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_on_stack.h"
   4 #include "btree_update.h"
   5 #include "extents.h"
   6 #include "inode.h"
   7 #include "io.h"
   8 #include "reflink.h"
   9
  10 #include <linux/sched/signal.h>
  11
  12 /* reflink pointers */
  13
  14 const char *bch2_reflink_p_invalid(const struct bch_fs *c, struct bkey_s_c k)
  15 {
  16         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  17
  18         if (bkey_val_bytes(p.k) != sizeof(*p.v))
  19                 return "incorrect value size";
  20
  21         return NULL;
  22 }
  23
  24 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  25                             struct bkey_s_c k)
  26 {
  27         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  28
  29         pr_buf(out, "idx %llu", le64_to_cpu(p.v->idx));
  30 }
  31
  32 enum merge_result bch2_reflink_p_merge(struct bch_fs *c,
  33                                        struct bkey_s _l, struct bkey_s _r)
  34 {
  35         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  36         struct bkey_s_reflink_p r = bkey_s_to_reflink_p(_r);
  37
  38         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  39                 return BCH_MERGE_NOMERGE;
  40
  41         if ((u64) l.k->size + r.k->size > KEY_SIZE_MAX) {
  42                 bch2_key_resize(l.k, KEY_SIZE_MAX);
  43                 bch2_cut_front_s(l.k->p, _r);
  44                 return BCH_MERGE_PARTIAL;
  45         }
  46
  47         bch2_key_resize(l.k, l.k->size + r.k->size);
  48
  49         return BCH_MERGE_MERGE;
  50 }
  51
  52 /* indirect extents */
  53
  54 const char *bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k)
  55 {
  56         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  57
  58         if (bkey_val_bytes(r.k) < sizeof(*r.v))
  59                 return "incorrect value size";
  60
  61         return bch2_bkey_ptrs_invalid(c, k);
  62 }
  63
  64 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  65                             struct bkey_s_c k)
  66 {
  67         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  68
  69         pr_buf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
  70
  71         bch2_bkey_ptrs_to_text(out, c, k);
  72 }
  73
  74 static int bch2_make_extent_indirect(struct btree_trans *trans,
  75                                      struct btree_iter *extent_iter,
  76                                      struct bkey_i_extent *e)
  77 {
  78         struct bch_fs *c = trans->c;
  79         struct btree_iter *reflink_iter;
  80         struct bkey_s_c k;
  81         struct bkey_i_reflink_v *r_v;
  82         struct bkey_i_reflink_p *r_p;
  83         int ret;
  84
  85         for_each_btree_key(trans, reflink_iter, BTREE_ID_REFLINK,
  86                            POS(0, c->reflink_hint),
  87                            BTREE_ITER_INTENT|BTREE_ITER_SLOTS, k, ret) {
  88                 if (reflink_iter->pos.inode) {
  89                         bch2_btree_iter_set_pos(reflink_iter, POS_MIN);
  90                         continue;
  91                 }
  92
  93                 if (bkey_deleted(k.k) && e->k.size <= k.k->size)
  94                         break;
  95         }
  96
  97         if (ret)
  98                 goto err;
  99
 100         /* rewind iter to start of hole, if necessary: */
 101         bch2_btree_iter_set_pos(reflink_iter, bkey_start_pos(k.k));
 102
 103         r_v = bch2_trans_kmalloc(trans, sizeof(*r_v) + bkey_val_bytes(&e->k));
 104         ret = PTR_ERR_OR_ZERO(r_v);
 105         if (ret)
 106                 goto err;
 107
 108         bkey_reflink_v_init(&r_v->k_i);
 109         r_v->k.p        = reflink_iter->pos;
 110         bch2_key_resize(&r_v->k, e->k.size);
 111         r_v->k.version  = e->k.version;
 112
 113         set_bkey_val_u64s(&r_v->k, bkey_val_u64s(&r_v->k) +
 114                           bkey_val_u64s(&e->k));
 115         r_v->v.refcount = 0;
