git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/migrate.c

   1 /*
   2  * Code for moving data off a device.
   3  */
   4
   5 #include "bcachefs.h"
   6 #include "btree_update.h"
   7 #include "btree_update_interior.h"
   8 #include "buckets.h"
   9 #include "extents.h"
  10 #include "io.h"
  11 #include "journal.h"
  12 #include "keylist.h"
  13 #include "migrate.h"
  14 #include "move.h"
  15 #include "replicas.h"
  16 #include "super-io.h"
  17
  18 static int drop_dev_ptrs(struct bch_fs *c, struct bkey_s k,
  19                          unsigned dev_idx, int flags, bool metadata)
  20 {
  21         unsigned replicas = metadata ? c->opts.metadata_replicas : c->opts.data_replicas;
  22         unsigned lost = metadata ? BCH_FORCE_IF_METADATA_LOST : BCH_FORCE_IF_DATA_LOST;
  23         unsigned degraded = metadata ? BCH_FORCE_IF_METADATA_DEGRADED : BCH_FORCE_IF_DATA_DEGRADED;
  24         unsigned nr_good;
  25
  26         bch2_bkey_drop_device(k, dev_idx);
  27
  28         nr_good = bch2_bkey_durability(c, k.s_c);
  29         if ((!nr_good && !(flags & lost)) ||
  30             (nr_good < replicas && !(flags & degraded)))
  31                 return -EINVAL;
  32
  33         return 0;
  34 }
  35
  36 static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
  37 {
  38         struct btree_trans trans;
  39         struct btree_iter *iter;
  40         struct bkey_s_c k;
  41         BKEY_PADDED(key) tmp;
  42         int ret = 0;
  43
  44         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
  45
  46         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
  47                                    POS_MIN, BTREE_ITER_PREFETCH);
  48
  49         while ((k = bch2_btree_iter_peek(iter)).k &&
  50                !(ret = bkey_err(k))) {
  51                 if (!bkey_extent_is_data(k.k) ||
  52                     !bch2_extent_has_device(bkey_s_c_to_extent(k), dev_idx)) {
  53                         ret = bch2_mark_bkey_replicas(c, k);
  54                         if (ret)
  55                                 break;
  56                         bch2_btree_iter_next(iter);
  57                         continue;
  58                 }
  59
  60                 bkey_reassemble(&tmp.key, k);
  61
  62                 ret = drop_dev_ptrs(c, bkey_i_to_s(&tmp.key),
  63                                     dev_idx, flags, false);
  64                 if (ret)
  65                         break;
  66
  67                 /*
  68                  * If the new extent no longer has any pointers, bch2_extent_normalize()
  69                  * will do the appropriate thing with it (turning it into a
  70                  * KEY_TYPE_error key, or just a discard if it was a cached extent)
  71                  */
  72                 bch2_extent_normalize(c, bkey_i_to_s(&tmp.key));
  73
  74                 /* XXX not sketchy at all */
  75                 iter->pos = bkey_start_pos(&tmp.key.k);
  76
  77                 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &tmp.key));
  78
  79                 ret = bch2_trans_commit(&trans, NULL, NULL,
  80                                         BTREE_INSERT_ATOMIC|
  81                                         BTREE_INSERT_NOFAIL);
  82
  83                 /*
  84                  * don't want to leave ret == -EINTR, since if we raced and
  85                  * something else overwrote the key we could spuriously return
  86                  * -EINTR below:
  87                  */
  88                 if (ret == -EINTR)
  89                         ret = 0;
  90                 if (ret)
  91                         break;
  92         }
  93
  94         ret = bch2_trans_exit(&trans) ?: ret;
  95
  96         BUG_ON(ret == -EINTR);
  97
  98         return ret;
  99 }
 100
 101 static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 102 {
 103         struct btree_trans trans;
 104         struct btree_iter *iter;
 105         struct closure cl;
 106         struct btree *b;
 107         unsigned id;
 108         int ret;
 109
 110         /* don't handle this yet: */
 111         if (flags & BCH_FORCE_IF_METADATA_LOST)
 112                 return -EINVAL;
 113
 114         bch2_trans_init(&trans, c, 0, 0);
 115         closure_init_stack(&cl);
 116
 117         for (id = 0; id < BTREE_ID_NR; id++) {
 118                 for_each_btree_node(&trans, iter, id, POS_MIN,
 119                                     BTREE_ITER_PREFETCH, b) {
 120                         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 121                         struct bkey_i_btree_ptr *new_key;
 122 retry:
 123                         if (!bch2_bkey_has_device(bkey_i_to_s_c(&b->key),
 124                                                   dev_idx)) {
 125                                 /*
 126                                  * we might have found a btree node key we
 127                                  * needed to update, and then tried to update it
 128                                  * but got -EINTR after upgrading the iter, but
 129                                  * then raced and the node is now gone:
 130                                  */
 131                                 bch2_btree_iter_downgrade(iter);
 132
 133                                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 134                                 if (ret)
 135                                         goto err;
 136                         } else {
 137                                 bkey_copy(&tmp.k, &b->key);
 138                                 new_key = bkey_i_to_btree_ptr(&tmp.k);
 139
 140                                 ret = drop_dev_ptrs(c, bkey_i_to_s(&new_key->k_i),
 141                                                     dev_idx, flags, true);
 142                                 if (ret)
 143                                         goto err;
 144
 145                                 ret = bch2_btree_node_update_key(c, iter, b, new_key);
 146                                 if (ret == -EINTR) {
 147                                         b = bch2_btree_iter_peek_node(iter);
 148                                         goto retry;
 149                                 }
 150                                 if (ret)
 151                                         goto err;
 152                         }
 153                 }
 154                 bch2_trans_iter_free(&trans, iter);
 155         }
 156
 157         /* flush relevant btree updates */
 158         while (1) {
 159                 closure_wait_event(&c->btree_interior_update_wait,
 160                                    !bch2_btree_interior_updates_nr_pending(c) ||
 161                                    c->btree_roots_dirty);
 162                 if (!bch2_btree_interior_updates_nr_pending(c))
 163                         break;
 164                 bch2_journal_meta(&c->journal);
 165         }
 166
 167         ret = 0;
 168 err:
 169         ret = bch2_trans_exit(&trans) ?: ret;
 170
 171         BUG_ON(ret == -EINTR);
 172
 173         return ret;
 174 }
 175
 176 int bch2_dev_data_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 177 {
 178         return bch2_dev_usrdata_drop(c, dev_idx, flags) ?:
 179                 bch2_dev_metadata_drop(c, dev_idx, flags) ?:
 180                 bch2_replicas_gc2(c);
 181 }