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                            unsigned flags, struct printbuf *err)
  30 {
  31         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  32
  33         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix &&
  34             le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad)) {
  35                 prt_printf(err, "idx < front_pad (%llu < %u)",
  36                        le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
  37                 return -EINVAL;
  38         }
  39
  40         return 0;
  41 }
  42
  43 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  44                             struct bkey_s_c k)
  45 {
  46         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  47
  48         prt_printf(out, "idx %llu front_pad %u back_pad %u",
  49                le64_to_cpu(p.v->idx),
  50                le32_to_cpu(p.v->front_pad),
  51                le32_to_cpu(p.v->back_pad));
  52 }
  53
  54 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  55 {
  56         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  57         struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
  58
  59         /*
  60          * Disabled for now, the triggers code needs to be reworked for merging
  61          * of reflink pointers to work:
  62          */
  63         return false;
  64
  65         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  66                 return false;
  67
  68         bch2_key_resize(l.k, l.k->size + r.k->size);
  69         return true;
  70 }
  71
  72 /* indirect extents */
  73
  74 int bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k,
  75                            unsigned flags, struct printbuf *err)
  76 {
  77         return bch2_bkey_ptrs_invalid(c, k, flags, err);
  78 }
  79
  80 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  81                             struct bkey_s_c k)
  82 {
  83         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  84
  85         prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
  86
  87         bch2_bkey_ptrs_to_text(out, c, k);
  88 }
  89
  90 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  91 {
  92         struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
  93         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
  94
  95         return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
  96 }
  97
  98 static inline void check_indirect_extent_deleting(struct bkey_i *new, unsigned *flags)
  99 {
 100         if ((*flags & BTREE_TRIGGER_INSERT) && !*bkey_refcount(new)) {
 101                 new->k.type = KEY_TYPE_deleted;
 102                 new->k.size = 0;
 103                 set_bkey_val_u64s(&new->k, 0);;
 104                 *flags &= ~BTREE_TRIGGER_INSERT;
 105         }
 106 }
 107
 108 int bch2_trans_mark_reflink_v(struct btree_trans *trans,
 109                               enum btree_id btree_id, unsigned level,
 110                               struct bkey_s_c old, struct bkey_i *new,
 111                               unsigned flags)
 112 {
 113         check_indirect_extent_deleting(new, &flags);
 114
 115         return bch2_trans_mark_extent(trans, btree_id, level, old, new, flags);
 116 }
 117
 118 /* indirect inline data */
 119
 120 int bch2_indirect_inline_data_invalid(const struct bch_fs *c, struct bkey_s_c k,
 121                                       unsigned flags, struct printbuf *err)
 122 {
 123         return 0;
 124 }
 125
 126 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 127                                        struct bch_fs *c, struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 130         unsigned datalen = bkey_inline_data_bytes(k.k);
 131
 132         prt_printf(out, "refcount %llu datalen %u: %*phN",
 133                le64_to_cpu(d.v->refcount), datalen,
 134                min(datalen, 32U), d.v->data);
 135 }
 136
 137 int bch2_trans_mark_indirect_inline_data(struct btree_trans *trans,
 138                               enum btree_id btree_id, unsigned level,
 139                               struct bkey_s_c old, struct bkey_i *new,
 140                               unsigned flags)
 141 {
 142         check_indirect_extent_deleting(new, &flags);
 143
 144         return 0;
 145 }
 146
 147 static int bch2_make_extent_indirect(struct btree_trans *trans,
 148                                      struct btree_iter *extent_iter,
 149                                      struct bkey_i *orig)
 150 {
 151         struct bch_fs *c = trans->c;
 152         struct btree_iter reflink_iter = { NULL };
 153         struct bkey_s_c k;
 154         struct bkey_i *r_v;
 155         struct bkey_i_reflink_p *r_p;
 156         __le64 *refcount;
 157         int ret;
 158
 159         if (orig->k.type == KEY_TYPE_inline_data)
 160                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 161
 162         bch2_trans_iter_init(trans, &reflink_iter, BTREE_ID_reflink, POS_MAX,
 163                              BTREE_ITER_INTENT);
 164         k = bch2_btree_iter_peek_prev(&reflink_iter);
 165         ret = bkey_err(k);
 166         if (ret)
 167                 goto err;
 168
 169         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 170         ret = PTR_ERR_OR_ZERO(r_v);
 171         if (ret)
 172                 goto err;
 173
 174         bkey_init(&r_v->k);
 175         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 176         r_v->k.p        = reflink_iter.pos;
 177         bch2_key_resize(&r_v->k, orig->k.size);
 178         r_v->k.version  = orig->k.version;
 179
 180         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 181
 182         refcount        = bkey_refcount(r_v);
 183         *refcount       = 0;
 184         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 185
 186         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 187         if (ret)
 188                 goto err;
 189
 190         /*
 191          * orig is in a bkey_buf which statically allocates 5 64s for the val,
 192          * so we know it will be big enough:
 193          */
 194         orig->k.type = KEY_TYPE_reflink_p;
 195         r_p = bkey_i_to_reflink_p(orig);
 196         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 197
 198         /* FORTIFY_SOURCE is broken here, and doesn't provide unsafe_memset() */
 199 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
 200         __underlying_memset(&r_p->v, 0, sizeof(r_p->v));
 201 #else
 202         memset(&r_p->v, 0, sizeof(r_p->v));
 203 #endif
 204
 205         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 206
 207         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 208                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 209 err:
 210         bch2_trans_iter_exit(trans, &reflink_iter);
 211
 212         return ret;
 213 }
 214
 215 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 216 {
 217         struct bkey_s_c k;
 218         int ret;
 219
 220         for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
 221                 if (bkey_extent_is_unwritten(k))
