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