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
  11 #include <linux/sched/signal.h>
  12
  13 static inline unsigned bkey_type_to_indirect(const struct bkey *k)
  14 {
  15         switch (k->type) {
  16         case KEY_TYPE_extent:
  17                 return KEY_TYPE_reflink_v;
  18         case KEY_TYPE_inline_data:
  19                 return KEY_TYPE_indirect_inline_data;
  20         default:
  21                 return 0;
  22         }
  23 }
  24
  25 /* reflink pointers */
  26
  27 const char *bch2_reflink_p_invalid(const struct bch_fs *c, struct bkey_s_c k)
  28 {
  29         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  30
  31         if (bkey_val_bytes(p.k) != sizeof(*p.v))
  32                 return "incorrect value size";
  33
  34         return NULL;
  35 }
  36
  37 void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
  38                             struct bkey_s_c k)
  39 {
  40         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
  41
  42         pr_buf(out, "idx %llu", le64_to_cpu(p.v->idx));
  43 }
  44
  45 bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  46 {
  47         struct bkey_s_reflink_p l = bkey_s_to_reflink_p(_l);
  48         struct bkey_s_c_reflink_p r = bkey_s_c_to_reflink_p(_r);
  49
  50         /*
  51          * Disabled for now, the triggers code needs to be reworked for merging
  52          * of reflink pointers to work:
  53          */
  54         return false;
  55
  56         if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
  57                 return false;
  58
  59         bch2_key_resize(l.k, l.k->size + r.k->size);
  60         return true;
  61 }
  62
  63 /* indirect extents */
  64
  65 const char *bch2_reflink_v_invalid(const struct bch_fs *c, struct bkey_s_c k)
  66 {
  67         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  68
  69         if (bkey_val_bytes(r.k) < sizeof(*r.v))
  70                 return "incorrect value size";
  71
  72         return bch2_bkey_ptrs_invalid(c, k);
  73 }
  74
  75 void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
  76                             struct bkey_s_c k)
  77 {
  78         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
  79
  80         pr_buf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
  81
  82         bch2_bkey_ptrs_to_text(out, c, k);
  83 }
  84
  85 bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
  86 {
  87         struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
  88         struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
  89
  90         return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
  91 }
  92
  93 /* indirect inline data */
  94
  95 const char *bch2_indirect_inline_data_invalid(const struct bch_fs *c,
  96                                               struct bkey_s_c k)
  97 {
  98         if (bkey_val_bytes(k.k) < sizeof(struct bch_indirect_inline_data))
  99                 return "incorrect value size";
 100         return NULL;
 101 }
 102
 103 void bch2_indirect_inline_data_to_text(struct printbuf *out,
 104                                         struct bch_fs *c, struct bkey_s_c k)
 105 {
 106         struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
 107         unsigned datalen = bkey_inline_data_bytes(k.k);
 108
 109         pr_buf(out, "refcount %llu datalen %u: %*phN",
 110                le64_to_cpu(d.v->refcount), datalen,
 111                min(datalen, 32U), d.v->data);
 112 }
 113
 114 static int bch2_make_extent_indirect(struct btree_trans *trans,
 115                                      struct btree_iter *extent_iter,
 116                                      struct bkey_i *orig)
 117 {
 118         struct bch_fs *c = trans->c;
 119         struct btree_iter *reflink_iter;
 120         struct bkey_s_c k;
 121         struct bkey_i *r_v;
 122         struct bkey_i_reflink_p *r_p;
 123         __le64 *refcount;
 124         int ret;
 125
 126         if (orig->k.type == KEY_TYPE_inline_data)
 127                 bch2_check_set_feature(c, BCH_FEATURE_reflink_inline_data);
