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 int bch2_trans_mark_reflink_v(struct btree_trans *trans,
 101                               enum btree_id btree_id, unsigned level,
 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, btree_id, level, old, new, flags);
 117 }
 118
 119 /* indirect inline data */
 120
 121 int bch2_indirect_inline_data_invalid(const struct bch_fs *c, struct bkey_s_c k,
 122                                       enum bkey_invalid_flags flags,
 123                                       struct printbuf *err)
 124 {
 125         return 0;
 126 }
 127
 128 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 129                                         struct bch_fs *c, struct bkey_s_c k)
 130 {
 131         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 132         unsigned datalen = bkey_inline_data_bytes(k.k);
 133
 134         prt_printf(out, "refcount %llu datalen %u: %*phN",
 135                le64_to_cpu(d.v->refcount), datalen,
 136                min(datalen, 32U), d.v->data);
 137 }
 138
 139 int bch2_trans_mark_indirect_inline_data(struct btree_trans *trans,
 140                               enum btree_id btree_id, unsigned level,
 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         bch2_trans_iter_init(trans, &reflink_iter, BTREE_ID_reflink, POS_MAX,
 174                              BTREE_ITER_INTENT);
 175         k = bch2_btree_iter_peek_prev(&reflink_iter);
 176         ret = bkey_err(k);
 177         if (ret)
 178                 goto err;
 179
 180         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 181         ret = PTR_ERR_OR_ZERO(r_v);
 182         if (ret)
 183                 goto err;
 184
 185         bkey_init(&r_v->k);
 186         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 187         r_v->k.p        = reflink_iter.pos;
 188         bch2_key_resize(&r_v->k, orig->k.size);
 189         r_v->k.version  = orig->k.version;
 190
 191         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 192
 193         refcount        = bkey_refcount(r_v);
 194         *refcount       = 0;
 195         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 196
 197         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 198         if (ret)
 199                 goto err;
 200
 201         /*
 202          * orig is in a bkey_buf which statically allocates 5 64s for the val,
 203          * so we know it will be big enough:
 204          */
 205         orig->k.type = KEY_TYPE_reflink_p;
 206         r_p = bkey_i_to_reflink_p(orig);
 207         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 208
 209         /* FORTIFY_SOURCE is broken here, and doesn't provide unsafe_memset() */
 210 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
 211         __underlying_memset(&r_p->v, 0, sizeof(r_p->v));
 212 #else
 213         memset(&r_p->v, 0, sizeof(r_p->v));
 214 #endif
 215
 216         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 217
 218         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 219                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 220 err:
 221         bch2_trans_iter_exit(trans, &reflink_iter);
 222
 223         return ret;
 224 }
 225
 226 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 227 {
 228         struct bkey_s_c k;
 229         int ret;
 230
 231         for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
 232                 if (bkey_extent_is_unwritten(k))
 233                         continue;
 234
 235                 if (bkey_extent_is_data(k.k))
 236                         return k;
 237         }
 238
 239         if (bkey_ge(iter->pos, end))
 240                 bch2_btree_iter_set_pos(iter, end);
 241         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 242 }
 243
 244 s64 bch2_remap_range(struct bch_fs *c,
 245                      subvol_inum dst_inum, u64 dst_offset,
 246                      subvol_inum src_inum, u64 src_offset,
 247                      u64 remap_sectors,
 248                      u64 new_i_size, s64 *i_sectors_delta)
 249 {
 250         struct btree_trans trans;
 251         struct btree_iter dst_iter, src_iter;
 252         struct bkey_s_c src_k;
 253         struct bkey_buf new_dst, new_src;
 254         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 255         struct bpos src_start = POS(src_inum.inum, src_offset);
 256         struct bpos dst_end = dst_start, src_end = src_start;
 257         struct bpos src_want;
 258         u64 dst_done;
 259         u32 dst_snapshot, src_snapshot;
 260         int ret = 0, ret2 = 0;
 261
 262         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
 263                 return -BCH_ERR_erofs_no_writes;
 264
 265         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 266
 267         dst_end.offset += remap_sectors;
 268         src_end.offset += remap_sectors;
 269
 270         bch2_bkey_buf_init(&new_dst);
 271         bch2_bkey_buf_init(&new_src);
 272         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 273
 274         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 275                              BTREE_ITER_INTENT);
 276         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 277                              BTREE_ITER_INTENT);
 278
