drivers/staging/lustre/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, 2015, 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/pagemap.h>
  37 #include <asm/div64.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/namei.h>
  40 #include "../include/lustre_intent.h"
  41 #include "../include/obd_support.h"
  42 #include "../include/lustre/lustre_idl.h"
  43 #include "../include/lustre_lib.h"
  44 #include "../include/lustre_net.h"
  45 #include "../include/lustre_dlm.h"
  46 #include "../include/lustre_mdc.h"
  47 #include "../include/obd_class.h"
  48 #include "../include/lprocfs_status.h"
  49 #include "lmv_internal.h"
  50
  51 static int lmv_intent_remote(struct obd_export *exp, struct lookup_intent *it,
  52                              const struct lu_fid *parent_fid,
  53                              struct ptlrpc_request **reqp,
  54                              ldlm_blocking_callback cb_blocking,
  55                              __u64 extra_lock_flags)
  56 {
  57         struct obd_device       *obd = exp->exp_obd;
  58         struct lmv_obd          *lmv = &obd->u.lmv;
  59         struct ptlrpc_request   *req = NULL;
  60         struct lustre_handle    plock;
  61         struct md_op_data       *op_data;
  62         struct lmv_tgt_desc     *tgt;
  63         struct mdt_body         *body;
  64         int                     pmode;
  65         int                     rc = 0;
  66
  67         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
  68         if (!body)
  69                 return -EPROTO;
  70
  71         LASSERT((body->mbo_valid & OBD_MD_MDS));
  72
  73         /*
  74          * Unfortunately, we have to lie to MDC/MDS to retrieve
  75          * attributes llite needs and provideproper locking.
  76          */
  77         if (it->it_op & IT_LOOKUP)
  78                 it->it_op = IT_GETATTR;
  79
  80         /*
  81          * We got LOOKUP lock, but we really need attrs.
  82          */
  83         pmode = it->it_lock_mode;
  84         if (pmode) {
  85                 plock.cookie = it->it_lock_handle;
  86                 it->it_lock_mode = 0;
  87                 it->it_request = NULL;
  88         }
  89
  90         LASSERT(fid_is_sane(&body->mbo_fid1));
  91
  92         tgt = lmv_find_target(lmv, &body->mbo_fid1);
  93         if (IS_ERR(tgt)) {
  94                 rc = PTR_ERR(tgt);
  95                 goto out;
  96         }
  97
  98         op_data = kzalloc(sizeof(*op_data), GFP_NOFS);
  99         if (!op_data) {
 100                 rc = -ENOMEM;
 101                 goto out;
 102         }
 103
 104         op_data->op_fid1 = body->mbo_fid1;
 105         /* Sent the parent FID to the remote MDT */
 106         if (parent_fid) {
 107                 /* The parent fid is only for remote open to
 108                  * check whether the open is from OBF,
 109                  * see mdt_cross_open
 110                  */
 111                 LASSERT(it->it_op & IT_OPEN);
 112                 op_data->op_fid2 = *parent_fid;
 113         }
 114
 115         op_data->op_bias = MDS_CROSS_REF;
 116         CDEBUG(D_INODE, "REMOTE_INTENT with fid=" DFID " -> mds #%u\n",
 117                PFID(&body->mbo_fid1), tgt->ltd_idx);
 118
 119         rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
 120                             extra_lock_flags);
 121         if (rc)
 122                 goto out_free_op_data;
 123
 124         /*
 125          * LLite needs LOOKUP lock to track dentry revocation in order to
 126          * maintain dcache consistency. Thus drop UPDATE|PERM lock here
 127          * and put LOOKUP in request.
