git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/dirent.c

   1
   2 #include "bcachefs.h"
   3 #include "bkey_methods.h"
   4 #include "btree_update.h"
   5 #include "extents.h"
   6 #include "dirent.h"
   7 #include "fs.h"
   8 #include "keylist.h"
   9 #include "str_hash.h"
  10
  11 #include <linux/dcache.h>
  12
  13 unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent d)
  14 {
  15         unsigned len = bkey_val_bytes(d.k) - sizeof(struct bch_dirent);
  16
  17         while (len && !d.v->d_name[len - 1])
  18                 --len;
  19
  20         return len;
  21 }
  22
  23 static unsigned dirent_val_u64s(unsigned len)
  24 {
  25         return DIV_ROUND_UP(sizeof(struct bch_dirent) + len, sizeof(u64));
  26 }
  27
  28 static u64 bch2_dirent_hash(const struct bch_hash_info *info,
  29                             const struct qstr *name)
  30 {
  31         struct bch_str_hash_ctx ctx;
  32
  33         bch2_str_hash_init(&ctx, info);
  34         bch2_str_hash_update(&ctx, info, name->name, name->len);
  35
  36         /* [0,2) reserved for dots */
  37         return max_t(u64, bch2_str_hash_end(&ctx, info), 2);
  38 }
  39
  40 static u64 dirent_hash_key(const struct bch_hash_info *info, const void *key)
  41 {
  42         return bch2_dirent_hash(info, key);
  43 }
  44
  45 static u64 dirent_hash_bkey(const struct bch_hash_info *info, struct bkey_s_c k)
  46 {
  47         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
  48         struct qstr name = QSTR_INIT(d.v->d_name, bch2_dirent_name_bytes(d));
  49
  50         return bch2_dirent_hash(info, &name);
  51 }
  52
  53 static bool dirent_cmp_key(struct bkey_s_c _l, const void *_r)
  54 {
  55         struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
  56         int len = bch2_dirent_name_bytes(l);
  57         const struct qstr *r = _r;
  58
  59         return len - r->len ?: memcmp(l.v->d_name, r->name, len);
  60 }
  61
  62 static bool dirent_cmp_bkey(struct bkey_s_c _l, struct bkey_s_c _r)
  63 {
  64         struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
  65         struct bkey_s_c_dirent r = bkey_s_c_to_dirent(_r);
  66         int l_len = bch2_dirent_name_bytes(l);
  67         int r_len = bch2_dirent_name_bytes(r);
  68
  69         return l_len - r_len ?: memcmp(l.v->d_name, r.v->d_name, l_len);
  70 }
  71
  72 const struct bch_hash_desc bch2_dirent_hash_desc = {
  73         .btree_id       = BTREE_ID_DIRENTS,
  74         .key_type       = BCH_DIRENT,
  75         .whiteout_type  = BCH_DIRENT_WHITEOUT,
  76         .hash_key       = dirent_hash_key,
  77         .hash_bkey      = dirent_hash_bkey,
  78         .cmp_key        = dirent_cmp_key,
  79         .cmp_bkey       = dirent_cmp_bkey,
  80 };
  81
  82 static const char *bch2_dirent_invalid(const struct bch_fs *c,
  83                                        struct bkey_s_c k)
  84 {
  85         struct bkey_s_c_dirent d;
  86         unsigned len;
  87
  88         switch (k.k->type) {
  89         case BCH_DIRENT:
  90                 if (bkey_val_bytes(k.k) < sizeof(struct bch_dirent))
  91                         return "value too small";
  92
  93                 d = bkey_s_c_to_dirent(k);
  94                 len = bch2_dirent_name_bytes(d);
  95
  96                 if (!len)
  97                         return "empty name";
  98
  99                 if (bkey_val_u64s(k.k) > dirent_val_u64s(len))
 100                         return "value too big";
 101
 102                 if (len > NAME_MAX)
 103                         return "dirent name too big";
 104
 105                 if (memchr(d.v->d_name, '/', len))
 106                         return "dirent name has invalid characters";
 107
 108                 return NULL;
 109         case BCH_DIRENT_WHITEOUT:
 110                 return bkey_val_bytes(k.k) != 0
 111                         ? "value size should be zero"
 112                         : NULL;
 113
 114         default:
 115                 return "invalid type";
 116         }
 117 }
 118
 119 static void bch2_dirent_to_text(struct bch_fs *c, char *buf,
 120                                 size_t size, struct bkey_s_c k)
 121 {
 122         struct bkey_s_c_dirent d;
 123         size_t n = 0;
