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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "bkey_buf.h"
   4 #include "btree_update.h"
   5 #include "buckets.h"
   6 #include "extents.h"
   7 #include "inode.h"
   8 #include "io.h"
   9 #include "reflink.h"
  10 #include "subvolume.h"
  11
  12 #include <linux/sched/signal.h>
  13
  14 static inline unsigned bkey_type_to_indirect(const struct bkey *k)
  15 {
  16         switch (k->type) {
  17         case KEY_TYPE_extent:
  18                 return KEY_TYPE_reflink_v;
  19         case KEY_TYPE_inline_data:
  20                 return KEY_TYPE_indirect_inline_data;
  21         default:
  22                 return 0;
  23         }
  24 }
  25
  26 /* reflink pointers */
  27
  28 int bch2_reflink_p_invalid(const struct bch_fs *c, struct bkey_s_c k,
  29                            int rw, struct printbuf *err)
  30 {
  31         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  32
  33         if (bkey_val_bytes(p.k) != sizeof(*p.v)) {
  34                 pr_buf(err, "incorrect value size (%zu != %zu)",
  35                        bkey_val_bytes(p.k), sizeof(*p.v));
  36                 return -EINVAL;
  37         }
  38
  39         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix &&
  40             le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad)) {
  41                 pr_buf(err, "idx < front_pad (%llu < %u)",
  42                        le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
  43                 return -EINVAL;
  44         }
  45
  46         return 0;
  47 }
  48
  49 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  50                             struct bkey_s_c k)
  51 {
  52         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  53
  54         pr_buf(out, "idx %llu front_pad %u back_pad %u",
  55                le64_to_cpu(p.v->idx),
  56                le32_to_cpu(p.v->front_pad),
  57                le32_to_cpu(p.v->back_pad));
  58 }
  59
  60 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  61 {
  62         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  63         struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
  64
  65         /*
  66          * Disabled for now, the triggers code needs to be reworked for merging
  67          * of reflink pointers to work:
  68          */
  69         return false;
  70
  71         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  72                 return false;
  73
  74         bch2_key_resize(l.k, l.k->size + r.k->size);
  75         return true;
  76 }
  77
  78 /* indirect extents */
  79
  80 int bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k,
  81                            int rw, struct printbuf *err)
  82 {
  83         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  84
  85         if (bkey_val_bytes(r.k) < sizeof(*r.v)) {
  86                 pr_buf(err, "incorrect value size (%zu < %zu)",
  87                        bkey_val_bytes(r.k), sizeof(*r.v));
  88                 return -EINVAL;
  89         }
  90
  91         return bch2_bkey_ptrs_invalid(c, k, rw, err);
  92 }
  93
  94 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  95                             struct bkey_s_c k)
  96 {
  97         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  98
  99         pr_buf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
 100
 101         bch2_bkey_ptrs_to_text(out, c, k);
 102 }
 103
 104 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
 105 {
 106         struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
 107         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
 108
 109         return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
 110 }
 111
 112 int bch2_trans_mark_reflink_v(struct btree_trans *trans,
 113                               struct bkey_s_c old, struct bkey_i *new,
 114                               unsigned flags)
 115 {
 116         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
 117                 struct bkey_i_reflink_v *r = bkey_i_to_reflink_v(new);
 118
 119                 if (!r->v.refcount) {
 120                         r->k.type = KEY_TYPE_deleted;
 121                         r->k.size = 0;
 122                         set_bkey_val_u64s(&r->k, 0);
 123                         return 0;
 124                 }
 125         }
 126
 127         return bch2_trans_mark_extent(trans, old, new, flags);
 128 }
 129
 130 /* indirect inline data */
 131
 132 int bch2_indirect_inline_data_invalid(const struct bch_fs *c, struct bkey_s_c k,
 133                                       int rw, struct printbuf *err)
 134 {
 135         if (bkey_val_bytes(k.k) < sizeof(struct bch_indirect_inline_data)) {
 136                 pr_buf(err, "incorrect value size (%zu < %zu)",
 137                        bkey_val_bytes(k.k), sizeof(struct bch_indirect_inline_data));
 138                 return -EINVAL;
 139         }
 140
 141         return 0;
 142 }
 143
 144 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 145                                         struct bch_fs *c, struct bkey_s_c k)
 146 {
 147         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 148         unsigned datalen = bkey_inline_data_bytes(k.k);
 149
 150         pr_buf(out, "refcount %llu datalen %u: %*phN",
 151                le64_to_cpu(d.v->refcount), datalen,
 152                min(datalen, 32U), d.v->data);
