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                            enum bkey_invalid_flags flags,
  30                            struct printbuf *err)
  31 {
  32         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  33
  34         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix &&
  35             le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad)) {
  36                 prt_printf(err, "idx < front_pad (%llu < %u)",
  37                        le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
  38                 return -EINVAL;
  39         }
  40
  41         return 0;
  42 }
  43
  44 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  45                             struct bkey_s_c k)
  46 {
  47         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  48
  49         prt_printf(out, "idx %llu front_pad %u back_pad %u",
  50                le64_to_cpu(p.v->idx),
  51                le32_to_cpu(p.v->front_pad),
  52                le32_to_cpu(p.v->back_pad));
  53 }
  54
  55 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  56 {
  57         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  58         struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
  59
  60         /*
  61          * Disabled for now, the triggers code needs to be reworked for merging
  62          * of reflink pointers to work:
  63          */
  64         return false;
  65
  66         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  67                 return false;
  68
  69         bch2_key_resize(l.k, l.k->size + r.k->size);
  70         return true;
  71 }
  72
  73 /* indirect extents */
  74
  75 int bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k,
  76                            enum bkey_invalid_flags flags,
  77                            struct printbuf *err)
  78 {
  79         return bch2_bkey_ptrs_invalid(c, k, flags, err);
  80 }
  81
  82 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  83                             struct bkey_s_c k)
  84 {
  85         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  86
  87         prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
  88
  89         bch2_bkey_ptrs_to_text(out, c, k);
  90 }
  91
  92 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  93 {
  94         struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
  95         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
  96
  97         return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
  98 }
  99
 100 static inline void check_indirect_extent_deleting(struct bkey_i *new, unsigned *flags)
 101 {
 102         if ((*flags & BTREE_TRIGGER_INSERT) && !*bkey_refcount(new)) {
 103                 new->k.type = KEY_TYPE_deleted;
 104                 new->k.size = 0;
 105                 set_bkey_val_u64s(&new->k, 0);;
 106                 *flags &= ~BTREE_TRIGGER_INSERT;
 107         }
 108 }
 109
 110 int bch2_trans_mark_reflink_v(struct btree_trans *trans,
 111                               enum btree_id btree_id, unsigned level,
 112                               struct bkey_s_c old, struct bkey_i *new,
 113                               unsigned flags)
 114 {
 115         check_indirect_extent_deleting(new, &flags);
 116
 117         return bch2_trans_mark_extent(trans, btree_id, level, old, new, flags);
 118 }
 119
 120 /* indirect inline data */
 121
 122 int bch2_indirect_inline_data_invalid(const struct bch_fs *c, struct bkey_s_c k,
 123                                       enum bkey_invalid_flags flags,
 124                                       struct printbuf *err)
 125 {
 126         return 0;
 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         prt_printf(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                               enum btree_id btree_id, unsigned level,
 142                               struct bkey_s_c old, struct bkey_i *new,
 143                               unsigned flags)
 144 {
 145         check_indirect_extent_deleting(new, &flags);
 146
 147         return 0;
 148 }
 149
 150 static int bch2_make_extent_indirect(struct btree_trans *trans,
 151                                      struct btree_iter *extent_iter,
 152                                      struct bkey_i *orig)
 153 {
 154         struct bch_fs *c = trans->c;
 155         struct btree_iter reflink_iter = { NULL };
 156         struct bkey_s_c k;
 157         struct bkey_i *r_v;
 158         struct bkey_i_reflink_p *r_p;
 159         __le64 *refcount;
 160         int ret;
 161
 162         if (orig->k.type == KEY_TYPE_inline_data)
 163                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 164
 165         bch2_trans_iter_init(trans, &reflink_iter, BTREE_ID_reflink, POS_MAX,
 166                              BTREE_ITER_INTENT);
 167         k = bch2_btree_iter_peek_prev(&reflink_iter);
 168         ret = bkey_err(k);
 169         if (ret)
 170                 goto err;
 171
 172         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 173         ret = PTR_ERR_OR_ZERO(r_v);
 174         if (ret)
 175                 goto err;
 176
 177         bkey_init(&r_v->k);
 178         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 179         r_v->k.p        = reflink_iter.pos;
 180         bch2_key_resize(&r_v->k, orig->k.size);
 181         r_v->k.version  = orig->k.version;
 182
 183         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 184
 185         refcount        = bkey_refcount(r_v);
 186         *refcount       = 0;
 187         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 188
 189         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 190         if (ret)
 191                 goto err;
 192
 193         /*
 194          * orig is in a bkey_buf which statically allocates 5 64s for the val,
 195          * so we know it will be big enough:
 196          */
 197         orig->k.type = KEY_TYPE_reflink_p;
 198         r_p = bkey_i_to_reflink_p(orig);
 199         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 200
 201         /* FORTIFY_SOURCE is broken here, and doesn't provide unsafe_memset() */
 202 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
 203         __underlying_memset(&r_p->v, 0, sizeof(r_p->v));
 204 #else
 205         memset(&r_p->v, 0, sizeof(r_p->v));
 206 #endif
 207
 208         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 209
 210         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 211                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 212 err:
 213         bch2_trans_iter_exit(trans, &reflink_iter);
 214
 215         return ret;
 216 }
 217
 218 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 219 {
 220         struct bkey_s_c k;
 221         int ret;
 222
 223         for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
 224                 if (bkey_extent_is_unwritten(k))
 225                         continue;
 226
 227                 if (bkey_extent_is_data(k.k))
 228                         return k;