 222                         continue;
 223
 224                 if (bkey_extent_is_data(k.k))
 225                         return k;
 226         }
 227
 228         if (bkey_ge(iter->pos, end))
 229                 bch2_btree_iter_set_pos(iter, end);
 230         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 231 }
 232
 233 s64 bch2_remap_range(struct bch_fs *c,
 234                      subvol_inum dst_inum, u64 dst_offset,
 235                      subvol_inum src_inum, u64 src_offset,
 236                      u64 remap_sectors,
 237                      u64 new_i_size, s64 *i_sectors_delta)
 238 {
 239         struct btree_trans trans;
 240         struct btree_iter dst_iter, src_iter;
 241         struct bkey_s_c src_k;
 242         struct bkey_buf new_dst, new_src;
 243         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 244         struct bpos src_start = POS(src_inum.inum, src_offset);
 245         struct bpos dst_end = dst_start, src_end = src_start;
 246         struct bpos src_want;
 247         u64 dst_done;
 248         u32 dst_snapshot, src_snapshot;
 249         int ret = 0, ret2 = 0;
 250
 251         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
 252                 return -BCH_ERR_erofs_no_writes;
 253
 254         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 255
 256         dst_end.offset += remap_sectors;
 257         src_end.offset += remap_sectors;
 258
 259         bch2_bkey_buf_init(&new_dst);
 260         bch2_bkey_buf_init(&new_src);
 261         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 262
 263         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 264                              BTREE_ITER_INTENT);
 265         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 266                              BTREE_ITER_INTENT);
 267
 268         while ((ret == 0 ||
 269                 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
 270                bkey_lt(dst_iter.pos, dst_end)) {
 271                 struct disk_reservation disk_res = { 0 };
 272
 273                 bch2_trans_begin(&trans);
 274
 275                 if (fatal_signal_pending(current)) {
 276                         ret = -EINTR;
 277                         break;
 278                 }
 279
 280                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 281                                                   &src_snapshot);
 282                 if (ret)
 283                         continue;
 284
 285                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 286
 287                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 288                                                   &dst_snapshot);
 289                 if (ret)
 290                         continue;
 291
 292                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 293
 294                 dst_done = dst_iter.pos.offset - dst_start.offset;
 295                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 296                 bch2_btree_iter_set_pos(&src_iter, src_want);
 297
 298                 src_k = get_next_src(&src_iter, src_end);
 299                 ret = bkey_err(src_k);
 300                 if (ret)
 301                         continue;
 302
 303                 if (bkey_lt(src_want, src_iter.pos)) {
 304                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 305                                         min(dst_end.offset,
 306                                             dst_iter.pos.offset +
 307                                             src_iter.pos.offset - src_want.offset),
 308                                         i_sectors_delta);
 309                         continue;
 310                 }
 311
 312                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 313                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 314
 315                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 316                         src_k = bkey_i_to_s_c(new_src.k);
 317
 318                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 319                                                 new_src.k);
 320                         if (ret)
 321                                 continue;
 322
 323                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 324                 }
 325
 326                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 327                         struct bkey_s_c_reflink_p src_p =
 328                                 bkey_s_c_to_reflink_p(src_k);
 329                         struct bkey_i_reflink_p *dst_p =
 330                                 bkey_reflink_p_init(new_dst.k);
 331
 332                         u64 offset = le64_to_cpu(src_p.v->idx) +
 333                                 (src_want.offset -
 334                                  bkey_start_offset(src_k.k));
 335
 336                         dst_p->v.idx = cpu_to_le64(offset);
 337                 } else {
 338                         BUG();
 339                 }
 340
 341                 new_dst.k->k.p = dst_iter.pos;
 342                 bch2_key_resize(&new_dst.k->k,
 343                                 min(src_k.k->p.offset - src_want.offset,
 344                                     dst_end.offset - dst_iter.pos.offset));
 345
 346                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 347                                          new_dst.k, &disk_res,
 348                                          new_i_size, i_sectors_delta,
 349                                          true);
 350                 bch2_disk_reservation_put(c, &disk_res);
 351         }
 352         bch2_trans_iter_exit(&trans, &dst_iter);
 353         bch2_trans_iter_exit(&trans, &src_iter);
 354
 355         BUG_ON(!ret && !bkey_eq(dst_iter.pos, dst_end));
 356         BUG_ON(bkey_gt(dst_iter.pos, dst_end));
 357
 358         dst_done = dst_iter.pos.offset - dst_start.offset;
 359         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 360
 361         do {
 362                 struct bch_inode_unpacked inode_u;
 363                 struct btree_iter inode_iter = { NULL };
 364
 365                 bch2_trans_begin(&trans);
 366
 367                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 368                                        dst_inum, BTREE_ITER_INTENT);
 369
 370                 if (!ret2 &&
 371                     inode_u.bi_size < new_i_size) {
 372                         inode_u.bi_size = new_i_size;
 373                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 374                                 bch2_trans_commit(&trans, NULL, NULL,
 375                                                   BTREE_INSERT_NOFAIL);
 376                 }
 377
 378                 bch2_trans_iter_exit(&trans, &inode_iter);
 379         } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
 380
 381         bch2_trans_exit(&trans);
 382         bch2_bkey_buf_exit(&new_src, c);
 383         bch2_bkey_buf_exit(&new_dst, c);
 384
 385         bch2_write_ref_put(c, BCH_WRITE_REF_reflink);
 386
 387         return dst_done ?: ret ?: ret2;
 388 }