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 int bch2_trans_mark_reflink_v(struct btree_trans *trans,
  99                               enum btree_id btree_id, unsigned level,
 100                               struct bkey_s_c old, struct bkey_i *new,
 101                               unsigned flags)
 102 {
 103         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
 104                 struct bkey_i_reflink_v *r = bkey_i_to_reflink_v(new);
 105
 106                 if (!r->v.refcount) {
 107                         r->k.type = KEY_TYPE_deleted;
 108                         r->k.size = 0;
 109                         set_bkey_val_u64s(&r->k, 0);
 110                         return 0;
 111                 }
 112         }
 113
 114         return bch2_trans_mark_extent(trans, btree_id, level, old, new, flags);
 115 }
 116
 117 /* indirect inline data */
 118
 119 int bch2_indirect_inline_data_invalid(const struct bch_fs *c, struct bkey_s_c k,
 120                                       unsigned flags, struct printbuf *err)
 121 {
 122         return 0;
 123 }
 124
 125 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 126                                         struct bch_fs *c, struct bkey_s_c k)
 127 {
 128         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 129         unsigned datalen = bkey_inline_data_bytes(k.k);
 130
 131         prt_printf(out, "refcount %llu datalen %u: %*phN",
 132                le64_to_cpu(d.v->refcount), datalen,
 133                min(datalen, 32U), d.v->data);
 134 }
 135
 136 int bch2_trans_mark_indirect_inline_data(struct btree_trans *trans,
 137                               enum btree_id btree_id, unsigned level,
 138                               struct bkey_s_c old, struct bkey_i *new,
 139                               unsigned flags)
 140 {
 141         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
 142                 struct bkey_i_indirect_inline_data *r =
 143                         bkey_i_to_indirect_inline_data(new);
 144
 145                 if (!r->v.refcount) {
 146                         r->k.type = KEY_TYPE_deleted;
 147                         r->k.size = 0;
 148                         set_bkey_val_u64s(&r->k, 0);
 149                 }
 150         }
 151
 152         return 0;
 153 }
 154
 155 static int bch2_make_extent_indirect(struct btree_trans *trans,
 156                                      struct btree_iter *extent_iter,
 157                                      struct bkey_i *orig)
 158 {
 159         struct bch_fs *c = trans->c;
 160         struct btree_iter reflink_iter = { NULL };
 161         struct bkey_s_c k;
 162         struct bkey_i *r_v;
 163         struct bkey_i_reflink_p *r_p;
 164         __le64 *refcount;
 165         int ret;
 166
 167         if (orig->k.type == KEY_TYPE_inline_data)
 168                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 169
 170         for_each_btree_key_norestart(trans, reflink_iter, BTREE_ID_reflink,
 171                            POS(0, c->reflink_hint),
 172                            BTREE_ITER_SLOTS, k, ret) {
 173                 if (reflink_iter.pos.inode) {
 174                         bch2_btree_iter_set_pos(&reflink_iter, POS_MIN);
 175                         continue;
 176                 }
 177
 178                 if (bkey_deleted(k.k) && orig->k.size <= k.k->size)
 179                         break;
 180         }
 181
 182         if (ret)
 183                 goto err;
 184
 185         /* rewind iter to start of hole, if necessary: */
 186         bch2_btree_iter_set_pos_to_extent_start(&reflink_iter);
 187
 188         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 189         ret = PTR_ERR_OR_ZERO(r_v);
 190         if (ret)
 191                 goto err;
 192
 193         bkey_init(&r_v->k);
 194         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 195         r_v->k.p        = reflink_iter.pos;
 196         bch2_key_resize(&r_v->k, orig->k.size);
 197         r_v->k.version  = orig->k.version;
 198
 199         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 200
 201         refcount        = bkey_refcount(r_v);
 202         *refcount       = 0;
 203         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 204
 205         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 206         if (ret)
 207                 goto err;
 208
 209         /*
 210          * orig is in a bkey_buf which statically allocates 5 64s for the val,
 211          * so we know it will be big enough:
 212          */
 213         orig->k.type = KEY_TYPE_reflink_p;
 214         r_p = bkey_i_to_reflink_p(orig);
 215         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 216
 217         /* FORTIFY_SOURCE is broken here, and doesn't provide unsafe_memset() */
 218 #if !defined(__NO_FORTIFY) && defined(__OPTIMIZE__) && defined(CONFIG_FORTIFY_SOURCE)
 219         __underlying_memset(&r_p->v, 0, sizeof(r_p->v));
 220 #else
 221         memset(&r_p->v, 0, sizeof(r_p->v));
 222 #endif
 223
 224         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 225
 226         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 227                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 228 err:
 229         c->reflink_hint = reflink_iter.pos.offset;
 230         bch2_trans_iter_exit(trans, &reflink_iter);
 231
 232         return ret;
 233 }
 234
 235 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 236 {
 237         struct bkey_s_c k;
 238         int ret;
 239
 240         for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
 241                 if (bkey_extent_is_unwritten(k))
 242                         continue;
 243
 244                 if (bkey_extent_is_data(k.k))
 245                         return k;
 246         }
 247
 248         if (bkey_ge(iter->pos, end))
 249                 bch2_btree_iter_set_pos(iter, end);
 250         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 251 }
 252
 253 s64 bch2_remap_range(struct bch_fs *c,
 254                      subvol_inum dst_inum, u64 dst_offset,
 255                      subvol_inum src_inum, u64 src_offset,
 256                      u64 remap_sectors,
 257                      u64 new_i_size, s64 *i_sectors_delta)
 258 {
 259         struct btree_trans trans;
 260         struct btree_iter dst_iter, src_iter;
 261         struct bkey_s_c src_k;