 229         }
 230
 231         if (bkey_ge(iter->pos, end))
 232                 bch2_btree_iter_set_pos(iter, end);
 233         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 234 }
 235
 236 s64 bch2_remap_range(struct bch_fs *c,
 237                      subvol_inum dst_inum, u64 dst_offset,
 238                      subvol_inum src_inum, u64 src_offset,
 239                      u64 remap_sectors,
 240                      u64 new_i_size, s64 *i_sectors_delta)
 241 {
 242         struct btree_trans trans;
 243         struct btree_iter dst_iter, src_iter;
 244         struct bkey_s_c src_k;
 245         struct bkey_buf new_dst, new_src;
 246         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 247         struct bpos src_start = POS(src_inum.inum, src_offset);
 248         struct bpos dst_end = dst_start, src_end = src_start;
 249         struct bpos src_want;
 250         u64 dst_done;
 251         u32 dst_snapshot, src_snapshot;
 252         int ret = 0, ret2 = 0;
 253
 254         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
 255                 return -BCH_ERR_erofs_no_writes;
 256
 257         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 258
 259         dst_end.offset += remap_sectors;
 260         src_end.offset += remap_sectors;
 261
 262         bch2_bkey_buf_init(&new_dst);
 263         bch2_bkey_buf_init(&new_src);
 264         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 265
 266         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 267                              BTREE_ITER_INTENT);
 268         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 269                              BTREE_ITER_INTENT);
 270
 271         while ((ret == 0 ||
 272                 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
 273                bkey_lt(dst_iter.pos, dst_end)) {
 274                 struct disk_reservation disk_res = { 0 };
 275
 276                 bch2_trans_begin(&trans);
 277
 278                 if (fatal_signal_pending(current)) {
 279                         ret = -EINTR;
 280                         break;
 281                 }
 282
 283                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 284                                                   &src_snapshot);
 285                 if (ret)
 286                         continue;
 287
 288                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 289
 290                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 291                                                   &dst_snapshot);
 292                 if (ret)
 293                         continue;
 294
 295                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 296
 297                 dst_done = dst_iter.pos.offset - dst_start.offset;
 298                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 299                 bch2_btree_iter_set_pos(&src_iter, src_want);
 300
 301                 src_k = get_next_src(&src_iter, src_end);
 302                 ret = bkey_err(src_k);
 303                 if (ret)
 304                         continue;
 305
 306                 if (bkey_lt(src_want, src_iter.pos)) {
 307                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 308                                         min(dst_end.offset,
 309                                             dst_iter.pos.offset +
 310                                             src_iter.pos.offset - src_want.offset),
 311                                         i_sectors_delta);
 312                         continue;
 313                 }
 314
 315                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 316                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 317
 318                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 319                         src_k = bkey_i_to_s_c(new_src.k);
 320
 321                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 322                                                 new_src.k);
 323                         if (ret)
 324                                 continue;
 325
 326                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 327                 }
 328
 329                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 330                         struct bkey_s_c_reflink_p src_p =
 331                                 bkey_s_c_to_reflink_p(src_k);
 332                         struct bkey_i_reflink_p *dst_p =
 333                                 bkey_reflink_p_init(new_dst.k);
 334
 335                         u64 offset = le64_to_cpu(src_p.v->idx) +
 336                                 (src_want.offset -
 337                                  bkey_start_offset(src_k.k));
 338
 339                         dst_p->v.idx = cpu_to_le64(offset);
 340                 } else {
 341                         BUG();
 342                 }
 343
 344                 new_dst.k->k.p = dst_iter.pos;
 345                 bch2_key_resize(&new_dst.k->k,
 346                                 min(src_k.k->p.offset - src_want.offset,
 347                                     dst_end.offset - dst_iter.pos.offset));
 348
 349                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 350                                          new_dst.k, &disk_res,
 351                                          new_i_size, i_sectors_delta,
 352                                          true);
 353                 bch2_disk_reservation_put(c, &disk_res);
 354         }
 355         bch2_trans_iter_exit(&trans, &dst_iter);
 356         bch2_trans_iter_exit(&trans, &src_iter);
 357
 358         BUG_ON(!ret && !bkey_eq(dst_iter.pos, dst_end));
 359         BUG_ON(bkey_gt(dst_iter.pos, dst_end));
 360
 361         dst_done = dst_iter.pos.offset - dst_start.offset;
 362         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 363
 364         do {
 365                 struct bch_inode_unpacked inode_u;
 366                 struct btree_iter inode_iter = { NULL };
 367
 368                 bch2_trans_begin(&trans);
 369
 370                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 371                                        dst_inum, BTREE_ITER_INTENT);
 372
 373                 if (!ret2 &&
 374                     inode_u.bi_size < new_i_size) {
 375                         inode_u.bi_size = new_i_size;
 376                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 377                                 bch2_trans_commit(&trans, NULL, NULL,
 378                                                   BTREE_INSERT_NOFAIL);
 379                 }
 380
 381                 bch2_trans_iter_exit(&trans, &inode_iter);
 382         } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
 383
 384         bch2_trans_exit(&trans);
 385         bch2_bkey_buf_exit(&new_src, c);
 386         bch2_bkey_buf_exit(&new_dst, c);
 387
 388         bch2_write_ref_put(c, BCH_WRITE_REF_reflink);
 389
 390         return dst_done ?: ret ?: ret2;
 391 }