 153 }
 154
 155 int bch2_trans_mark_indirect_inline_data(struct btree_trans *trans,
 156                               struct bkey_s_c old, struct bkey_i *new,
 157                               unsigned flags)
 158 {
 159         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
 160                 struct bkey_i_indirect_inline_data *r =
 161                         bkey_i_to_indirect_inline_data(new);
 162
 163                 if (!r->v.refcount) {
 164                         r->k.type = KEY_TYPE_deleted;
 165                         r->k.size = 0;
 166                         set_bkey_val_u64s(&r->k, 0);
 167                 }
 168         }
 169
 170         return 0;
 171 }
 172
 173 static int bch2_make_extent_indirect(struct btree_trans *trans,
 174                                      struct btree_iter *extent_iter,
 175                                      struct bkey_i *orig)
 176 {
 177         struct bch_fs *c = trans->c;
 178         struct btree_iter reflink_iter = { NULL };
 179         struct bkey_s_c k;
 180         struct bkey_i *r_v;
 181         struct bkey_i_reflink_p *r_p;
 182         __le64 *refcount;
 183         int ret;
 184
 185         if (orig->k.type == KEY_TYPE_inline_data)
 186                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 187
 188         for_each_btree_key_norestart(trans, reflink_iter, BTREE_ID_reflink,
 189                            POS(0, c->reflink_hint),
 190                            BTREE_ITER_INTENT|BTREE_ITER_SLOTS, k, ret) {
 191                 if (reflink_iter.pos.inode) {
 192                         bch2_btree_iter_set_pos(&reflink_iter, POS_MIN);
 193                         continue;
 194                 }
 195
 196                 if (bkey_deleted(k.k) && orig->k.size <= k.k->size)
 197                         break;
 198         }
 199
 200         if (ret)
 201                 goto err;
 202
 203         /* rewind iter to start of hole, if necessary: */
 204         bch2_btree_iter_set_pos_to_extent_start(&reflink_iter);
 205
 206         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 207         ret = PTR_ERR_OR_ZERO(r_v);
 208         if (ret)
 209                 goto err;
 210
 211         bkey_init(&r_v->k);
 212         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 213         r_v->k.p        = reflink_iter.pos;
 214         bch2_key_resize(&r_v->k, orig->k.size);
 215         r_v->k.version  = orig->k.version;
 216
 217         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 218
 219         refcount        = bkey_refcount(r_v);
 220         *refcount       = 0;
 221         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 222
 223         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 224         if (ret)
 225                 goto err;
 226
 227         /*
 228          * orig is in a bkey_buf which statically allocates 5 64s for the val,
 229          * so we know it will be big enough:
 230          */
 231         orig->k.type = KEY_TYPE_reflink_p;
 232         r_p = bkey_i_to_reflink_p(orig);
 233         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 234         memset(&r_p->v, 0, sizeof(r_p->v));
 235
 236         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 237
 238         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 239                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 240 err:
 241         c->reflink_hint = reflink_iter.pos.offset;
 242         bch2_trans_iter_exit(trans, &reflink_iter);
 243
 244         return ret;
 245 }
 246
 247 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 248 {
 249         struct bkey_s_c k;
 250         int ret;
 251
 252         for_each_btree_key_continue_norestart(*iter, 0, k, ret) {
 253                 if (bkey_cmp(iter->pos, end) >= 0)
 254                         break;
 255
 256                 if (bkey_extent_is_data(k.k))
 257                         return k;
 258         }
 259
 260         if (bkey_cmp(iter->pos, end) >= 0)
 261                 bch2_btree_iter_set_pos(iter, end);
 262         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 263 }
 264
 265 s64 bch2_remap_range(struct bch_fs *c,
 266                      subvol_inum dst_inum, u64 dst_offset,
 267                      subvol_inum src_inum, u64 src_offset,
 268                      u64 remap_sectors,
 269                      u64 new_i_size, s64 *i_sectors_delta)
 270 {
 271         struct btree_trans trans;
 272         struct btree_iter dst_iter, src_iter;
 273         struct bkey_s_c src_k;
 274         struct bkey_buf new_dst, new_src;
 275         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 276         struct bpos src_start = POS(src_inum.inum, src_offset);
 277         struct bpos dst_end = dst_start, src_end = src_start;
 278         struct bpos src_want;
 279         u64 dst_done;
 280         u32 dst_snapshot, src_snapshot;
 281         int ret = 0, ret2 = 0;
 282
 283         if (!percpu_ref_tryget(&c->writes))
 284                 return -EROFS;
 285
 286         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 287
 288         dst_end.offset += remap_sectors;
 289         src_end.offset += remap_sectors;
 290
 291         bch2_bkey_buf_init(&new_dst);
 292         bch2_bkey_buf_init(&new_src);
 293         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 294