 116         memcpy(r_v->v.start, e->v.start, bkey_val_bytes(&e->k));
 117
 118         bch2_trans_update(trans, reflink_iter, &r_v->k_i);
 119
 120         r_p = bch2_trans_kmalloc(trans, sizeof(*r_p));
 121         if (IS_ERR(r_p))
 122                 return PTR_ERR(r_p);
 123
 124         e->k.type = KEY_TYPE_reflink_p;
 125         r_p = bkey_i_to_reflink_p(&e->k_i);
 126         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 127         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 128
 129         bch2_trans_update(trans, extent_iter, &r_p->k_i);
 130 err:
 131         if (!IS_ERR(reflink_iter)) {
 132                 c->reflink_hint = reflink_iter->pos.offset;
 133                 bch2_trans_iter_put(trans, reflink_iter);
 134         }
 135
 136         return ret;
 137 }
 138
 139 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 140 {
 141         struct bkey_s_c k = bch2_btree_iter_peek(iter);
 142         int ret;
 143
 144         for_each_btree_key_continue(iter, 0, k, ret) {
 145                 if (bkey_cmp(iter->pos, end) >= 0)
 146                         return bkey_s_c_null;
 147
 148                 if (k.k->type == KEY_TYPE_extent ||
 149                     k.k->type == KEY_TYPE_reflink_p)
 150                         break;
 151         }
 152
 153         return k;
 154 }
 155
 156 s64 bch2_remap_range(struct bch_fs *c,
 157                      struct bpos dst_start, struct bpos src_start,
 158                      u64 remap_sectors, u64 *journal_seq,
 159                      u64 new_i_size, s64 *i_sectors_delta)
 160 {
 161         struct btree_trans trans;
 162         struct btree_iter *dst_iter, *src_iter;
 163         struct bkey_s_c src_k;
 164         BKEY_PADDED(k) new_dst;
 165         struct bkey_on_stack new_src;
 166         struct bpos dst_end = dst_start, src_end = src_start;
 167         struct bpos dst_want, src_want;
 168         u64 src_done, dst_done;
 169         int ret = 0, ret2 = 0;
 170
 171         if (!percpu_ref_tryget(&c->writes))
 172                 return -EROFS;
 173
 174         if (!(c->sb.features & (1ULL << BCH_FEATURE_REFLINK))) {
 175                 mutex_lock(&c->sb_lock);
 176                 if (!(c->sb.features & (1ULL << BCH_FEATURE_REFLINK))) {
 177                         c->disk_sb.sb->features[0] |=
 178                                 cpu_to_le64(1ULL << BCH_FEATURE_REFLINK);
 179
 180                         bch2_write_super(c);
 181                 }
 182                 mutex_unlock(&c->sb_lock);
 183         }
 184
 185         dst_end.offset += remap_sectors;
 186         src_end.offset += remap_sectors;
 187
 188         bkey_on_stack_init(&new_src);
 189         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 190
 191         src_iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, src_start,
 192                                        BTREE_ITER_INTENT);
 193         dst_iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, dst_start,
 194                                        BTREE_ITER_INTENT);
 195
 196         while (1) {
 197                 bch2_trans_begin_updates(&trans);
 198                 trans.mem_top = 0;
 199
 200                 if (fatal_signal_pending(current)) {
 201                         ret = -EINTR;
 202                         goto err;
 203                 }
 204
 205                 src_k = get_next_src(src_iter, src_end);
 206                 ret = bkey_err(src_k);
 207                 if (ret)
 208                         goto btree_err;
 209
 210                 src_done = bpos_min(src_iter->pos, src_end).offset -
 211                         src_start.offset;
 212                 dst_want = POS(dst_start.inode, dst_start.offset + src_done);
 213
 214                 if (bkey_cmp(dst_iter->pos, dst_want) < 0) {
 215                         ret = bch2_fpunch_at(&trans, dst_iter, dst_want,
 216                                              journal_seq, i_sectors_delta);
 217                         if (ret)
 218                                 goto btree_err;
