lustre/liblustre/dir.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/liblustre/dir.c
  39  *
  40  * Lustre Light directory handling
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_LLITE
  44
  45 #include <unistd.h>
  46 #include <stdlib.h>
  47 #include <string.h>
  48 #include <assert.h>
  49 #include <time.h>
  50 #include <sys/types.h>
  51 #include <sys/stat.h>
  52 #include <fcntl.h>
  53 #include <sys/queue.h>
  54
  55 #ifdef HAVE_LINUX_UNISTD_H
  56 #include <linux/unistd.h>
  57 #elif defined(HAVE_UNISTD_H)
  58 #include <unistd.h>
  59 #endif
  60
  61
  62 #include "llite_lib.h"
  63 #include <dirent.h>
  64
  65 /* (new) readdir implementation overview can be found in lustre/llite/dir.c */
  66
  67 static int llu_dir_do_readpage(struct inode *inode, struct page *page)
  68 {
  69         struct llu_inode_info *lli = llu_i2info(inode);
  70         struct intnl_stat     *st = llu_i2stat(inode);
  71         struct llu_sb_info    *sbi = llu_i2sbi(inode);
  72         struct ptlrpc_request *request;
  73         struct lustre_handle   lockh;
  74         struct mdt_body       *body;
  75         struct lookup_intent   it = { .it_op = IT_READDIR };
  76         struct md_op_data      op_data = {{ 0 }};
  77         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_UPDATE } };
  78         int rc = 0;
  79         ENTRY;
  80
  81         llu_prep_md_op_data(&op_data, inode, NULL, NULL, 0, 0, LUSTRE_OPC_ANY);
  82         rc = md_lock_match(sbi->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
  83                            &lli->lli_fid, LDLM_IBITS, &policy, LCK_CR, &lockh);
  84         if (!rc) {
  85                 struct ldlm_enqueue_info einfo = {LDLM_IBITS, LCK_CR,
  86                         llu_md_blocking_ast, ldlm_completion_ast, NULL, NULL,
  87                         inode};
  88
  89                 rc = md_enqueue(sbi->ll_md_exp, &einfo, &it,
  90                                 &op_data, &lockh, NULL, 0, NULL,
  91                                 LDLM_FL_CANCEL_ON_BLOCK);
  92                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
  93                 if (request)
  94                         ptlrpc_req_finished(request);
  95                 if (rc < 0) {
  96                         CERROR("lock enqueue: err: %d\n", rc);
  97                         RETURN(rc);
  98                 }
  99         }
 100         ldlm_lock_dump_handle(D_OTHER, &lockh);
 101
 102         op_data.op_offset = (__u64)hash_x_index(page->index, 0);
 103         op_data.op_npages = 1;
 104         rc = md_readpage(sbi->ll_md_exp, &op_data, &page, &request);
 105         if (!rc) {
 106                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 107                 LASSERT(body != NULL);         /* checked by md_readpage() */
 108
 109                 if (body->valid & OBD_MD_FLSIZE)
 110                         st->st_size = body->size;
 111         } else {
 112                 CERROR("read_dir_page(%ld) error %d\n", page->index, rc);
 113         }
 114         ptlrpc_req_finished(request);
 115         EXIT;
 116
 117         ldlm_lock_decref(&lockh, LCK_CR);
 118         return rc;
 119 }
 120
 121 static cfs_page_t *llu_dir_read_page(struct inode *ino, __u64 hash,
 122                                      int exact, struct ll_dir_chain *chain)
 123 {
 124         cfs_page_t *page;
 125         int rc;
 126         ENTRY;
 127
 128         OBD_PAGE_ALLOC(page, 0);
 129         if (!page)
 130                 RETURN(ERR_PTR(-ENOMEM));
 131         page->index = hash_x_index(hash, 0);
 132
 133         rc = llu_dir_do_readpage(ino, page);
 134         if (rc) {
 135                 OBD_PAGE_FREE(page);
 136                 RETURN(ERR_PTR(rc));
 137         }
 138
 139         return page;
 140 }
 141
 142 void *(*memmover)(void *, const void *, size_t) = memmove;
 143
 144 #define NAME_OFFSET(de) ((int) ((de)->d_name - (char *) (de)))
 145 #define ROUND_UP64(x)   (((x)+sizeof(__u64)-1) & ~(sizeof(__u64)-1))
 146 static int filldir(char *buf, int buflen,
 147                    const char *name, int namelen, loff_t offset,
 148                    ino_t ino, unsigned int d_type, int *filled)
 149 {
 150         cfs_dirent_t *dirent = (cfs_dirent_t *) (buf + *filled);
 151         cfs_dirent_t  holder;
 152         int reclen = ROUND_UP64(NAME_OFFSET(dirent) + namelen + 1);
 153
 154         /*
 155          * @buf is not guaranteed to be properly aligned. To work around,
 156          * first fill stack-allocated @holder, then copy @holder into @buf by
 157          * memmove().