 295         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 296                              BTREE_ITER_INTENT);
 297         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 298                              BTREE_ITER_INTENT);
 299
 300         while ((ret == 0 || ret == -EINTR) &&
 301                bkey_cmp(dst_iter.pos, dst_end) < 0) {
 302                 struct disk_reservation disk_res = { 0 };
 303
 304                 bch2_trans_begin(&trans);
 305
 306                 if (fatal_signal_pending(current)) {
 307                         ret = -EINTR;
 308                         break;
 309                 }
 310
 311                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 312                                                   &src_snapshot);
 313                 if (ret)
 314                         continue;
 315
 316                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 317
 318                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 319                                                   &dst_snapshot);
 320                 if (ret)
 321                         continue;
 322
 323                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 324
 325                 dst_done = dst_iter.pos.offset - dst_start.offset;
 326                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 327                 bch2_btree_iter_set_pos(&src_iter, src_want);
 328
 329                 src_k = get_next_src(&src_iter, src_end);
 330                 ret = bkey_err(src_k);
 331                 if (ret)
 332                         continue;
 333
 334                 if (bkey_cmp(src_want, src_iter.pos) < 0) {
 335                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 336                                         min(dst_end.offset,
 337                                             dst_iter.pos.offset +
 338                                             src_iter.pos.offset - src_want.offset),
 339                                         i_sectors_delta);
 340                         continue;
 341                 }
 342
 343                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 344                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 345
 346                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 347                         src_k = bkey_i_to_s_c(new_src.k);
 348
 349                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 350                                                 new_src.k);
 351                         if (ret)
 352                                 continue;
 353
 354                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 355                 }
 356
 357                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 358                         struct bkey_s_c_reflink_p src_p =
 359                                 bkey_s_c_to_reflink_p(src_k);
 360                         struct bkey_i_reflink_p *dst_p =
 361                                 bkey_reflink_p_init(new_dst.k);
 362
 363                         u64 offset = le64_to_cpu(src_p.v->idx) +
 364                                 (src_want.offset -
 365                                  bkey_start_offset(src_k.k));
 366
 367                         dst_p->v.idx = cpu_to_le64(offset);
 368                 } else {
 369                         BUG();
 370                 }
 371
 372                 new_dst.k->k.p = dst_iter.pos;
 373                 bch2_key_resize(&new_dst.k->k,
 374                                 min(src_k.k->p.offset - src_want.offset,
 375                                     dst_end.offset - dst_iter.pos.offset));
 376
 377                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 378                                          new_dst.k, &disk_res, NULL,
 379                                          new_i_size, i_sectors_delta,
 380                                          true);
 381                 bch2_disk_reservation_put(c, &disk_res);
 382         }
 383         bch2_trans_iter_exit(&trans, &dst_iter);
 384         bch2_trans_iter_exit(&trans, &src_iter);
 385
 386         BUG_ON(!ret && bkey_cmp(dst_iter.pos, dst_end));
 387         BUG_ON(bkey_cmp(dst_iter.pos, dst_end) > 0);
 388
 389         dst_done = dst_iter.pos.offset - dst_start.offset;
 390         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 391
 392         do {
 393                 struct bch_inode_unpacked inode_u;
 394                 struct btree_iter inode_iter = { NULL };
 395
 396                 bch2_trans_begin(&trans);
 397
 398                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 399                                        dst_inum, BTREE_ITER_INTENT);
 400
 401                 if (!ret2 &&
 402                     inode_u.bi_size < new_i_size) {
 403                         inode_u.bi_size = new_i_size;
 404                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 405                                 bch2_trans_commit(&trans, NULL, NULL,
 406                                                   BTREE_INSERT_NOFAIL);
 407                 }
 408
 409                 bch2_trans_iter_exit(&trans, &inode_iter);
 410         } while (ret2 == -EINTR);
 411
 412         bch2_trans_exit(&trans);
 413         bch2_bkey_buf_exit(&new_src, c);
 414         bch2_bkey_buf_exit(&new_dst, c);
 415
 416         percpu_ref_put(&c->writes);
 417
 418         return dst_done ?: ret ?: ret2;
 419 }