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 "bkey_on_stack.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "buckets.h"
  11 #include "extents.h"
  12 #include "io.h"
  13 #include "journal.h"
  14 #include "keylist.h"
  15 #include "migrate.h"
  16 #include "move.h"
  17 #include "replicas.h"
  18 #include "super-io.h"
  19
  20 static int drop_dev_ptrs(struct bch_fs *c, struct bkey_s k,
  21                          unsigned dev_idx, int flags, bool metadata)
  22 {
  23         unsigned replicas = metadata ? c->opts.metadata_replicas : c->opts.data_replicas;
  24         unsigned lost = metadata ? BCH_FORCE_IF_METADATA_LOST : BCH_FORCE_IF_DATA_LOST;
  25         unsigned degraded = metadata ? BCH_FORCE_IF_METADATA_DEGRADED : BCH_FORCE_IF_DATA_DEGRADED;
  26         unsigned nr_good;
  27
  28         bch2_bkey_drop_device(k, dev_idx);
  29
  30         nr_good = bch2_bkey_durability(c, k.s_c);
  31         if ((!nr_good && !(flags & lost)) ||
  32             (nr_good < replicas && !(flags & degraded)))
  33                 return -EINVAL;
  34
  35         return 0;
  36 }
  37
  38 static int __bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags,
  39                                    enum btree_id btree_id)
  40 {
  41         struct btree_trans trans;
  42         struct btree_iter *iter;
  43         struct bkey_s_c k;
  44         struct bkey_on_stack sk;
  45         int ret = 0;
  46
  47         bkey_on_stack_init(&sk);
  48         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
  49
  50         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
  51                                    BTREE_ITER_PREFETCH);
  52
  53         while ((k = bch2_btree_iter_peek(iter)).k &&
  54                !(ret = bkey_err(k))) {
  55                 if (!bch2_bkey_has_device(k, dev_idx)) {
  56                         ret = bch2_mark_bkey_replicas(c, k);
  57                         if (ret)
  58                                 break;
  59                         bch2_btree_iter_next(iter);
  60                         continue;
  61                 }
  62
  63                 bkey_on_stack_reassemble(&sk, c, k);
  64
  65                 ret = drop_dev_ptrs(c, bkey_i_to_s(sk.k),
  66                                     dev_idx, flags, false);
  67                 if (ret)
  68                         break;
  69
  70                 /*
  71                  * If the new extent no longer has any pointers, bch2_extent_normalize()
  72                  * will do the appropriate thing with it (turning it into a
  73                  * KEY_TYPE_error key, or just a discard if it was a cached extent)
  74                  */
  75                 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
  76
  77                 bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
  78
  79                 bch2_trans_update(&trans, iter, sk.k);
  80
  81                 ret = bch2_trans_commit(&trans, NULL, NULL,
  82                                         BTREE_INSERT_ATOMIC|
  83                                         BTREE_INSERT_NOFAIL);
  84
  85                 /*
  86                  * don't want to leave ret == -EINTR, since if we raced and
  87                  * something else overwrote the key we could spuriously return
  88                  * -EINTR below:
  89                  */
  90                 if (ret == -EINTR)
  91                         ret = 0;
  92                 if (ret)
  93                         break;
  94         }
  95
  96         ret = bch2_trans_exit(&trans) ?: ret;
  97         bkey_on_stack_exit(&sk, c);
  98
  99         BUG_ON(ret == -EINTR);
 100
 101         return ret;
 102 }
 103
 104 static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 105 {
 106         return  __bch2_dev_usrdata_drop(c, dev_idx, flags, BTREE_ID_EXTENTS) ?:
 107                 __bch2_dev_usrdata_drop(c, dev_idx, flags, BTREE_ID_REFLINK);
 108 }
 109
 110 static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 111 {
 112         struct btree_trans trans;
 113         struct btree_iter *iter;
 114         struct closure cl;
 115         struct btree *b;
 116         unsigned id;
 117         int ret;
 118
 119         /* don't handle this yet: */
 120         if (flags & BCH_FORCE_IF_METADATA_LOST)
 121                 return -EINVAL;
 122
 123         bch2_trans_init(&trans, c, 0, 0);
 124         closure_init_stack(&cl);
 125
 126         for (id = 0; id < BTREE_ID_NR; id++) {
 127                 for_each_btree_node(&trans, iter, id, POS_MIN,
 128                                     BTREE_ITER_PREFETCH, b) {
 129                         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 130                         struct bkey_i_btree_ptr *new_key;
 131 retry:
 132                         if (!bch2_bkey_has_device(bkey_i_to_s_c(&b->key),
 133                                                   dev_idx)) {
 134                                 /*
 135                                  * we might have found a btree node key we
 136                                  * needed to update, and then tried to update it
 137                                  * but got -EINTR after upgrading the iter, but
 138                                  * then raced and the node is now gone:
 139                                  */
 140                                 bch2_btree_iter_downgrade(iter);
 141
 142                                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 143                                 if (ret)
 144                                         goto err;
 145                         } else {
 146                                 bkey_copy(&tmp.k, &b->key);
 147                                 new_key = bkey_i_to_btree_ptr(&tmp.k);
 148
 149                                 ret = drop_dev_ptrs(c, bkey_i_to_s(&new_key->k_i),
 150                                                     dev_idx, flags, true);
 151                                 if (ret)
 152                                         goto err;
 153
 154                                 ret = bch2_btree_node_update_key(c, iter, b, new_key);
 155                                 if (ret == -EINTR) {
 156                                         b = bch2_btree_iter_peek_node(iter);
 157                                         goto retry;
 158                                 }
 159                                 if (ret)
 160                                         goto err;
 161                         }
 162                 }
 163                 bch2_trans_iter_free(&trans, iter);
 164         }
 165
 166         /* flush relevant btree updates */
 167         while (1) {
 168                 closure_wait_event(&c->btree_interior_update_wait,
 169                                    !bch2_btree_interior_updates_nr_pending(c) ||
 170                                    c->btree_roots_dirty);
 171                 if (!bch2_btree_interior_updates_nr_pending(c))
 172                         break;
 173                 bch2_journal_meta(&c->journal);
 174         }
 175
 176         ret = 0;
 177 err:
 178         ret = bch2_trans_exit(&trans) ?: ret;
 179
 180         BUG_ON(ret == -EINTR);
 181
 182         return ret;
 183 }
 184
 185 int bch2_dev_data_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 186 {
 187         return bch2_dev_usrdata_drop(c, dev_idx, flags) ?:
 188                 bch2_dev_metadata_drop(c, dev_idx, flags) ?:
 189                 bch2_replicas_gc2(c);
 190 }