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                            int rw, struct printbuf *err)
  30 {
  31         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  32
  33         if (bkey_val_bytes(p.k) != sizeof(*p.v)) {
  34                 prt_printf(err, "incorrect value size (%zu != %zu)",
  35                        bkey_val_bytes(p.k), sizeof(*p.v));
  36                 return -EINVAL;
  37         }
  38
  39         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix &&
  40             le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad)) {
  41                 prt_printf(err, "idx < front_pad (%llu < %u)",
  42                        le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
  43                 return -EINVAL;
  44         }
  45
  46         return 0;
  47 }
  48
  49 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  50                             struct bkey_s_c k)
  51 {
  52         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  53
  54         prt_printf(out, "idx %llu front_pad %u back_pad %u",
  55                le64_to_cpu(p.v->idx),
  56                le32_to_cpu(p.v->front_pad),
  57                le32_to_cpu(p.v->back_pad));
  58 }
  59
  60 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  61 {
  62         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  63         struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
  64
  65         /*
  66          * Disabled for now, the triggers code needs to be reworked for merging
  67          * of reflink pointers to work:
  68          */
  69         return false;
  70
  71         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  72                 return false;
  73
  74         bch2_key_resize(l.k, l.k->size + r.k->size);
  75         return true;
  76 }
  77
  78 /* indirect extents */
  79
  80 int bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k,
  81                            int rw, struct printbuf *err)
  82 {
  83         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  84
  85         if (bkey_val_bytes(r.k) < sizeof(*r.v)) {
  86                 prt_printf(err, "incorrect value size (%zu < %zu)",
  87                        bkey_val_bytes(r.k), sizeof(*r.v));
  88                 return -EINVAL;
  89         }
  90
  91         return bch2_bkey_ptrs_invalid(c, k, rw, err);
  92 }
  93
  94 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  95                             struct bkey_s_c k)
  96 {
  97         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  98
  99         prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
 100
 101         bch2_bkey_ptrs_to_text(out, c, k);
 102 }
 103
 104 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
 105 {
 106         struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
 107         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
 108
 109         return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
 110 }
 111
 112 int bch2_trans_mark_reflink_v(struct btree_trans *trans,
 113                               enum btree_id btree_id, unsigned level,
 114                               struct bkey_s_c old, struct bkey_i *new,
 115                               unsigned flags)
 116 {
 117         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
 118                 struct bkey_i_reflink_v *r = bkey_i_to_reflink_v(new);
 119
 120                 if (!r->v.refcount) {
 121                         r->k.type = KEY_TYPE_deleted;
 122                         r->k.size = 0;
 123                         set_bkey_val_u64s(&r->k, 0);
 124                         return 0;
 125                 }
 126         }
 127
 128         return bch2_trans_mark_extent(trans, btree_id, level, old, new, flags);
 129 }
 130
 131 /* indirect inline data */
 132
 133 int bch2_indirect_inline_data_invalid(const struct bch_fs *c, struct bkey_s_c k,
 134                                       int rw, struct printbuf *err)
 135 {
 136         if (bkey_val_bytes(k.k) < sizeof(struct bch_indirect_inline_data)) {
 137                 prt_printf(err, "incorrect value size (%zu < %zu)",
 138                        bkey_val_bytes(k.k), sizeof(struct bch_indirect_inline_data));
 139                 return -EINVAL;
 140         }
 141
 142         return 0;
 143 }
 144
 145 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 146                                         struct bch_fs *c, struct bkey_s_c k)
 147 {
 148         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 149         unsigned datalen = bkey_inline_data_bytes(k.k);
 150
 151         prt_printf(out, "refcount %llu datalen %u: %*phN",
 152                le64_to_cpu(d.v->refcount), datalen,
 153                min(datalen, 32U), d.v->data);
 154 }
 155
 156 int bch2_trans_mark_indirect_inline_data(struct btree_trans *trans,
 157                               enum btree_id btree_id, unsigned level,
 158                               struct bkey_s_c old, struct bkey_i *new,
 159                               unsigned flags)
 160 {
 161         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
 162                 struct bkey_i_indirect_inline_data *r =
 163                         bkey_i_to_indirect_inline_data(new);
 164
