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 "buckets.h"
   8 #include "extents.h"
   9 #include "io.h"
  10 #include "journal.h"
  11 #include "keylist.h"
  12 #include "migrate.h"
  13 #include "move.h"
  14 #include "replicas.h"
  15 #include "super-io.h"
  16
  17 static int drop_dev_ptrs(struct bch_fs *c, struct bkey_s_extent e,
  18                          unsigned dev_idx, int flags, bool metadata)
  19 {
  20         unsigned replicas = metadata ? c->opts.metadata_replicas : c->opts.data_replicas;
  21         unsigned lost = metadata ? BCH_FORCE_IF_METADATA_LOST : BCH_FORCE_IF_DATA_LOST;
  22         unsigned degraded = metadata ? BCH_FORCE_IF_METADATA_DEGRADED : BCH_FORCE_IF_DATA_DEGRADED;
  23         unsigned nr_good;
  24
  25         bch2_extent_drop_device(e, dev_idx);
  26
  27         nr_good = bch2_extent_durability(c, e.c);
  28         if ((!nr_good && !(flags & lost)) ||
  29             (nr_good < replicas && !(flags & degraded)))
  30                 return -EINVAL;
  31
  32         return 0;
  33 }
  34
  35 static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
  36 {
  37         struct bkey_s_c k;
  38         struct bkey_s_extent e;
  39         BKEY_PADDED(key) tmp;
  40         struct btree_iter iter;
  41         int ret = 0;
  42
  43         mutex_lock(&c->replicas_gc_lock);
  44         bch2_replicas_gc_start(c, (1 << BCH_DATA_USER)|(1 << BCH_DATA_CACHED));
  45
  46         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS,
  47                              POS_MIN, BTREE_ITER_PREFETCH);
  48
  49         while ((k = bch2_btree_iter_peek(&iter)).k &&
  50                !(ret = btree_iter_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, BCH_DATA_USER, k);
  54                         if (ret)
  55                                 break;
  56                         bch2_btree_iter_next(&iter);
  57                         continue;
  58                 }
  59
  60                 bkey_reassemble(&tmp.key, k);
  61                 e = bkey_i_to_s_extent(&tmp.key);
  62
  63                 ret = drop_dev_ptrs(c, e, 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, e.s);
  73
  74                 ret = bch2_mark_bkey_replicas(c, BCH_DATA_USER,
  75                                               bkey_i_to_s_c(&tmp.key));
  76                 if (ret)
  77                         break;
  78
  79                 iter.pos = bkey_start_pos(&tmp.key.k);
  80
  81                 ret = bch2_btree_insert_at(c, NULL, NULL,
  82                                            BTREE_INSERT_ATOMIC|
  83                                            BTREE_INSERT_NOFAIL,
  84                                            BTREE_INSERT_ENTRY(&iter, &tmp.key));
  85
  86                 /*
  87                  * don't want to leave ret == -EINTR, since if we raced and
  88                  * something else overwrote the key we could spuriously return
  89                  * -EINTR below:
  90                  */
  91                 if (ret == -EINTR)
  92                         ret = 0;
  93                 if (ret)
  94                         break;
  95         }
  96
  97         bch2_btree_iter_unlock(&iter);
  98
  99         bch2_replicas_gc_end(c, ret);
 100         mutex_unlock(&c->replicas_gc_lock);
 101
 102         return ret;
 103 }
 104
 105 static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 106 {
 107         struct btree_iter iter;
 108         struct closure cl;
 109         struct btree *b;
 110         unsigned id;
 111         int ret;
 112
 113         /* don't handle this yet: */
 114         if (flags & BCH_FORCE_IF_METADATA_LOST)
 115                 return -EINVAL;
 116
 117         closure_init_stack(&cl);
 118
 119         mutex_lock(&c->replicas_gc_lock);
 120         bch2_replicas_gc_start(c, 1 << BCH_DATA_BTREE);
 121
 122         for (id = 0; id < BTREE_ID_NR; id++) {
 123                 for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 124                         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 125                         struct bkey_i_extent *new_key;
 126 retry:
 127                         if (!bch2_extent_has_device(bkey_i_to_s_c_extent(&b->key),
 128                                                     dev_idx)) {
 129                                 /*
 130                                  * we might have found a btree node key we
 131                                  * needed to update, and then tried to update it
 132                                  * but got -EINTR after upgrading the iter, but
 133                                  * then raced and the node is now gone:
 134                                  */
 135                                 bch2_btree_iter_downgrade(&iter);
 136
 137                                 ret = bch2_mark_bkey_replicas(c, BCH_DATA_BTREE,
 138                                                               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_extent(&tmp.k);
 144
 145                                 ret = drop_dev_ptrs(c, extent_i_to_s(new_key),
 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         ret = 0;
 163 out:
 164         ret = bch2_replicas_gc_end(c, ret);
 165         mutex_unlock(&c->replicas_gc_lock);
 166
 167         return ret;
 168 err:
 169         bch2_btree_iter_unlock(&iter);
 170         goto out;
 171 }
 172
 173 int bch2_dev_data_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 174 {
 175         return bch2_dev_usrdata_drop(c, dev_idx, flags) ?:
 176                 bch2_dev_metadata_drop(c, dev_idx, flags);
 177 }