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);
  45
  46         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
  47                                    POS_MIN, BTREE_ITER_PREFETCH);
  48
  49         mutex_lock(&c->replicas_gc_lock);
  50         bch2_replicas_gc_start(c, (1 << BCH_DATA_USER)|(1 << BCH_DATA_CACHED));
  51
  52
  53         while ((k = bch2_btree_iter_peek(iter)).k &&
  54                !(ret = btree_iter_err(k))) {
  55                 if (!bkey_extent_is_data(k.k) ||
  56                     !bch2_extent_has_device(bkey_s_c_to_extent(k), dev_idx)) {
  57                         ret = bch2_mark_bkey_replicas(c, k);
  58                         if (ret)
  59                                 break;
  60                         bch2_btree_iter_next(iter);
  61                         continue;
  62                 }
  63
  64                 bkey_reassemble(&tmp.key, k);
  65
  66                 ret = drop_dev_ptrs(c, bkey_i_to_s(&tmp.key),
  67                                     dev_idx, flags, false);
  68                 if (ret)
  69                         break;
  70
  71                 /*
  72                  * If the new extent no longer has any pointers, bch2_extent_normalize()
  73                  * will do the appropriate thing with it (turning it into a
  74                  * KEY_TYPE_error key, or just a discard if it was a cached extent)
  75                  */
  76                 bch2_extent_normalize(c, bkey_i_to_s(&tmp.key));
  77
  78                 /* XXX not sketchy at all */
  79                 iter->pos = bkey_start_pos(&tmp.key.k);
  80
  81                 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &tmp.key));
  82
  83                 ret = bch2_trans_commit(&trans, NULL, NULL,
  84                                         BTREE_INSERT_ATOMIC|
  85                                         BTREE_INSERT_NOFAIL);
  86
  87                 /*
  88                  * don't want to leave ret == -EINTR, since if we raced and
  89                  * something else overwrote the key we could spuriously return
  90                  * -EINTR below:
  91                  */
  92                 if (ret == -EINTR)
  93                         ret = 0;
  94                 if (ret)
  95                         break;
  96         }
  97
  98         bch2_trans_exit(&trans);
  99
 100         bch2_replicas_gc_end(c, ret);
 101         mutex_unlock(&c->replicas_gc_lock);
 102
 103         return ret;
 104 }
 105
 106 static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 107 {
 108         struct btree_iter iter;
 109         struct closure cl;
 110         struct btree *b;
 111         unsigned id;
 112         int ret;
 113
 114         /* don't handle this yet: */
 115         if (flags & BCH_FORCE_IF_METADATA_LOST)
 116                 return -EINVAL;
 117
 118         closure_init_stack(&cl);
 119
 120         mutex_lock(&c->replicas_gc_lock);
 121         bch2_replicas_gc_start(c, 1 << BCH_DATA_BTREE);
 122
 123         for (id = 0; id < BTREE_ID_NR; id++) {
 124                 for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 125                         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 126                         struct bkey_i_btree_ptr *new_key;
 127 retry:
 128                         if (!bch2_bkey_has_device(bkey_i_to_s_c(&b->key),
 129                                                   dev_idx)) {
 130                                 /*
 131                                  * we might have found a btree node key we
 132                                  * needed to update, and then tried to update it
 133                                  * but got -EINTR after upgrading the iter, but
 134                                  * then raced and the node is now gone:
 135                                  */
 136                                 bch2_btree_iter_downgrade(&iter);
 137
 138                                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 139                                 if (ret)
 140                                         goto err;
 141                         } else {
 142                                 bkey_copy(&tmp.k, &b->key);
 143                                 new_key = bkey_i_to_btree_ptr(&tmp.k);
 144
 145                                 ret = drop_dev_ptrs(c, bkey_i_to_s(&new_key->k_i),
 146                                                     dev_idx, flags, true);
 147                                 if (ret)
 148                                         goto err;
 149
 150                                 ret = bch2_btree_node_update_key(c, &iter, b, new_key);
 151                                 if (ret == -EINTR) {
 152                                         b = bch2_btree_iter_peek_node(&iter);
 153                                         goto retry;
 154                                 }
 155                                 if (ret)
 156                                         goto err;
 157                         }
 158                 }
 159                 bch2_btree_iter_unlock(&iter);
 160         }
 161
 162         /* flush relevant btree updates */
 163         while (1) {
 164                 closure_wait_event(&c->btree_interior_update_wait,
 165                                    !bch2_btree_interior_updates_nr_pending(c) ||
 166                                    c->btree_roots_dirty);
 167                 if (!bch2_btree_interior_updates_nr_pending(c))
 168                         break;
 169                 bch2_journal_meta(&c->journal);
 170         }
 171
 172         ret = 0;
 173 out:
 174         ret = bch2_replicas_gc_end(c, ret);
 175         mutex_unlock(&c->replicas_gc_lock);
 176
 177         return ret;
 178 err:
 179         bch2_btree_iter_unlock(&iter);
 180         goto out;
 181 }
 182
 183 int bch2_dev_data_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 184 {
 185         return bch2_dev_usrdata_drop(c, dev_idx, flags) ?:
 186                 bch2_dev_metadata_drop(c, dev_idx, flags);
 187 }