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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for moving data off a device.
   4  */
   5
   6 #include "bcachefs.h"
   7 #include "btree_update.h"
   8 #include "btree_update_interior.h"
   9 #include "buckets.h"
  10 #include "extents.h"
  11 #include "io.h"
  12 #include "journal.h"
  13 #include "keylist.h"
  14 #include "migrate.h"
  15 #include "move.h"
  16 #include "replicas.h"
  17 #include "super-io.h"
  18
  19 static int drop_dev_ptrs(struct bch_fs *c, struct bkey_s k,
  20                          unsigned dev_idx, int flags, bool metadata)
  21 {
  22         unsigned replicas = metadata ? c->opts.metadata_replicas : c->opts.data_replicas;
  23         unsigned lost = metadata ? BCH_FORCE_IF_METADATA_LOST : BCH_FORCE_IF_DATA_LOST;
  24         unsigned degraded = metadata ? BCH_FORCE_IF_METADATA_DEGRADED : BCH_FORCE_IF_DATA_DEGRADED;
  25         unsigned nr_good;
  26
  27         bch2_bkey_drop_device(k, dev_idx);
  28
  29         nr_good = bch2_bkey_durability(c, k.s_c);
  30         if ((!nr_good && !(flags & lost)) ||
  31             (nr_good < replicas && !(flags & degraded)))
  32                 return -EINVAL;
  33
  34         return 0;
  35 }
  36
  37 static int __bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags,
  38                                    enum btree_id btree_id)
  39 {
  40         struct btree_trans trans;
  41         struct btree_iter *iter;
  42         struct bkey_s_c k;
  43         BKEY_PADDED(key) tmp;
  44         int ret = 0;
  45
  46         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
  47
  48         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
  49                                    BTREE_ITER_PREFETCH);
  50
  51         while ((k = bch2_btree_iter_peek(iter)).k &&
  52                !(ret = bkey_err(k))) {
  53                 if (!bch2_bkey_has_device(k, dev_idx)) {
  54                         ret = bch2_mark_bkey_replicas(c, k);
  55                         if (ret)
  56                                 break;
  57                         bch2_btree_iter_next(iter);
  58                         continue;
  59                 }
  60
  61                 bkey_reassemble(&tmp.key, k);
  62
  63                 ret = drop_dev_ptrs(c, bkey_i_to_s(&tmp.key),
  64                                     dev_idx, flags, false);
  65                 if (ret)
  66                         break;
  67
  68                 /*
  69                  * If the new extent no longer has any pointers, bch2_extent_normalize()
  70                  * will do the appropriate thing with it (turning it into a
  71                  * KEY_TYPE_error key, or just a discard if it was a cached extent)
  72                  */
  73                 bch2_extent_normalize(c, bkey_i_to_s(&tmp.key));
  74
  75                 /* XXX not sketchy at all */
  76                 iter->pos = bkey_start_pos(&tmp.key.k);
  77
  78                 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &tmp.key));
  79
  80                 ret = bch2_trans_commit(&trans, NULL, NULL,
  81                                         BTREE_INSERT_ATOMIC|
  82                                         BTREE_INSERT_NOFAIL);
  83
  84                 /*
  85                  * don't want to leave ret == -EINTR, since if we raced and
  86                  * something else overwrote the key we could spuriously return
  87                  * -EINTR below:
  88                  */
  89                 if (ret == -EINTR)
  90                         ret = 0;
  91                 if (ret)
  92                         break;
  93         }
  94
  95         ret = bch2_trans_exit(&trans) ?: ret;
  96
  97         BUG_ON(ret == -EINTR);
  98
  99         return ret;
 100 }
 101
 102 static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 103 {
 104         return  __bch2_dev_usrdata_drop(c, dev_idx, flags, BTREE_ID_EXTENTS) ?:
 105                 __bch2_dev_usrdata_drop(c, dev_idx, flags, BTREE_ID_REFLINK);
 106 }
 107
 108 static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 109 {
 110         struct btree_trans trans;
 111         struct btree_iter *iter;
 112         struct closure cl;
 113         struct btree *b;
 114         unsigned id;
 115         int ret;
 116
 117         /* don't handle this yet: */
 118         if (flags & BCH_FORCE_IF_METADATA_LOST)
 119                 return -EINVAL;
 120
 121         bch2_trans_init(&trans, c, 0, 0);
 122         closure_init_stack(&cl);
 123
 124         for (id = 0; id < BTREE_ID_NR; id++) {
 125                 for_each_btree_node(&trans, iter, id, POS_MIN,
 126                                     BTREE_ITER_PREFETCH, b) {
 127                         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 128                         struct bkey_i_btree_ptr *new_key;
 129 retry:
 130                         if (!bch2_bkey_has_device(bkey_i_to_s_c(&b->key),
 131                                                   dev_idx)) {
 132                                 /*
 133                                  * we might have found a btree node key we
 134                                  * needed to update, and then tried to update it
 135                                  * but got -EINTR after upgrading the iter, but
 136                                  * then raced and the node is now gone:
 137                                  */
 138                                 bch2_btree_iter_downgrade(iter);
 139
 140                                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 141                                 if (ret)
 142                                         goto err;
 143                         } else {
 144                                 bkey_copy(&tmp.k, &b->key);
 145                                 new_key = bkey_i_to_btree_ptr(&tmp.k);
 146
 147                                 ret = drop_dev_ptrs(c, bkey_i_to_s(&new_key->k_i),
 148                                                     dev_idx, flags, true);
 149                                 if (ret)
 150                                         goto err;
 151
 152                                 ret = bch2_btree_node_update_key(c, iter, b, new_key);
 153                                 if (ret == -EINTR) {
 154                                         b = bch2_btree_iter_peek_node(iter);
 155                                         goto retry;
 156                                 }
 157                                 if (ret)
 158                                         goto err;
 159                         }
 160                 }
 161                 bch2_trans_iter_free(&trans, iter);
 162         }
 163
 164         /* flush relevant btree updates */
 165         while (1) {
 166                 closure_wait_event(&c->btree_interior_update_wait,
 167                                    !bch2_btree_interior_updates_nr_pending(c) ||
 168                                    c->btree_roots_dirty);
 169                 if (!bch2_btree_interior_updates_nr_pending(c))
 170                         break;
 171                 bch2_journal_meta(&c->journal);
 172         }
 173
 174         ret = 0;
 175 err:
 176         ret = bch2_trans_exit(&trans) ?: ret;
 177
 178         BUG_ON(ret == -EINTR);
 179
 180         return ret;
 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                 bch2_replicas_gc2(c);
 188 }