 158          */
 159
 160         /* check overflow */
 161         if ((*filled + reclen) > buflen)
 162                 return 1;
 163
 164         holder.d_ino = ino;
 165 #ifdef _DIRENT_HAVE_D_OFF
 166         holder.d_off = offset;
 167 #endif
 168         holder.d_reclen = reclen;
 169 #ifdef _DIRENT_HAVE_D_TYPE
 170         holder.d_type = (unsigned short) d_type;
 171 #endif
 172         /* gcc unrolls memcpy() of structs into field-wise assignments,
 173          * assuming proper alignment. Humor it. */
 174         (*memmover)(dirent, &holder, NAME_OFFSET(dirent));
 175         memcpy(dirent->d_name, name, namelen);
 176         dirent->d_name[namelen] = 0;
 177
 178         *filled += reclen;
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * TODO: much of the code here is similar/identical to llite ll_readdir().
 185  * These code can be factored out and shared in a common module.
 186  */
 187
 188 ssize_t llu_iop_filldirentries(struct inode *dir, _SYSIO_OFF_T *basep,
 189                                char *buf, size_t nbytes)
 190 {
 191         struct llu_inode_info *lli = llu_i2info(dir);
 192         struct intnl_stat     *st = llu_i2stat(dir);
 193         loff_t                 pos = *basep;
 194         struct ll_dir_chain    chain;
 195         cfs_page_t            *page;
 196         int filled = 0;
 197         int rc;
 198         int done;
 199         __u16 type;
 200         ENTRY;
 201
 202         liblustre_wait_event(0);
 203
 204         if (st->st_size == 0) {
 205                 CWARN("dir size is 0?\n");
 206                 RETURN(0);
 207         }
 208
 209         if (pos == MDS_DIR_END_OFF)
 210                 /*
 211                  * end-of-file.
 212                  */
 213                 RETURN(0);
 214
 215         rc    = 0;
 216         done  = 0;
 217         ll_dir_chain_init(&chain);
 218
 219         page = llu_dir_read_page(dir, pos, 0, &chain);
 220         while (rc == 0 && !done) {
 221                 struct lu_dirpage *dp;
 222                 struct lu_dirent  *ent;
 223
 224                 if (!IS_ERR(page)) {
 225                         /*
 226                          * If page is empty (end of directoryis reached),
 227                          * use this value.
 228                          */
 229                         __u64 hash = MDS_DIR_END_OFF;
 230                         __u64 next;
 231
 232                         dp = page->addr;
 233                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 234                              ent = lu_dirent_next(ent)) {
 235                                 char          *name;
 236                                 int            namelen;
 237                                 struct lu_fid  fid;
 238                                 __u64          ino;
 239
 240                                 hash    = le64_to_cpu(ent->lde_hash);
 241                                 namelen = le16_to_cpu(ent->lde_namelen);
 242
 243                                 if (hash < pos)
 244                                         /*
 245                                          * Skip until we find target hash
 246                                          * value.
 247                                          */
 248                                         continue;
 249
 250                                 if (namelen == 0)
 251                                         /*
 252                                          * Skip dummy record.
 253                                          */
 254                                         continue;
 255
 256                                 fid  = ent->lde_fid;
 257                                 name = ent->lde_name;
 258                                 fid_le_to_cpu(&fid, &fid);
 259                                 ino  = cl_fid_build_ino(&fid, 0);
 260                                 type = ll_dirent_type_get(ent);
 261                                 done = filldir(buf, nbytes, name, namelen,
 262                                                (loff_t)hash, ino, type,
 263                                                &filled);
 264                         }
 265                         next = le64_to_cpu(dp->ldp_hash_end);
 266                         OBD_PAGE_FREE(page);
 267                         if (!done) {
 268                                 pos = next;
 269                                 if (pos == MDS_DIR_END_OFF)
 270                                         /*
 271                                          * End of directory reached.
 272                                          */
 273                                         done = 1;
 274                                 else if (1 /* chain is exhausted*/)
 275                                         /*
 276                                          * Normal case: continue to the next
 277                                          * page.
 278                                          */
 279                                         page = llu_dir_read_page(dir, pos, 1,
 280                                                                &chain);
 281                                 else {
 282                                         /*
 283                                          * go into overflow page.
 284                                          */
 285                                 }
 286                         } else {
 287                                 pos = hash;
 288                                 if (filled == 0)
 289                                         GOTO(out, filled = -EINVAL);
 290                         }
 291                 } else {
 292                         rc = PTR_ERR(page);
 293                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 294                                PFID(&lli->lli_fid), (unsigned long)pos, rc);
 295                 }
 296         }
 297         lli->lli_dir_pos = (loff_t)pos;
 298         *basep = lli->lli_dir_pos;
 299 out:
 300         ll_dir_chain_fini(&chain);
 301         liblustre_wait_event(0);
 302         RETURN(filled);
 303 }