 124
 125         switch (k.k->type) {
 126         case BCH_DIRENT:
 127                 d = bkey_s_c_to_dirent(k);
 128
 129                 n += bch_scnmemcpy(buf + n, size - n, d.v->d_name,
 130                                    bch2_dirent_name_bytes(d));
 131                 n += scnprintf(buf + n, size - n, " -> %llu", d.v->d_inum);
 132                 break;
 133         case BCH_DIRENT_WHITEOUT:
 134                 scnprintf(buf, size, "whiteout");
 135                 break;
 136         }
 137 }
 138
 139 const struct bkey_ops bch2_bkey_dirent_ops = {
 140         .key_invalid    = bch2_dirent_invalid,
 141         .val_to_text    = bch2_dirent_to_text,
 142 };
 143
 144 static struct bkey_i_dirent *dirent_create_key(u8 type,
 145                                 const struct qstr *name, u64 dst)
 146 {
 147         struct bkey_i_dirent *dirent;
 148         unsigned u64s = BKEY_U64s + dirent_val_u64s(name->len);
 149
 150         dirent = kmalloc(u64s * sizeof(u64), GFP_NOFS);
 151         if (!dirent)
 152                 return NULL;
 153
 154         bkey_dirent_init(&dirent->k_i);
 155         dirent->k.u64s = u64s;
 156         dirent->v.d_inum = cpu_to_le64(dst);
 157         dirent->v.d_type = type;
 158
 159         memcpy(dirent->v.d_name, name->name, name->len);
 160         memset(dirent->v.d_name + name->len, 0,
 161                bkey_val_bytes(&dirent->k) -
 162                (sizeof(struct bch_dirent) + name->len));
 163
 164         EBUG_ON(bch2_dirent_name_bytes(dirent_i_to_s_c(dirent)) != name->len);
 165
 166         return dirent;
 167 }
 168
 169 int bch2_dirent_create(struct bch_fs *c, u64 dir_inum,
 170                        const struct bch_hash_info *hash_info,
 171                        u8 type, const struct qstr *name, u64 dst_inum,
 172                        u64 *journal_seq, int flags)
 173 {
 174         struct bkey_i_dirent *dirent;
 175         int ret;
 176
 177         dirent = dirent_create_key(type, name, dst_inum);
 178         if (!dirent)
 179                 return -ENOMEM;
 180
 181         ret = bch2_hash_set(bch2_dirent_hash_desc, hash_info, c, dir_inum,
 182                            journal_seq, &dirent->k_i, flags);
 183         kfree(dirent);
 184
 185         return ret;
 186 }
 187
 188 static void dirent_copy_target(struct bkey_i_dirent *dst,
 189                                struct bkey_s_c_dirent src)
 190 {
 191         dst->v.d_inum = src.v->d_inum;
 192         dst->v.d_type = src.v->d_type;
 193 }
 194
 195 static struct bpos bch2_dirent_pos(struct bch_inode_info *inode,
 196                                    const struct qstr *name)
 197 {
 198         return POS(inode->v.i_ino, bch2_dirent_hash(&inode->ei_str_hash, name));
 199 }
 200
 201 int bch2_dirent_rename(struct bch_fs *c,
 202                 struct bch_inode_info *src_dir, const struct qstr *src_name,
 203                 struct bch_inode_info *dst_dir, const struct qstr *dst_name,
 204                 u64 *journal_seq, enum bch_rename_mode mode)
 205 {
 206         struct btree_iter src_iter, dst_iter, whiteout_iter;
 207         struct bkey_s_c old_src, old_dst;
 208         struct bkey delete;
 209         struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
 210         struct bpos src_pos = bch2_dirent_pos(src_dir, src_name);
 211         struct bpos dst_pos = bch2_dirent_pos(dst_dir, dst_name);
 212         bool need_whiteout;
 213         int ret = -ENOMEM;
 214
 215         bch2_btree_iter_init(&src_iter, c, BTREE_ID_DIRENTS, src_pos,
 216                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 217         bch2_btree_iter_init(&dst_iter, c, BTREE_ID_DIRENTS, dst_pos,
 218                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 219         bch2_btree_iter_link(&src_iter, &dst_iter);
 220
 221         bch2_btree_iter_init(&whiteout_iter, c, BTREE_ID_DIRENTS, src_pos,
 222                              BTREE_ITER_SLOTS);
 223         bch2_btree_iter_link(&src_iter, &whiteout_iter);
 224
 225         if (mode == BCH_RENAME_EXCHANGE) {
 226                 new_src = dirent_create_key(0, src_name, 0);
 227                 if (!new_src)
 228                         goto err;
 229         } else {