 219                         continue;
 220                 }
 221
 222                 BUG_ON(bkey_cmp(dst_iter->pos, dst_want));
 223
 224                 if (!bkey_cmp(dst_iter->pos, dst_end))
 225                         break;
 226
 227                 if (src_k.k->type == KEY_TYPE_extent) {
 228                         bkey_on_stack_realloc(&new_src, c, src_k.k->u64s);
 229                         bkey_reassemble(new_src.k, src_k);
 230                         src_k = bkey_i_to_s_c(new_src.k);
 231
 232                         bch2_cut_front(src_iter->pos,   new_src.k);
 233                         bch2_cut_back(src_end,          new_src.k);
 234
 235                         ret = bch2_make_extent_indirect(&trans, src_iter,
 236                                                 bkey_i_to_extent(new_src.k));
 237                         if (ret)
 238                                 goto btree_err;
 239
 240                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 241                 }
 242
 243                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 244                         struct bkey_s_c_reflink_p src_p =
 245                                 bkey_s_c_to_reflink_p(src_k);
 246                         struct bkey_i_reflink_p *dst_p =
 247                                 bkey_reflink_p_init(&new_dst.k);
 248
 249                         u64 offset = le64_to_cpu(src_p.v->idx) +
 250                                 (src_iter->pos.offset -
 251                                  bkey_start_offset(src_k.k));
 252
 253                         dst_p->v.idx = cpu_to_le64(offset);
 254                 } else {
 255                         BUG();
 256                 }
 257
 258                 new_dst.k.k.p = dst_iter->pos;
 259                 bch2_key_resize(&new_dst.k.k,
 260                                 min(src_k.k->p.offset - src_iter->pos.offset,
 261                                     dst_end.offset - dst_iter->pos.offset));
 262
 263                 ret = bch2_extent_update(&trans, dst_iter, &new_dst.k,
 264                                          NULL, journal_seq,
 265                                          new_i_size, i_sectors_delta);
 266                 if (ret)
 267                         goto btree_err;
 268
 269                 dst_done = dst_iter->pos.offset - dst_start.offset;
 270                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 271                 bch2_btree_iter_set_pos(src_iter, src_want);
 272 btree_err:
 273                 if (ret == -EINTR)
 274                         ret = 0;
 275                 if (ret)
 276                         goto err;
 277         }
 278
 279         BUG_ON(bkey_cmp(dst_iter->pos, dst_end));
 280 err:
 281         BUG_ON(bkey_cmp(dst_iter->pos, dst_end) > 0);
 282
 283         dst_done = dst_iter->pos.offset - dst_start.offset;
 284         new_i_size = min(dst_iter->pos.offset << 9, new_i_size);
 285
 286         bch2_trans_begin(&trans);
 287
 288         do {
 289                 struct bch_inode_unpacked inode_u;
 290                 struct btree_iter *inode_iter;
 291
 292                 inode_iter = bch2_inode_peek(&trans, &inode_u,
 293                                 dst_start.inode, BTREE_ITER_INTENT);
 294                 ret2 = PTR_ERR_OR_ZERO(inode_iter);
 295
 296                 if (!ret2 &&
 297                     inode_u.bi_size < new_i_size) {
 298                         inode_u.bi_size = new_i_size;
 299                         ret2  = bch2_inode_write(&trans, inode_iter, &inode_u) ?:
 300                                 bch2_trans_commit(&trans, NULL, journal_seq,
 301                                                   BTREE_INSERT_ATOMIC);
 302                 }
 303         } while (ret2 == -EINTR);
 304
 305         ret = bch2_trans_exit(&trans) ?: ret;
 306         bkey_on_stack_exit(&new_src, c);
 307
 308         percpu_ref_put(&c->writes);
 309
 310         return dst_done ?: ret ?: ret2;
 311 }