 128          */
 129         if (it->it_lock_mode != 0) {
 130                 it->it_remote_lock_handle =
 131                                         it->it_lock_handle;
 132                 it->it_remote_lock_mode = it->it_lock_mode;
 133         }
 134
 135         if (pmode) {
 136                 it->it_lock_handle = plock.cookie;
 137                 it->it_lock_mode = pmode;
 138         }
 139
 140 out_free_op_data:
 141         kfree(op_data);
 142 out:
 143         if (rc && pmode)
 144                 ldlm_lock_decref(&plock, pmode);
 145
 146         ptlrpc_req_finished(*reqp);
 147         *reqp = req;
 148         return rc;
 149 }
 150
 151 int lmv_revalidate_slaves(struct obd_export *exp,
 152                           const struct lmv_stripe_md *lsm,
 153                           ldlm_blocking_callback cb_blocking,
 154                           int extra_lock_flags)
 155 {
 156         struct obd_device *obd = exp->exp_obd;
 157         struct lmv_obd *lmv = &obd->u.lmv;
 158         struct ptlrpc_request *req = NULL;
 159         struct mdt_body *body;
 160         struct md_op_data *op_data;
 161         int rc = 0, i;
 162
 163         /**
 164          * revalidate slaves has some problems, temporarily return,
 165          * we may not need that
 166          */
 167         op_data = kzalloc(sizeof(*op_data), GFP_NOFS);
 168         if (!op_data)
 169                 return -ENOMEM;
 170
 171         /**
 172          * Loop over the stripe information, check validity and update them
 173          * from MDS if needed.
 174          */
 175         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
 176                 struct lookup_intent it = { .it_op = IT_GETATTR };
 177                 struct lustre_handle *lockh = NULL;
 178                 struct lmv_tgt_desc *tgt = NULL;
 179                 struct inode *inode;
 180                 struct lu_fid fid;
 181
 182                 fid = lsm->lsm_md_oinfo[i].lmo_fid;
 183                 inode = lsm->lsm_md_oinfo[i].lmo_root;
 184
 185                 /*
 186                  * Prepare op_data for revalidating. Note that @fid2 shluld be
 187                  * defined otherwise it will go to server and take new lock
 188                  * which is not needed here.
 189                  */
 190                 memset(op_data, 0, sizeof(*op_data));
 191                 op_data->op_fid1 = fid;
 192                 op_data->op_fid2 = fid;
 193
 194                 tgt = lmv_locate_mds(lmv, op_data, &fid);
 195                 if (IS_ERR(tgt)) {
 196                         rc = PTR_ERR(tgt);
 197                         goto cleanup;
 198                 }
 199
 200                 CDEBUG(D_INODE, "Revalidate slave " DFID " -> mds #%u\n",
 201                        PFID(&fid), tgt->ltd_idx);
 202
 203                 if (req) {
 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;
 212
 213                 lockh = (struct lustre_handle *)&it.it_lock_handle;
 214                 if (rc > 0 && !req) {
 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                         LASSERT(body);
 224
 225                         if (unlikely(body->mbo_nlink < 2)) {
 226                                 CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID":" DFID"\n",
 227                                        obd->obd_name, body->mbo_nlink, i,
 228                                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
 229                                        PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
 230
 231                                 if (it.it_lock_mode && lockh) {
 232                                         ldlm_lock_decref(lockh, it.it_lock_mode);
 233                                         it.it_lock_mode = 0;
 234                                 }
 235
 236                                 rc = -EIO;
 237                                 goto cleanup;
 238                         }
 239
 240                         i_size_write(inode, body->mbo_size);
 241                         inode->i_blocks = body->mbo_blocks;
 242                         set_nlink(inode, body->mbo_nlink);
 243                         LTIME_S(inode->i_atime) = body->mbo_atime;
 244                         LTIME_S(inode->i_ctime) = body->mbo_ctime;
 245                         LTIME_S(inode->i_mtime) = body->mbo_mtime;
 246                 }
 247
 248                 md_set_lock_data(tgt->ltd_exp, lockh, inode, NULL);
 249
 250                 if (it.it_lock_mode && lockh) {
 251                         ldlm_lock_decref(lockh, it.it_lock_mode);
 252                         it.it_lock_mode = 0;
 253                 }
 254         }
 255
 256 cleanup:
 257         if (req)
 258                 ptlrpc_req_finished(req);
 259
 260         kfree(op_data);
 261         return rc;
 262 }
 263
 264 /*
 265  * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
 266  * may be split dir.
 267  */
 268 static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
 269                            struct lookup_intent *it,
 270                            struct ptlrpc_request **reqp,
 271                            ldlm_blocking_callback cb_blocking,
 272                            __u64 extra_lock_flags)
 273 {
 274         struct obd_device       *obd = exp->exp_obd;
 275         struct lmv_obd          *lmv = &obd->u.lmv;
 276         struct lmv_tgt_desc     *tgt;
 277         struct mdt_body         *body;
 278         int                     rc;
 279
 280         if (it->it_flags & MDS_OPEN_BY_FID) {
 281                 LASSERT(fid_is_sane(&op_data->op_fid2));
 282
 283                 /*
 284                  * for striped directory, we can't know parent stripe fid
 285                  * without name, but we can set it to child fid, and MDT
 286                  * will obtain it from linkea in open in such case.
