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 const char *bch2_reflink_p_invalid(const struct bch_fs *c, struct bkey_s_c k)
  29 {
  30         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  31
  32         if (bkey_val_bytes(p.k) != sizeof(*p.v))
  33                 return "incorrect value size";
  34
  35         return NULL;
  36 }
  37
  38 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  39                             struct bkey_s_c k)
  40 {
  41         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  42
  43         pr_buf(out, "idx %llu", le64_to_cpu(p.v->idx));
  44 }
  45
  46 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  47 {
  48         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  49         struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
  50
  51         /*
  52          * Disabled for now, the triggers code needs to be reworked for merging
  53          * of reflink pointers to work:
  54          */
  55         return false;
  56
  57         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  58                 return false;
  59
  60         bch2_key_resize(l.k, l.k->size + r.k->size);
  61         return true;
  62 }
  63
  64 /* indirect extents */
  65
  66 const char *bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k)
  67 {
  68         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  69
  70         if (bkey_val_bytes(r.k) < sizeof(*r.v))
  71                 return "incorrect value size";
  72
  73         return bch2_bkey_ptrs_invalid(c, k);
  74 }
  75
  76 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  77                             struct bkey_s_c k)
  78 {
  79         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  80
  81         pr_buf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
  82
  83         bch2_bkey_ptrs_to_text(out, c, k);
  84 }
  85
  86 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  87 {
  88         struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
  89         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
  90
  91         return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
  92 }
  93
  94 /* indirect inline data */
  95
  96 const char *bch2_indirect_inline_data_invalid(const struct bch_fs *c,
  97                                               struct bkey_s_c k)
  98 {
  99         if (bkey_val_bytes(k.k) < sizeof(struct bch_indirect_inline_data))
 100                 return "incorrect value size";
 101         return NULL;
 102 }
 103
 104 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 105                                         struct bch_fs *c, struct bkey_s_c k)
 106 {
 107         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 108         unsigned datalen = bkey_inline_data_bytes(k.k);
 109
 110         pr_buf(out, "refcount %llu datalen %u: %*phN",
 111                le64_to_cpu(d.v->refcount), datalen,
 112                min(datalen, 32U), d.v->data);
 113 }
 114
 115 static int bch2_make_extent_indirect(struct btree_trans *trans,
 116                                      struct btree_iter *extent_iter,
 117                                      struct bkey_i *orig)
 118 {
 119         struct bch_fs *c = trans->c;
 120         struct btree_iter reflink_iter = { NULL };
 121         struct bkey_s_c k;
 122         struct bkey_i *r_v;
 123         struct bkey_i_reflink_p *r_p;
 124         __le64 *refcount;
 125         int ret;
 126
 127         if (orig->k.type == KEY_TYPE_inline_data)
 128                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 129
 130         for_each_btree_key(trans, reflink_iter, BTREE_ID_reflink,
 131                            POS(0, c->reflink_hint),
 132                            BTREE_ITER_INTENT|BTREE_ITER_SLOTS, k, ret) {
 133                 if (reflink_iter.pos.inode) {
 134                         bch2_btree_iter_set_pos(&reflink_iter, POS_MIN);
 135                         continue;
 136                 }
 137
 138                 if (bkey_deleted(k.k) && orig->k.size <= k.k->size)
 139                         break;
 140         }
 141
 142         if (ret)
 143                 goto err;
 144
 145         /* rewind iter to start of hole, if necessary: */
 146         bch2_btree_iter_set_pos_to_extent_start(&reflink_iter);
 147
 148         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 149         ret = PTR_ERR_OR_ZERO(r_v);
 150         if (ret)
 151                 goto err;
 152
 153         bkey_init(&r_v->k);
 154         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 155         r_v->k.p        = reflink_iter.pos;
 156         bch2_key_resize(&r_v->k, orig->k.size);
 157         r_v->k.version  = orig->k.version;
 158
 159         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 160
 161         refcount        = bkey_refcount(r_v);
 162         *refcount       = 0;
 163         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 164
 165         ret = bch2_trans_update(trans, &reflink_iter, r_v, 0);
 166         if (ret)
 167                 goto err;
 168
 169         orig->k.type = KEY_TYPE_reflink_p;
 170         r_p = bkey_i_to_reflink_p(orig);
 171         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 172         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 173
 174         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i, 0);
 175 err:
 176         c->reflink_hint = reflink_iter.pos.offset;
 177         bch2_trans_iter_exit(trans, &reflink_iter);
 178
 179         return ret;
 180 }
 181
 182 static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 183 {
 184         struct bkey_s_c k;
 185         int ret;
 186
 187         for_each_btree_key_continue(*iter, 0, k, ret) {
 188                 if (bkey_cmp(iter->pos, end) >= 0)
 189                         break;
 190
 191                 if (bkey_extent_is_data(k.k))
 192                         return k;
 193         }
 194
 195         if (bkey_cmp(iter->pos, end) >= 0)
 196                 bch2_btree_iter_set_pos(iter, end);
 197         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 198 }
 199
 200 s64 bch2_remap_range(struct bch_fs *c,
 201                      subvol_inum dst_inum, u64 dst_offset,
 202                      subvol_inum src_inum, u64 src_offset,
 203                      u64 remap_sectors, u64 *journal_seq,
 204                      u64 new_i_size, s64 *i_sectors_delta)
 205 {
 206         struct btree_trans trans;