 262         struct bkey_buf new_dst, new_src;
 263         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 264         struct bpos src_start = POS(src_inum.inum, src_offset);
 265         struct bpos dst_end = dst_start, src_end = src_start;
 266         struct bpos src_want;
 267         u64 dst_done;
 268         u32 dst_snapshot, src_snapshot;
 269         int ret = 0, ret2 = 0;
 270
 271         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
 272                 return -BCH_ERR_erofs_no_writes;
 273
 274         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 275
 276         dst_end.offset += remap_sectors;
 277         src_end.offset += remap_sectors;
 278
 279         bch2_bkey_buf_init(&new_dst);
 280         bch2_bkey_buf_init(&new_src);
 281         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 282
 283         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 284                              BTREE_ITER_INTENT);
 285         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 286                              BTREE_ITER_INTENT);
 287
 288         while ((ret == 0 ||
 289                 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
 290                bkey_lt(dst_iter.pos, dst_end)) {
 291                 struct disk_reservation disk_res = { 0 };
 292
 293                 bch2_trans_begin(&trans);
 294
 295                 if (fatal_signal_pending(current)) {
 296                         ret = -EINTR;
 297                         break;
 298                 }
 299
 300                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 301                                                   &src_snapshot);
 302                 if (ret)
 303                         continue;
 304
 305                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 306
 307                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 308                                                   &dst_snapshot);
 309                 if (ret)
 310                         continue;
 311
 312                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 313
 314                 dst_done = dst_iter.pos.offset - dst_start.offset;
 315                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 316                 bch2_btree_iter_set_pos(&src_iter, src_want);
 317
 318                 src_k = get_next_src(&src_iter, src_end);
 319                 ret = bkey_err(src_k);
 320                 if (ret)
 321                         continue;
 322
 323                 if (bkey_lt(src_want, src_iter.pos)) {
 324                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 325                                         min(dst_end.offset,
 326                                             dst_iter.pos.offset +
 327                                             src_iter.pos.offset - src_want.offset),
 328                                         i_sectors_delta);
 329                         continue;
 330                 }
 331
 332                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 333                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 334
 335                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 336                         src_k = bkey_i_to_s_c(new_src.k);
 337
 338                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 339                                                 new_src.k);
 340                         if (ret)
 341                                 continue;
 342
 343                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 344                 }
 345
 346                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 347                         struct bkey_s_c_reflink_p src_p =
 348                                 bkey_s_c_to_reflink_p(src_k);
 349                         struct bkey_i_reflink_p *dst_p =
 350                                 bkey_reflink_p_init(new_dst.k);
 351
 352                         u64 offset = le64_to_cpu(src_p.v->idx) +
 353                                 (src_want.offset -
 354                                  bkey_start_offset(src_k.k));
 355
 356                         dst_p->v.idx = cpu_to_le64(offset);
 357                 } else {
 358                         BUG();
 359                 }
 360
 361                 new_dst.k->k.p = dst_iter.pos;
 362                 bch2_key_resize(&new_dst.k->k,
 363                                 min(src_k.k->p.offset - src_want.offset,
 364                                     dst_end.offset - dst_iter.pos.offset));
 365
 366                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 367                                          new_dst.k, &disk_res,
 368                                          new_i_size, i_sectors_delta,
 369                                          true);
 370                 bch2_disk_reservation_put(c, &disk_res);
 371         }
 372         bch2_trans_iter_exit(&trans, &dst_iter);
 373         bch2_trans_iter_exit(&trans, &src_iter);
 374
 375         BUG_ON(!ret && !bkey_eq(dst_iter.pos, dst_end));
 376         BUG_ON(bkey_gt(dst_iter.pos, dst_end));
 377
 378         dst_done = dst_iter.pos.offset - dst_start.offset;
 379         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 380
 381         do {
 382                 struct bch_inode_unpacked inode_u;
 383                 struct btree_iter inode_iter = { NULL };
 384
 385                 bch2_trans_begin(&trans);
 386
 387                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 388                                        dst_inum, BTREE_ITER_INTENT);
 389
 390                 if (!ret2 &&
 391                     inode_u.bi_size < new_i_size) {
 392                         inode_u.bi_size = new_i_size;
 393                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 394                                 bch2_trans_commit(&trans, NULL, NULL,
 395                                                   BTREE_INSERT_NOFAIL);
 396                 }
 397
 398                 bch2_trans_iter_exit(&trans, &inode_iter);
 399         } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
 400
 401         bch2_trans_exit(&trans);
 402         bch2_bkey_buf_exit(&new_src, c);
 403         bch2_bkey_buf_exit(&new_dst, c);
 404
 405         bch2_write_ref_put(c, BCH_WRITE_REF_reflink);
 406
 407         return dst_done ?: ret ?: ret2;
 408 }