         if ((it->it_disposition & DISP_LOOKUP_NEG) &&
 322             !(it->it_disposition & DISP_OPEN_CREATE) &&
 323             !(it->it_disposition & DISP_OPEN_OPEN))
 324                 RETURN(rc);
 325
 326         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 327         if (body == NULL)
 328                 RETURN(-EPROTO);
 329
 330         /* Not cross-ref case, just get out of here. */
 331         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 332                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 333                                        cb_blocking, extra_lock_flags);
 334                 if (rc != 0)
 335                         RETURN(rc);
 336
 337                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 338                 if (body == NULL)
 339                         RETURN(-EPROTO);
 340         }
 341
 342         RETURN(rc);
 343 }
 344
 345 /*
 346  * Handler for: getattr, lookup and revalidate cases.
 347  */
 348 static int
 349 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
 350                   struct lookup_intent *it, struct ptlrpc_request **reqp,
 351                   ldlm_blocking_callback cb_blocking,
 352                   __u64 extra_lock_flags)
 353 {
 354         struct obd_device       *obd = exp->exp_obd;
 355         struct lmv_obd          *lmv = &obd->u.lmv;
 356         struct lmv_tgt_desc     *tgt = NULL;
 357         struct mdt_body         *body;
 358         struct lmv_stripe_md    *lsm = op_data->op_mea1;
 359         int                     rc = 0;
 360         ENTRY;
 361
 362         /* If it returns ERR_PTR(-EBADFD) then it is an unknown hash type
 363          * it will try all stripes to locate the object */
 364         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 365         if (IS_ERR(tgt) && (PTR_ERR(tgt) != -EBADFD))
 366                 RETURN(PTR_ERR(tgt));
 367
 368         /* Both migrating dir and unknown hash dir need to try
 369          * all of sub-stripes */
 370         if (lsm != NULL && !lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
 371                 struct lmv_oinfo *oinfo;
 372
 373                 oinfo = &lsm->lsm_md_oinfo[0];
 374
 375                 op_data->op_fid1 = oinfo->lmo_fid;
 376                 op_data->op_mds = oinfo->lmo_mds;
 377                 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
 378                 if (IS_ERR(tgt))
 379                         RETURN(PTR_ERR(tgt));
 380         }
 381
 382         if (!fid_is_sane(&op_data->op_fid2))
 383                 fid_zero(&op_data->op_fid2);
 384
 385         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
 386                ", name='%s' -> mds #%u lsm=%p lsm_magic=%x\n",
 387                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 388                op_data->op_name ? op_data->op_name : "<NULL>",
 389                tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
 390
 391         op_data->op_bias &= ~MDS_CROSS_REF;
 392
 393         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 394                             extra_lock_flags);
 395         if (rc < 0)
 396                 RETURN(rc);
 397
 398         if (*reqp == NULL) {
 399                 /* If RPC happens, lsm information will be revalidated
 400                  * during update_inode process (see ll_update_lsm_md) */
 401                 if (op_data->op_mea2 != NULL) {
 402                         rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
 403                                                    cb_blocking,
 404                                                    extra_lock_flags);
 405                         if (rc != 0)
 406                                 RETURN(rc);
 407                 }
 408                 RETURN(rc);
 409         } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
 410                    lmv_need_try_all_stripes(lsm)) {
 411                 /* For migrating and unknown hash type directory, it will
 412                  * try to target the entry on other stripes */
 413                 int stripe_index;
 414
 415                 for (stripe_index = 1;
 416                      stripe_index < lsm->lsm_md_stripe_count &&
 417                      it_disposition(it, DISP_LOOKUP_NEG); stripe_index++) {
 418                         struct lmv_oinfo *oinfo;
 419
 420                         /* release the previous request */
 421                         ptlrpc_req_finished(*reqp);
 422                         it->it_request = NULL;
 423                         *reqp = NULL;
 424
 425                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
 426                         tgt = lmv_find_target(lmv, &oinfo->lmo_fid);
 427                         if (IS_ERR(tgt))
 428                                 RETURN(PTR_ERR(tgt));
 429
 430                         CDEBUG(D_INODE, "Try other stripes " DFID"\n",
 431                                PFID(&oinfo->lmo_fid));
 432
 433                         op_data->op_fid1 = oinfo->lmo_fid;
 434                         it->it_disposition &= ~DISP_ENQ_COMPLETE;
 435                         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
 436                                             cb_blocking, extra_lock_flags);
 437                         if (rc != 0)
 438                                 RETURN(rc);
 439                 }
 440         }
 441
 442         if (!it_has_reply_body(it))
 443                 RETURN(0);
 444
 445         /*
 446          * MDS has returned success. Probably name has been resolved in
 447          * remote inode. Let's check this.
 448          */
 449         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 450         if (body == NULL)
 451                 RETURN(-EPROTO);
 452
 453         /* Not cross-ref case, just get out of here. */
 454         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 455                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 456                                        extra_lock_flags);
 457                 if (rc != 0)
 458                         RETURN(rc);
 459                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 460                 if (body == NULL)
 461                         RETURN(-EPROTO);
 462         }
 463
 464         RETURN(rc);
 465 }
 466
 467 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 468                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 469                     ldlm_blocking_callback cb_blocking,
 470                     __u64 extra_lock_flags)
 471 {
 472         int rc;
 473         ENTRY;
 474
 475         LASSERT(it != NULL);
 476         LASSERT(fid_is_sane(&op_data->op_fid1));
 477
 478         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
 479                 LL_IT2STR(it), PFID(&op_data->op_fid2),
 480                 (int)op_data->op_namelen, op_data->op_name,
 481                 PFID(&op_data->op_fid1));
 482
 483         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT | IT_GETXATTR))
 484                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 485                                        extra_lock_flags);
 486         else if (it->it_op & IT_OPEN)
 487                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 488                                      extra_lock_flags);
 489         else
 490                 LBUG();
 491
 492         if (rc < 0) {
 493                 struct lustre_handle lock_handle;
 494
 495                 if (it->it_lock_mode != 0) {
 496                         lock_handle.cookie = it->it_lock_handle;
 497                         ldlm_lock_decref_and_cancel(&lock_handle,
 498                                                     it->it_lock_mode);
 499                 }
 500
 501                 it->it_lock_handle = 0;
 502                 it->it_lock_mode = 0;
 503
 504                 if (it->it_remote_lock_mode != 0) {
 505                         lock_handle.cookie = it->it_remote_lock_handle;
 506                         ldlm_lock_decref_and_cancel(&lock_handle,
 507                                                     it->it_remote_lock_mode);
 508                 }
 509
 510                 it->it_remote_lock_handle = 0;
 511                 it->it_remote_lock_mode = 0;
 512         }
 513
 514         RETURN(rc);
 515 }