 287                  */
 288                 if (op_data->op_mea1)
 289                         op_data->op_fid1 = op_data->op_fid2;
 290
 291                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
 292                 if (IS_ERR(tgt))
 293                         return PTR_ERR(tgt);
 294
 295                 op_data->op_mds = tgt->ltd_idx;
 296         } else {
 297                 LASSERT(fid_is_sane(&op_data->op_fid1));
 298                 LASSERT(fid_is_zero(&op_data->op_fid2));
 299                 LASSERT(op_data->op_name);
 300
 301                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 302                 if (IS_ERR(tgt))
 303                         return PTR_ERR(tgt);
 304         }
 305
 306         /* If it is ready to open the file by FID, do not need
 307          * allocate FID at all, otherwise it will confuse MDT
 308          */
 309         if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
 310                 /*
 311                  * For lookup(IT_CREATE) cases allocate new fid and setup FLD
 312                  * for it.
 313                  */
 314                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 315                 if (rc != 0)
 316                         return rc;
 317         }
 318
 319         CDEBUG(D_INODE, "OPEN_INTENT with fid1=" DFID ", fid2=" DFID ", name='%s' -> mds #%u\n",
 320                PFID(&op_data->op_fid1),
 321                PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
 322
 323         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 324                             extra_lock_flags);
 325         if (rc != 0)
 326                 return rc;
 327         /*
 328          * Nothing is found, do not access body->mbo_fid1 as it is zero and thus
 329          * pointless.
 330          */
 331         if ((it->it_disposition & DISP_LOOKUP_NEG) &&
 332             !(it->it_disposition & DISP_OPEN_CREATE) &&
 333             !(it->it_disposition & DISP_OPEN_OPEN))
 334                 return rc;
 335
 336         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 337         if (!body)
 338                 return -EPROTO;
 339
 340         /* Not cross-ref case, just get out of here. */
 341         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 342                 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
 343                                        cb_blocking, extra_lock_flags);
 344                 if (rc != 0)
 345                         return rc;
 346
 347                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 348                 if (!body)
 349                         return -EPROTO;
 350         }
 351
 352         return rc;
 353 }
 354
 355 /*
 356  * Handler for: getattr, lookup and revalidate cases.
 357  */
 358 static int lmv_intent_lookup(struct obd_export *exp,
 359                              struct md_op_data *op_data,
 360                              struct lookup_intent *it,
 361                              struct ptlrpc_request **reqp,
 362                              ldlm_blocking_callback cb_blocking,
 363                              __u64 extra_lock_flags)
 364 {
 365         struct lmv_stripe_md *lsm = op_data->op_mea1;
 366         struct obd_device      *obd = exp->exp_obd;
 367         struct lmv_obd   *lmv = &obd->u.lmv;
 368         struct lmv_tgt_desc    *tgt = NULL;
 369         struct mdt_body *body;
 370         int                  rc = 0;
 371
 372         /*
 373          * If it returns ERR_PTR(-EBADFD) then it is an unknown hash type
 374          * it will try all stripes to locate the object
 375          */
 376         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
 377         if (IS_ERR(tgt) && (PTR_ERR(tgt) != -EBADFD))
 378                 return PTR_ERR(tgt);
 379
 380         /*
 381          * Both migrating dir and unknown hash dir need to try
 382          * all of sub-stripes
 383          */
 384         if (lsm && !lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
 385                 struct lmv_oinfo *oinfo = &lsm->lsm_md_oinfo[0];
 386
 387                 op_data->op_fid1 = oinfo->lmo_fid;
 388                 op_data->op_mds = oinfo->lmo_mds;
 389                 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
 390                 if (IS_ERR(tgt))
 391                         return PTR_ERR(tgt);
 392         }
 393
 394         if (!fid_is_sane(&op_data->op_fid2))
 395                 fid_zero(&op_data->op_fid2);
 396
 397         CDEBUG(D_INODE, "LOOKUP_INTENT with fid1=" DFID ", fid2=" DFID ", name='%s' -> mds #%u lsm=%p lsm_magic=%x\n",
 398                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
 399                op_data->op_name ? op_data->op_name : "<NULL>",
 400                tgt->ltd_idx, lsm, !lsm ? -1 : lsm->lsm_md_magic);
 401
 402         op_data->op_bias &= ~MDS_CROSS_REF;
 403
 404         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
 405                             extra_lock_flags);
 406         if (rc < 0)
 407                 return rc;
 408
 409         if (!*reqp) {
 410                 /*