 165                 if (!r->v.refcount) {
 166                         r->k.type = KEY_TYPE_deleted;
 167                         r->k.size = 0;
 168                         set_bkey_val_u64s(&r->k, 0);
 169                 }
 170         }
 171
 172         return 0;
 173 }
 174
 175 static int bch2_make_extent_indirect(struct btree_trans *trans,
 176                                      struct btree_iter *extent_iter,
 177                                      struct bkey_i *orig)
 178 {
 179         struct bch_fs *c = trans->c;
 180         struct btree_iter reflink_iter = { NULL };
 181         struct bkey_s_c k;
 182         struct bkey_i *r_v;
 183         struct bkey_i_reflink_p *r_p;
 184         __le64 *refcount;
 185         int ret;
 186
 187         if (orig->k.type == KEY_TYPE_inline_data)
 188                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 189
 190         for_each_btree_key_norestart(trans, reflink_iter, BTREE_ID_reflink,
 191                            POS(0, c->reflink_hint),
 192                            BTREE_ITER_INTENT|BTREE_ITER_SLOTS, k, ret) {
 193                 if (reflink_iter.pos.inode) {
 194                         bch2_btree_iter_set_pos(&reflink_iter, POS_MIN);
 195                         continue;
 196                 }
 197
 198                 if (bkey_deleted(k.k) && orig->k.size <= k.k->size)
 199                         break;
 200         }
 201
 202         if (ret)
 203                 goto err;
 204
 205         /* rewind iter to start of hole, if necessary: */
 206         bch2_btree_iter_set_pos_to_extent_start(&reflink_iter);
 207
 208         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 209         ret = PTR_ERR_OR_ZERO(r_v);
 210         if (ret)
 211                 goto err;
 212
 213         bkey_init(&r_v->k);
 214         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 215         r_v->k.p        = reflink_iter.pos;
 216         bch2_key_resize(&r_v->k, orig->k.size);
 217         r_v->k.version  = orig->k.version;
 218
 219         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 220
 221         refcount        = bkey_refcount(r_v);
 222         *refcount       = 0;
 223         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 224
 225         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 226         if (ret)
 227                 goto err;
 228
 229         /*
 230          * orig is in a bkey_buf which statically allocates 5 64s for the val,
 231          * so we know it will be big enough:
 232          */
 233         orig->k.type = KEY_TYPE_reflink_p;
 234         r_p = bkey_i_to_reflink_p(orig);
 235         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 236         memset(&r_p->v, 0, sizeof(r_p->v));
 237
 238         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 239
 240         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 241                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 242 err:
 243         c->reflink_hint = reflink_iter.pos.offset;
 244         bch2_trans_iter_exit(trans, &reflink_iter);
 245
 246         return ret;
 247 }
 248
 249 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 250 {
 251         struct bkey_s_c k;
 252         int ret;
 253
 254         for_each_btree_key_continue_norestart(*iter, 0, k, ret) {
 255                 if (bkey_cmp(iter->pos, end) >= 0)
 256                         break;
 257
 258                 if (bkey_extent_is_unwritten(k))
 259                         continue;
 260
 261                 if (bkey_extent_is_data(k.k))
 262                         return k;
 263         }
 264
 265         if (bkey_cmp(iter->pos, end) >= 0)
 266                 bch2_btree_iter_set_pos(iter, end);
 267         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 268 }
 269
 270 s64 bch2_remap_range(struct bch_fs *c,
 271                      subvol_inum dst_inum, u64 dst_offset,
 272                      subvol_inum src_inum, u64 src_offset,
 273                      u64 remap_sectors,
 274                      u64 new_i_size, s64 *i_sectors_delta)
 275 {
 276         struct btree_trans trans;
 277         struct btree_iter dst_iter, src_iter;
 278         struct bkey_s_c src_k;
 279         struct bkey_buf new_dst, new_src;
 280         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 281         struct bpos src_start = POS(src_inum.inum, src_offset);
 282         struct bpos dst_end = dst_start, src_end = src_start;
 283         struct bpos src_want;
 284         u64 dst_done;
 285         u32 dst_snapshot, src_snapshot;
 286         int ret = 0, ret2 = 0;
 287
 288         if (!percpu_ref_tryget_live(&c->writes))
 289                 return -EROFS;
 290
 291         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 292
 293         dst_end.offset += remap_sectors;
 294         src_end.offset += remap_sectors;
 295
 296         bch2_bkey_buf_init(&new_dst);
 297         bch2_bkey_buf_init(&new_src);
 298         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 299
 300         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 301                              BTREE_ITER_INTENT);
 302         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 303                              BTREE_ITER_INTENT);
 304
 305         while ((ret == 0 ||