 128
 129         for_each_btree_key(trans, reflink_iter, BTREE_ID_reflink,
 130                            POS(0, c->reflink_hint),
 131                            BTREE_ITER_INTENT|BTREE_ITER_SLOTS, k, ret) {
 132                 if (reflink_iter->pos.inode) {
 133                         bch2_btree_iter_set_pos(reflink_iter, POS_MIN);
 134                         continue;
 135                 }
 136
 137                 if (bkey_deleted(k.k) && orig->k.size <= k.k->size)
 138                         break;
 139         }
 140
 141         if (ret)
 142                 goto err;
 143
 144         /* rewind iter to start of hole, if necessary: */
 145         bch2_btree_iter_set_pos_to_extent_start(reflink_iter);
 146
 147         r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 148         ret = PTR_ERR_OR_ZERO(r_v);
 149         if (ret)
 150                 goto err;
 151
 152         bkey_init(&r_v->k);
 153         r_v->k.type     = bkey_type_to_indirect(&orig->k);
 154         r_v->k.p        = reflink_iter->pos;
 155         bch2_key_resize(&r_v->k, orig->k.size);
 156         r_v->k.version  = orig->k.version;
 157
 158         set_bkey_val_bytes(&r_v->k, sizeof(__le64) + bkey_val_bytes(&orig->k));
 159
 160         refcount        = bkey_refcount(r_v);
 161         *refcount       = 0;
 162         memcpy(refcount + 1, &orig->v, bkey_val_bytes(&orig->k));
 163
 164         ret = bch2_trans_update(trans, reflink_iter, r_v, 0);
 165         if (ret)
 166                 goto err;
 167
 168         orig->k.type = KEY_TYPE_reflink_p;
 169         r_p = bkey_i_to_reflink_p(orig);
 170         set_bkey_val_bytes(&r_p->k, sizeof(r_p->v));
 171         r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
 172
 173         ret = bch2_trans_update(trans, extent_iter, &r_p->k_i, 0);
 174 err:
 175         if (!IS_ERR(reflink_iter))
 176                 c->reflink_hint = reflink_iter->pos.offset;
 177         bch2_trans_iter_put(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         bch2_btree_iter_set_pos(iter, end);
 196         return bkey_s_c_null;
 197 }
 198
 199 s64 bch2_remap_range(struct bch_fs *c,
 200                      struct bpos dst_start, struct bpos src_start,
 201                      u64 remap_sectors, u64 *journal_seq,
 202                      u64 new_i_size, s64 *i_sectors_delta)
 203 {
 204         struct btree_trans trans;
 205         struct btree_iter *dst_iter, *src_iter;
 206         struct bkey_s_c src_k;
 207         struct bkey_buf new_dst, new_src;
 208         struct bpos dst_end = dst_start, src_end = src_start;
 209         struct bpos src_want;
 210         u64 dst_done;
 211         int ret = 0, ret2 = 0;
 212
 213         if (!percpu_ref_tryget(&c->writes))
 214                 return -EROFS;
 215
 216         bch2_check_set_feature(c, BCH_FEATURE_reflink);
 217
 218         dst_end.offset += remap_sectors;
 219         src_end.offset += remap_sectors;
 220
 221         bch2_bkey_buf_init(&new_dst);
 222         bch2_bkey_buf_init(&new_src);
 223         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 4096);
 224
 225         src_iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, src_start,
 226                                        BTREE_ITER_INTENT);
 227         dst_iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, dst_start,
 228                                        BTREE_ITER_INTENT);
 229
 230         while ((ret == 0 || ret == -EINTR) &&
 231                bkey_cmp(dst_iter->pos, dst_end) < 0) {
 232                 struct disk_reservation disk_res = { 0 };
 233
 234                 bch2_trans_begin(&trans);
 235
 236                 if (fatal_signal_pending(current)) {
 237                         ret = -EINTR;
 238                         break;
 239                 }
 240
 241                 dst_done = dst_iter->pos.offset - dst_start.offset;
 242                 src_want = POS(src_start.inode, src_start.offset + dst_done);
 243                 bch2_btree_iter_set_pos(src_iter, src_want);
 244
