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