 207         struct btree_iter dst_iter, src_iter;
 208         struct bkey_s_c src_k;
 209         struct bkey_buf new_dst, new_src;
 210         struct bpos dst_start = POS(dst_inum.inum, dst_offset);
 211         struct bpos src_start = POS(src_inum.inum, src_offset);
 212         struct bpos dst_end = dst_start, src_end = src_start;
 213         struct bpos src_want;
 214         u64 dst_done;
 215         u32 dst_snapshot, src_snapshot;
 216         int ret = 0, ret2 = 0;
 217
 218         if (!percpu_ref_tryget(&c->writes))
 219                 return -EROFS;
 220
 221         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 222
 223         dst_end.offset += remap_sectors;
 224         src_end.offset += remap_sectors;
 225
 226         bch2_bkey_buf_init(&new_dst);
 227         bch2_bkey_buf_init(&new_src);
 228         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 229
 230         bch2_trans_iter_init(&trans, &src_iter, BTREE_ID_extents, src_start,
 231                              BTREE_ITER_INTENT);
 232         bch2_trans_iter_init(&trans, &dst_iter, BTREE_ID_extents, dst_start,
 233                              BTREE_ITER_INTENT);
 234
 235         while ((ret == 0 || ret == -EINTR) &&
 236                bkey_cmp(dst_iter.pos, dst_end) < 0) {
 237                 struct disk_reservation disk_res = { 0 };
 238
 239                 bch2_trans_begin(&trans);
 240
 241                 if (fatal_signal_pending(current)) {
 242                         ret = -EINTR;
 243                         break;
 244                 }
 245
 246                 ret = bch2_subvolume_get_snapshot(&trans, src_inum.subvol,
 247                                                   &src_snapshot);
 248                 if (ret)
 249                         continue;
 250
 251                 bch2_btree_iter_set_snapshot(&src_iter, src_snapshot);
 252
 253                 ret = bch2_subvolume_get_snapshot(&trans, dst_inum.subvol,
 254                                                   &dst_snapshot);
 255                 if (ret)
 256                         continue;
 257
 258                 bch2_btree_iter_set_snapshot(&dst_iter, dst_snapshot);
 259
 260                 dst_done = dst_iter.pos.offset - dst_start.offset;
 261                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 262                 bch2_btree_iter_set_pos(&src_iter, src_want);
 263
 264                 src_k = get_next_src(&src_iter, src_end);
 265                 ret = bkey_err(src_k);
 266                 if (ret)
 267                         continue;
 268
 269                 if (bkey_cmp(src_want, src_iter.pos) < 0) {
 270                         ret = bch2_fpunch_at(&trans, &dst_iter, dst_inum,
 271                                         min(dst_end.offset,
 272                                             dst_iter.pos.offset +
 273                                             src_iter.pos.offset - src_want.offset),
 274                                         journal_seq, i_sectors_delta);
 275                         continue;
 276                 }
 277
 278                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 279                         bch2_btree_iter_set_pos_to_extent_start(&src_iter);
 280
 281                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 282                         src_k = bkey_i_to_s_c(new_src.k);
 283
 284                         ret = bch2_make_extent_indirect(&trans, &src_iter,
 285                                                 new_src.k);
 286                         if (ret)
 287                                 continue;
 288
 289                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 290                 }
 291
 292                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 293                         struct bkey_s_c_reflink_p src_p =
 294                                 bkey_s_c_to_reflink_p(src_k);
 295                         struct bkey_i_reflink_p *dst_p =
 296                                 bkey_reflink_p_init(new_dst.k);
 297
 298                         u64 offset = le64_to_cpu(src_p.v->idx) +
 299                                 (src_want.offset -
 300                                  bkey_start_offset(src_k.k));
 301
 302                         dst_p->v.idx = cpu_to_le64(offset);
 303                 } else {
 304                         BUG();
 305                 }
 306
 307                 new_dst.k->k.p = dst_iter.pos;
 308                 bch2_key_resize(&new_dst.k->k,
 309                                 min(src_k.k->p.offset - src_want.offset,
 310                                     dst_end.offset - dst_iter.pos.offset));
 311
 312                 ret = bch2_extent_update(&trans, dst_inum, &dst_iter,
 313                                          new_dst.k, &disk_res, journal_seq,
 314                                          new_i_size, i_sectors_delta,
 315                                          true);
 316                 bch2_disk_reservation_put(c, &disk_res);
 317         }
 318         bch2_trans_iter_exit(&trans, &dst_iter);
 319         bch2_trans_iter_exit(&trans, &src_iter);
 320
 321         BUG_ON(!ret && bkey_cmp(dst_iter.pos, dst_end));
 322         BUG_ON(bkey_cmp(dst_iter.pos, dst_end) > 0);
 323
 324         dst_done = dst_iter.pos.offset - dst_start.offset;
 325         new_i_size = min(dst_iter.pos.offset << 9, new_i_size);
 326
 327         do {
 328                 struct bch_inode_unpacked inode_u;
 329                 struct btree_iter inode_iter = { NULL };
 330
 331                 bch2_trans_begin(&trans);
 332
 333                 ret2 = bch2_inode_peek(&trans, &inode_iter, &inode_u,
 334                                        dst_inum, BTREE_ITER_INTENT);
 335
 336                 if (!ret2 &&
 337                     inode_u.bi_size < new_i_size) {
 338                         inode_u.bi_size = new_i_size;
 339                         ret2  = bch2_inode_write(&trans, &inode_iter, &inode_u) ?:
 340                                 bch2_trans_commit(&trans, NULL, journal_seq, 0);
 341                 }
 342
 343                 bch2_trans_iter_exit(&trans, &inode_iter);
 344         } while (ret2 == -EINTR);
 345
 346         ret = bch2_trans_exit(&trans) ?: ret;
 347         bch2_bkey_buf_exit(&new_src, c);
 348         bch2_bkey_buf_exit(&new_dst, c);
 349
 350         percpu_ref_put(&c->writes);
 351
 352         return dst_done ?: ret ?: ret2;
 353 }