 245                 src_k = get_next_src(src_iter, src_end);
 246                 ret = bkey_err(src_k);
 247                 if (ret)
 248                         continue;
 249
 250                 if (bkey_cmp(src_want, src_iter->pos) < 0) {
 251                         ret = bch2_fpunch_at(&trans, dst_iter,
 252                                         bpos_min(dst_end,
 253                                                  POS(dst_iter->pos.inode, dst_iter->pos.offset +
 254                                                      src_iter->pos.offset - src_want.offset)),
 255                                                  journal_seq, i_sectors_delta);
 256                         continue;
 257                 }
 258
 259                 if (src_k.k->type != KEY_TYPE_reflink_p) {
 260                         bch2_btree_iter_set_pos_to_extent_start(src_iter);
 261
 262                         bch2_bkey_buf_reassemble(&new_src, c, src_k);
 263                         src_k = bkey_i_to_s_c(new_src.k);
 264
 265                         ret = bch2_make_extent_indirect(&trans, src_iter,
 266                                                 new_src.k);
 267                         if (ret)
 268                                 continue;
 269
 270                         BUG_ON(src_k.k->type != KEY_TYPE_reflink_p);
 271                 }
 272
 273                 if (src_k.k->type == KEY_TYPE_reflink_p) {
 274                         struct bkey_s_c_reflink_p src_p =
 275                                 bkey_s_c_to_reflink_p(src_k);
 276                         struct bkey_i_reflink_p *dst_p =
 277                                 bkey_reflink_p_init(new_dst.k);
 278
 279                         u64 offset = le64_to_cpu(src_p.v->idx) +
 280                                 (src_want.offset -
 281                                  bkey_start_offset(src_k.k));
 282
 283                         dst_p->v.idx = cpu_to_le64(offset);
 284                 } else {
 285                         BUG();
 286                 }
 287
 288                 new_dst.k->k.p = dst_iter->pos;
 289                 bch2_key_resize(&new_dst.k->k,
 290                                 min(src_k.k->p.offset - src_want.offset,
 291                                     dst_end.offset - dst_iter->pos.offset));
 292                 ret = bch2_extent_update(&trans, dst_iter, new_dst.k,
 293                                          &disk_res, journal_seq,
 294                                          new_i_size, i_sectors_delta,
 295                                          true);
 296                 bch2_disk_reservation_put(c, &disk_res);
 297         }
 298         bch2_trans_iter_put(&trans, dst_iter);
 299         bch2_trans_iter_put(&trans, src_iter);
 300
 301         BUG_ON(!ret && bkey_cmp(dst_iter->pos, dst_end));
 302         BUG_ON(bkey_cmp(dst_iter->pos, dst_end) > 0);
 303
 304         dst_done = dst_iter->pos.offset - dst_start.offset;
 305         new_i_size = min(dst_iter->pos.offset << 9, new_i_size);
 306
 307         bch2_trans_begin(&trans);
 308
 309         do {
 310                 struct bch_inode_unpacked inode_u;
 311                 struct btree_iter *inode_iter;
 312
 313                 inode_iter = bch2_inode_peek(&trans, &inode_u,
 314                                 dst_start.inode, BTREE_ITER_INTENT);
 315                 ret2 = PTR_ERR_OR_ZERO(inode_iter);
 316
 317                 if (!ret2 &&
 318                     inode_u.bi_size < new_i_size) {
 319                         inode_u.bi_size = new_i_size;
 320                         ret2  = bch2_inode_write(&trans, inode_iter, &inode_u) ?:
 321                                 bch2_trans_commit(&trans, NULL, journal_seq, 0);
 322                 }
 323
 324                 bch2_trans_iter_put(&trans, inode_iter);
 325         } while (ret2 == -EINTR);
 326
 327         ret = bch2_trans_exit(&trans) ?: ret;
 328         bch2_bkey_buf_exit(&new_src, c);
 329         bch2_bkey_buf_exit(&new_dst, c);
 330
 331         percpu_ref_put(&c->writes);
 332
 333         return dst_done ?: ret ?: ret2;
 334 }