 230                 new_src = (void *) &delete;
 231         }
 232
 233         new_dst = dirent_create_key(0, dst_name, 0);
 234         if (!new_dst)
 235                 goto err;
 236 retry:
 237         /*
 238          * Note that on -EINTR/dropped locks we're not restarting the lookup
 239          * from the original hashed position (like we do when creating dirents,
 240          * in bch_hash_set) -  we never move existing dirents to different slot:
 241          */
 242         old_src = bch2_hash_lookup_at(bch2_dirent_hash_desc,
 243                                      &src_dir->ei_str_hash,
 244                                      &src_iter, src_name);
 245         if ((ret = btree_iter_err(old_src)))
 246                 goto err;
 247
 248         ret = bch2_hash_needs_whiteout(bch2_dirent_hash_desc,
 249                                 &src_dir->ei_str_hash,
 250                                 &whiteout_iter, &src_iter);
 251         if (ret < 0)
 252                 goto err;
 253         need_whiteout = ret;
 254
 255         /*
 256          * Note that in BCH_RENAME mode, we're _not_ checking if
 257          * the target already exists - we're relying on the VFS
 258          * to do that check for us for correctness:
 259          */
 260         old_dst = mode == BCH_RENAME
 261                 ? bch2_hash_hole_at(bch2_dirent_hash_desc, &dst_iter)
 262                 : bch2_hash_lookup_at(bch2_dirent_hash_desc,
 263                                      &dst_dir->ei_str_hash,
 264                                      &dst_iter, dst_name);
 265         if ((ret = btree_iter_err(old_dst)))
 266                 goto err;
 267
 268         switch (mode) {
 269         case BCH_RENAME:
 270                 bkey_init(&new_src->k);
 271                 dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 272
 273                 if (bkey_cmp(dst_pos, src_iter.pos) <= 0 &&
 274                     bkey_cmp(src_iter.pos, dst_iter.pos) < 0) {
 275                         /*
 276                          * If we couldn't insert new_dst at its hashed
 277                          * position (dst_pos) due to a hash collision,
 278                          * and we're going to be deleting in
 279                          * between the hashed position and first empty
 280                          * slot we found - just overwrite the pos we
 281                          * were going to delete:
 282                          *
 283                          * Note: this is a correctness issue, in this
 284                          * situation bch2_hash_needs_whiteout() could
 285                          * return false when the whiteout would have
 286                          * been needed if we inserted at the pos
 287                          * __dirent_find_hole() found
 288                          */
 289                         new_dst->k.p = src_iter.pos;
 290                         ret = bch2_btree_insert_at(c, NULL, NULL,
 291                                         journal_seq,
 292                                         BTREE_INSERT_ATOMIC,
 293                                         BTREE_INSERT_ENTRY(&src_iter,
 294                                                            &new_dst->k_i));
 295                         goto err;
 296                 }
 297
 298                 if (need_whiteout)
 299                         new_src->k.type = BCH_DIRENT_WHITEOUT;
 300                 break;
 301         case BCH_RENAME_OVERWRITE:
 302                 bkey_init(&new_src->k);
 303                 dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 304
 305                 if (bkey_cmp(dst_pos, src_iter.pos) <= 0 &&
 306                     bkey_cmp(src_iter.pos, dst_iter.pos) < 0) {
 307                         /*
 308                          * Same case described above -
 309                          * bch_hash_needs_whiteout could spuriously
 310                          * return false, but we have to insert at
 311                          * dst_iter.pos because we're overwriting
 312                          * another dirent:
 313                          */
 314                         new_src->k.type = BCH_DIRENT_WHITEOUT;
 315                 } else if (need_whiteout)