 279         while ((ret == 0 ||
 280                 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
 281                bkey_lt(dst_iter.pos, dst_end)) {
 282                 struct disk_reservation disk_res = { 0 };
 283
 284                 bch2_trans_begin(&trans);
 285
 286                 if (fatal_signal_pending(current)) {
 287                         ret = -EINTR;
 288                         break;
 289                 }
 290
 291                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 292                                                   &src_snapshot);
 293                 if (ret)
 294                         continue;
 295
 296                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 297
 298                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 299                                                   &dst_snapshot);
 300                 if (ret)
 301                         continue;
 302
 303                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 304
 305                 dst_done = dst_iter.pos.offset - dst_start.offset;
 306                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 307                 bch2_btree_iter_set_pos(&src_iter, src_want);
 308
 309                 src_k = get_next_src(&src_iter, src_end);
 310                 ret = bkey_err(src_k);
 311                 if (ret)
 312                         continue;
 313
 314                 if (bkey_lt(src_want, src_iter.pos)) {
 315                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 316                                         min(dst_end.offset,
 317                                             dst_iter.pos.offset +
 318                                             src_iter.pos.offset - src_want.offset),
 319                                         i_sectors_delta);
 320                         continue;
 321                 }
 322
 323                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 324                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 325
 326                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 327                         src_k = bkey_i_to_s_c(new_src.k);
 328
 329                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 330                                                 new_src.k);
 331                         if (ret)
 332                                 continue;
 333
 334                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 335                 }
 336
 337                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 338                         struct bkey_s_c_reflink_p src_p =
 339                                 bkey_s_c_to_reflink_p(src_k);
 340                         struct bkey_i_reflink_p *dst_p =
 341                                 bkey_reflink_p_init(new_dst.k);
 342
 343                         u64 offset = le64_to_cpu(src_p.v->idx) +
 344                                 (src_want.offset -
 345                                  bkey_start_offset(src_k.k));
 346
 347                         dst_p->v.idx = cpu_to_le64(offset);
 348                 } else {
 349                         BUG();
 350                 }
 351
 352                 new_dst.k->k.p = dst_iter.pos;
 353                 bch2_key_resize(&new_dst.k->k,
 354                                 min(src_k.k->p.offset - src_want.offset,
 355                                     dst_end.offset - dst_iter.pos.offset));
 356
 357                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 358                                          new_dst.k, &disk_res,
 359                                          new_i_size, i_sectors_delta,
 360                                          true);
 361                 bch2_disk_reservation_put(c, &disk_res);
 362         }
 363         bch2_trans_iter_exit(&trans, &dst_iter);
 364         bch2_trans_iter_exit(&trans, &src_iter);
 365
 366         BUG_ON(!ret && !bkey_eq(dst_iter.pos, dst_end));
 367         BUG_ON(bkey_gt(dst_iter.pos, dst_end));
 368
 369         dst_done = dst_iter.pos.offset - dst_start.offset;
 370         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 371
 372         do {
 373                 struct bch_inode_unpacked inode_u;
 374                 struct btree_iter inode_iter = { NULL };
 375
 376                 bch2_trans_begin(&trans);
 377
 378                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 379                                        dst_inum, BTREE_ITER_INTENT);
 380
 381                 if (!ret2 &&
 382                     inode_u.bi_size < new_i_size) {
 383                         inode_u.bi_size = new_i_size;
 384                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 385                                 bch2_trans_commit(&trans, NULL, NULL,
 386                                                   BTREE_INSERT_NOFAIL);
 387                 }
 388
 389                 bch2_trans_iter_exit(&trans, &inode_iter);
 390         } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
 391
 392         bch2_trans_exit(&trans);
 393         bch2_bkey_buf_exit(&new_src, c);
 394         bch2_bkey_buf_exit(&new_dst, c);
 395
 396         bch2_write_ref_put(c, BCH_WRITE_REF_reflink);
 397
 398         return dst_done ?: ret ?: ret2;
 399 }