 306                 bch2_err_matches(ret, BCH_ERR_transaction_restart)) &&
 307                bkey_cmp(dst_iter.pos, dst_end) < 0) {
 308                 struct disk_reservation disk_res = { 0 };
 309
 310                 bch2_trans_begin(&trans);
 311
 312                 if (fatal_signal_pending(current)) {
 313                         ret = -EINTR;
 314                         break;
 315                 }
 316
 317                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 318                                                   &src_snapshot);
 319                 if (ret)
 320                         continue;
 321
 322                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 323
 324                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 325                                                   &dst_snapshot);
 326                 if (ret)
 327                         continue;
 328
 329                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 330
 331                 dst_done = dst_iter.pos.offset - dst_start.offset;
 332                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 333                 bch2_btree_iter_set_pos(&src_iter, src_want);
 334
 335                 src_k = get_next_src(&src_iter, src_end);
 336                 ret = bkey_err(src_k);
 337                 if (ret)
 338                         continue;
 339
 340                 if (bkey_cmp(src_want, src_iter.pos) < 0) {
 341                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 342                                         min(dst_end.offset,
 343                                             dst_iter.pos.offset +
 344                                             src_iter.pos.offset - src_want.offset),
 345                                         i_sectors_delta);
 346                         continue;
 347                 }
 348
 349                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 350                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 351
 352                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 353                         src_k = bkey_i_to_s_c(new_src.k);
 354
 355                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 356                                                 new_src.k);
 357                         if (ret)
 358                                 continue;
 359
 360                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 361                 }
 362
 363                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 364                         struct bkey_s_c_reflink_p src_p =
 365                                 bkey_s_c_to_reflink_p(src_k);
 366                         struct bkey_i_reflink_p *dst_p =
 367                                 bkey_reflink_p_init(new_dst.k);
 368
 369                         u64 offset = le64_to_cpu(src_p.v->idx) +
 370                                 (src_want.offset -
 371                                  bkey_start_offset(src_k.k));
 372
 373                         dst_p->v.idx = cpu_to_le64(offset);
 374                 } else {
 375                         BUG();
 376                 }
 377
 378                 new_dst.k->k.p = dst_iter.pos;
 379                 bch2_key_resize(&new_dst.k->k,
 380                                 min(src_k.k->p.offset - src_want.offset,
 381                                     dst_end.offset - dst_iter.pos.offset));
 382
 383                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 384                                          new_dst.k, &disk_res,
 385                                          new_i_size, i_sectors_delta,
 386                                          true);
 387                 bch2_disk_reservation_put(c, &disk_res);
 388         }
 389         bch2_trans_iter_exit(&trans, &dst_iter);
 390         bch2_trans_iter_exit(&trans, &src_iter);
 391
 392         BUG_ON(!ret && bkey_cmp(dst_iter.pos, dst_end));
 393         BUG_ON(bkey_cmp(dst_iter.pos, dst_end) > 0);
 394
 395         dst_done = dst_iter.pos.offset - dst_start.offset;
 396         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 397
 398         do {
 399                 struct bch_inode_unpacked inode_u;
 400                 struct btree_iter inode_iter = { NULL };
 401
 402                 bch2_trans_begin(&trans);
 403
 404                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 405                                        dst_inum, BTREE_ITER_INTENT);
 406
 407                 if (!ret2 &&
 408                     inode_u.bi_size < new_i_size) {
 409                         inode_u.bi_size = new_i_size;
 410                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 411                                 bch2_trans_commit(&trans, NULL, NULL,
 412                                                   BTREE_INSERT_NOFAIL);
 413                 }
 414
 415                 bch2_trans_iter_exit(&trans, &inode_iter);
 416         } while (bch2_err_matches(ret2, BCH_ERR_transaction_restart));
 417
 418         bch2_trans_exit(&trans);
 419         bch2_bkey_buf_exit(&new_src, c);
 420         bch2_bkey_buf_exit(&new_dst, c);
 421
 422         percpu_ref_put(&c->writes);
 423
 424         return dst_done ?: ret ?: ret2;
 425 }