 411                  * If RPC happens, lsm information will be revalidated
 412                  * during update_inode process (see ll_update_lsm_md)
 413                  */
 414                 if (op_data->op_mea2) {
 415                         rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
 416                                                    cb_blocking,
 417                                                    extra_lock_flags);
 418                         if (rc != 0)
 419                                 return rc;
 420                 }
 421                 return rc;
 422         } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm &&
 423                    lmv_need_try_all_stripes(lsm)) {
 424                 /*
 425                  * For migrating and unknown hash type directory, it will
 426                  * try to target the entry on other stripes
 427                  */
 428                 int stripe_index;
 429
 430                 for (stripe_index = 1;
 431                      stripe_index < lsm->lsm_md_stripe_count &&
 432                      it_disposition(it, DISP_LOOKUP_NEG); stripe_index++) {
 433                         struct lmv_oinfo *oinfo;
 434
 435                         /* release the previous request */
 436                         ptlrpc_req_finished(*reqp);
 437                         it->it_request = NULL;
 438                         *reqp = NULL;
 439
 440                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
 441                         tgt = lmv_find_target(lmv, &oinfo->lmo_fid);
 442                         if (IS_ERR(tgt))
 443                                 return PTR_ERR(tgt);
 444
 445                         CDEBUG(D_INODE, "Try other stripes " DFID"\n",
 446                                PFID(&oinfo->lmo_fid));
 447
 448                         op_data->op_fid1 = oinfo->lmo_fid;
 449                         it->it_disposition &= ~DISP_ENQ_COMPLETE;
 450                         rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
 451                                             cb_blocking, extra_lock_flags);
 452                         if (rc)
 453                                 return rc;
 454                 }
 455         }
 456
 457         /*
 458          * MDS has returned success. Probably name has been resolved in
 459          * remote inode. Let's check this.
 460          */
 461         body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 462         if (!body)
 463                 return -EPROTO;
 464
 465         /* Not cross-ref case, just get out of here. */
 466         if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
 467                 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
 468                                        extra_lock_flags);
 469                 if (rc != 0)
 470                         return rc;
 471                 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
 472                 if (!body)
 473                         return -EPROTO;
 474         }
 475
 476         return rc;
 477 }
 478
 479 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 480                     struct lookup_intent *it, struct ptlrpc_request **reqp,
 481                     ldlm_blocking_callback cb_blocking,
 482                     __u64 extra_lock_flags)
 483 {
 484         struct obd_device *obd = exp->exp_obd;
 485         int             rc;
 486
 487         LASSERT(fid_is_sane(&op_data->op_fid1));
 488
 489         CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%*s' on "DFID"\n",
 490                LL_IT2STR(it), PFID(&op_data->op_fid2),
 491                (int)op_data->op_namelen, op_data->op_name,
 492                PFID(&op_data->op_fid1));
 493
 494         rc = lmv_check_connect(obd);
 495         if (rc)
 496                 return rc;
 497
 498         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
 499                 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
 500                                        extra_lock_flags);
 501         else if (it->it_op & IT_OPEN)
 502                 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
 503                                      extra_lock_flags);
 504         else
 505                 LBUG();
 506
 507         if (rc < 0) {
 508                 struct lustre_handle lock_handle;
 509
 510                 if (it->it_lock_mode) {
 511                         lock_handle.cookie = it->it_lock_handle;
 512                         ldlm_lock_decref(&lock_handle, it->it_lock_mode);
 513                 }
 514
 515                 it->it_lock_handle = 0;
 516                 it->it_lock_mode = 0;
 517
 518                 if (it->it_remote_lock_mode) {
 519                         lock_handle.cookie = it->it_remote_lock_handle;
 520                         ldlm_lock_decref(&lock_handle,
 521                                          it->it_remote_lock_mode);
 522                 }
 523
 524                 it->it_remote_lock_handle = 0;
 525                 it->it_remote_lock_mode = 0;
 526         }
 527
 528         return rc;
 529 }