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         bch2_trans_preload_iters(&trans);
  46
  47         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
  48                                    POS_MIN, BTREE_ITER_PREFETCH);
  49
  50         mutex_lock(&c->replicas_gc_lock);
  51         bch2_replicas_gc_start(c, (1 << BCH_DATA_USER)|(1 << BCH_DATA_CACHED));
  52
  53
  54         while ((k = bch2_btree_iter_peek(iter)).k &&
  55                !(ret = bkey_err(k))) {
  56                 if (!bkey_extent_is_data(k.k) ||
  57                     !bch2_extent_has_device(bkey_s_c_to_extent(k), dev_idx)) {
  58                         ret = bch2_mark_bkey_replicas(c, k);
  59                         if (ret)
  60                                 break;
  61                         bch2_btree_iter_next(iter);
  62                         continue;
  63                 }
  64
  65                 bkey_reassemble(&tmp.key, k);
  66
  67                 ret = drop_dev_ptrs(c, bkey_i_to_s(&tmp.key),
  68                                     dev_idx, flags, false);
  69                 if (ret)
  70                         break;
  71
  72                 /*
  73                  * If the new extent no longer has any pointers, bch2_extent_normalize()
  74                  * will do the appropriate thing with it (turning it into a
  75                  * KEY_TYPE_error key, or just a discard if it was a cached extent)
  76                  */
  77                 bch2_extent_normalize(c, bkey_i_to_s(&tmp.key));
  78
  79                 /* XXX not sketchy at all */
  80                 iter->pos = bkey_start_pos(&tmp.key.k);
  81
  82                 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &tmp.key));
  83
  84                 ret = bch2_trans_commit(&trans, NULL, NULL,
  85                                         BTREE_INSERT_ATOMIC|
  86                                         BTREE_INSERT_NOFAIL);
  87
  88                 /*
  89                  * don't want to leave ret == -EINTR, since if we raced and
  90                  * something else overwrote the key we could spuriously return
  91                  * -EINTR below:
  92                  */
  93                 if (ret == -EINTR)
  94                         ret = 0;
  95                 if (ret)
  96                         break;
  97         }
  98
  99         BUG_ON(ret == -EINTR);
 100
 101         bch2_trans_exit(&trans);
 102
 103         bch2_replicas_gc_end(c, ret);
 104         mutex_unlock(&c->replicas_gc_lock);
 105
 106         return ret;
 107 }
 108
 109 static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 110 {
 111         struct btree_trans trans;
 112         struct btree_iter *iter;
 113         struct closure cl;
 114         struct btree *b;
 115         unsigned id;
 116         int ret;
 117
 118         /* don't handle this yet: */
 119         if (flags & BCH_FORCE_IF_METADATA_LOST)
 120                 return -EINVAL;
 121
 122         bch2_trans_init(&trans, c);
 123         closure_init_stack(&cl);
 124
 125         mutex_lock(&c->replicas_gc_lock);
 126         bch2_replicas_gc_start(c, 1 << BCH_DATA_BTREE);
 127
 128         for (id = 0; id < BTREE_ID_NR; id++) {
 129                 for_each_btree_node(&trans, iter, id, POS_MIN,
 130                                     BTREE_ITER_PREFETCH, b) {
 131                         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 132                         struct bkey_i_btree_ptr *new_key;
 133 retry:
 134                         if (!bch2_bkey_has_device(bkey_i_to_s_c(&b->key),
 135                                                   dev_idx)) {
 136                                 /*
 137                                  * we might have found a btree node key we
 138                                  * needed to update, and then tried to update it
 139                                  * but got -EINTR after upgrading the iter, but
 140                                  * then raced and the node is now gone:
 141                                  */
 142                                 bch2_btree_iter_downgrade(iter);
 143
 144                                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 145                                 if (ret)
 146                                         goto err;
 147                         } else {
 148                                 bkey_copy(&tmp.k, &b->key);
 149                                 new_key = bkey_i_to_btree_ptr(&tmp.k);
 150
 151                                 ret = drop_dev_ptrs(c, bkey_i_to_s(&new_key->k_i),
 152                                                     dev_idx, flags, true);
 153                                 if (ret)
 154                                         goto err;
 155
 156                                 ret = bch2_btree_node_update_key(c, iter, b, new_key);
 157                                 if (ret == -EINTR) {
 158                                         b = bch2_btree_iter_peek_node(iter);
 159                                         goto retry;
 160                                 }
 161                                 if (ret)
 162                                         goto err;
 163                         }
 164                 }
 165                 bch2_trans_iter_free(&trans, iter);
 166         }
 167
 168         /* flush relevant btree updates */
 169         while (1) {
 170                 closure_wait_event(&c->btree_interior_update_wait,
 171                                    !bch2_btree_interior_updates_nr_pending(c) ||
 172                                    c->btree_roots_dirty);
 173                 if (!bch2_btree_interior_updates_nr_pending(c))
 174                         break;
 175                 bch2_journal_meta(&c->journal);
 176         }
 177
 178         ret = 0;
 179 err:
 180         bch2_trans_exit(&trans);
 181
 182         ret = bch2_replicas_gc_end(c, ret);
 183         mutex_unlock(&c->replicas_gc_lock);
 184
 185         return ret;
 186 }
 187
 188 int bch2_dev_data_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 189 {
 190         return bch2_dev_usrdata_drop(c, dev_idx, flags) ?:
 191                 bch2_dev_metadata_drop(c, dev_idx, flags);
 192 }