 316                         new_src->k.type = BCH_DIRENT_WHITEOUT;
 317                 break;
 318         case BCH_RENAME_EXCHANGE:
 319                 dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
 320                 dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 321                 break;
 322         }
 323
 324         new_src->k.p = src_iter.pos;
 325         new_dst->k.p = dst_iter.pos;
 326         ret = bch2_btree_insert_at(c, NULL, NULL, journal_seq,
 327                         BTREE_INSERT_ATOMIC,
 328                         BTREE_INSERT_ENTRY(&src_iter, &new_src->k_i),
 329                         BTREE_INSERT_ENTRY(&dst_iter, &new_dst->k_i));
 330 err:
 331         if (ret == -EINTR)
 332                 goto retry;
 333
 334         bch2_btree_iter_unlock(&whiteout_iter);
 335         bch2_btree_iter_unlock(&dst_iter);
 336         bch2_btree_iter_unlock(&src_iter);
 337
 338         if (new_src != (void *) &delete)
 339                 kfree(new_src);
 340         kfree(new_dst);
 341         return ret;
 342 }
 343
 344 int bch2_dirent_delete(struct bch_fs *c, u64 dir_inum,
 345                        const struct bch_hash_info *hash_info,
 346                        const struct qstr *name,
 347                        u64 *journal_seq)
 348 {
 349         return bch2_hash_delete(bch2_dirent_hash_desc, hash_info,
 350                                c, dir_inum, journal_seq, name);
 351 }
 352
 353 u64 bch2_dirent_lookup(struct bch_fs *c, u64 dir_inum,
 354                        const struct bch_hash_info *hash_info,
 355                        const struct qstr *name)
 356 {
 357         struct btree_iter iter;
 358         struct bkey_s_c k;
 359         u64 inum;
 360
 361         k = bch2_hash_lookup(bch2_dirent_hash_desc, hash_info, c,
 362                             dir_inum, &iter, name);
 363         if (IS_ERR(k.k)) {
 364                 bch2_btree_iter_unlock(&iter);
 365                 return 0;
 366         }
 367
 368         inum = le64_to_cpu(bkey_s_c_to_dirent(k).v->d_inum);
 369         bch2_btree_iter_unlock(&iter);
 370
 371         return inum;
 372 }
 373
 374 int bch2_empty_dir(struct bch_fs *c, u64 dir_inum)
 375 {
 376         struct btree_iter iter;
 377         struct bkey_s_c k;
 378         int ret = 0;
 379
 380         for_each_btree_key(&iter, c, BTREE_ID_DIRENTS, POS(dir_inum, 0), 0, k) {
 381                 if (k.k->p.inode > dir_inum)
 382                         break;
 383
 384                 if (k.k->type == BCH_DIRENT) {
 385                         ret = -ENOTEMPTY;
 386                         break;
 387                 }
 388         }
 389         bch2_btree_iter_unlock(&iter);
 390
 391         return ret;
 392 }
 393
 394 int bch2_readdir(struct bch_fs *c, struct file *file,
 395                  struct dir_context *ctx)
 396 {
 397         struct bch_inode_info *inode = file_bch_inode(file);
 398         struct btree_iter iter;
 399         struct bkey_s_c k;
 400         struct bkey_s_c_dirent dirent;
 401         unsigned len;
 402
 403         if (!dir_emit_dots(file, ctx))
 404                 return 0;
 405
 406         for_each_btree_key(&iter, c, BTREE_ID_DIRENTS,
 407                            POS(inode->v.i_ino, ctx->pos), 0, k) {
 408                 if (k.k->type != BCH_DIRENT)
 409                         continue;
 410
 411                 dirent = bkey_s_c_to_dirent(k);
 412
 413                 if (bkey_cmp(k.k->p, POS(inode->v.i_ino, ctx->pos)) < 0)
 414                         continue;
 415
 416                 if (k.k->p.inode > inode->v.i_ino)
 417                         break;
 418
 419                 len = bch2_dirent_name_bytes(dirent);
 420
 421                 /*
 422                  * XXX: dir_emit() can fault and block, while we're holding
 423                  * locks
 424                  */
 425                 if (!dir_emit(ctx, dirent.v->d_name, len,
 426                               le64_to_cpu(dirent.v->d_inum),
 427                               dirent.v->d_type))
 428                         break;
 429
 430                 ctx->pos = k.k->p.offset + 1;
 431         }
 432         bch2_btree_iter_unlock(&iter);
 433
 434         return 0;
 435 }