]> git.sesse.net Git - bcachefs-tools-debian/blobdiff - libbcachefs/move.c
Move c_src dirs back to toplevel
[bcachefs-tools-debian] / libbcachefs / move.c
diff --git a/libbcachefs/move.c b/libbcachefs/move.c
new file mode 100644 (file)
index 0000000..7a33319
--- /dev/null
@@ -0,0 +1,1161 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
+#include "backpointers.h"
+#include "bkey_buf.h"
+#include "btree_gc.h"
+#include "btree_update.h"
+#include "btree_update_interior.h"
+#include "btree_write_buffer.h"
+#include "disk_groups.h"
+#include "ec.h"
+#include "errcode.h"
+#include "error.h"
+#include "inode.h"
+#include "io_read.h"
+#include "io_write.h"
+#include "journal_reclaim.h"
+#include "keylist.h"
+#include "move.h"
+#include "replicas.h"
+#include "snapshot.h"
+#include "super-io.h"
+#include "trace.h"
+
+#include <linux/ioprio.h>
+#include <linux/kthread.h>
+
+const char * const bch2_data_ops_strs[] = {
+#define x(t, n, ...) [n] = #t,
+       BCH_DATA_OPS()
+#undef x
+       NULL
+};
+
+static void trace_move_extent2(struct bch_fs *c, struct bkey_s_c k)
+{
+       if (trace_move_extent_enabled()) {
+               struct printbuf buf = PRINTBUF;
+
+               bch2_bkey_val_to_text(&buf, c, k);
+               trace_move_extent(c, buf.buf);
+               printbuf_exit(&buf);
+       }
+}
+
+static void trace_move_extent_read2(struct bch_fs *c, struct bkey_s_c k)
+{
+       if (trace_move_extent_read_enabled()) {
+               struct printbuf buf = PRINTBUF;
+
+               bch2_bkey_val_to_text(&buf, c, k);
+               trace_move_extent_read(c, buf.buf);
+               printbuf_exit(&buf);
+       }
+}
+
+struct moving_io {
+       struct list_head                read_list;
+       struct list_head                io_list;
+       struct move_bucket_in_flight    *b;
+       struct closure                  cl;
+       bool                            read_completed;
+
+       unsigned                        read_sectors;
+       unsigned                        write_sectors;
+
+       struct bch_read_bio             rbio;
+
+       struct data_update              write;
+       /* Must be last since it is variable size */
+       struct bio_vec                  bi_inline_vecs[];
+};
+
+static void move_free(struct moving_io *io)
+{
+       struct moving_context *ctxt = io->write.ctxt;
+
+       if (io->b)
+               atomic_dec(&io->b->count);
+
+       bch2_data_update_exit(&io->write);
+
+       mutex_lock(&ctxt->lock);
+       list_del(&io->io_list);
+       wake_up(&ctxt->wait);
+       mutex_unlock(&ctxt->lock);
+
+       kfree(io);
+}
+
+static void move_write_done(struct bch_write_op *op)
+{
+       struct moving_io *io = container_of(op, struct moving_io, write.op);
+       struct moving_context *ctxt = io->write.ctxt;
+
+       if (io->write.op.error)
+               ctxt->write_error = true;
+
+       atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
+       atomic_dec(&io->write.ctxt->write_ios);
+       move_free(io);
+       closure_put(&ctxt->cl);
+}
+
+static void move_write(struct moving_io *io)
+{
+       if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
+               move_free(io);
+               return;
+       }
+
+       closure_get(&io->write.ctxt->cl);
+       atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
+       atomic_inc(&io->write.ctxt->write_ios);
+
+       bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
+}
+
+struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
+{
+       struct moving_io *io =
+               list_first_entry_or_null(&ctxt->reads, struct moving_io, read_list);
+
+       return io && io->read_completed ? io : NULL;
+}
+
+static void move_read_endio(struct bio *bio)
+{
+       struct moving_io *io = container_of(bio, struct moving_io, rbio.bio);
+       struct moving_context *ctxt = io->write.ctxt;
+
+       atomic_sub(io->read_sectors, &ctxt->read_sectors);
+       atomic_dec(&ctxt->read_ios);
+       io->read_completed = true;
+
+       wake_up(&ctxt->wait);
+       closure_put(&ctxt->cl);
+}
+
+void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt)
+{
+       struct moving_io *io;
+
+       while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
+               bch2_trans_unlock_long(ctxt->trans);
+               list_del(&io->read_list);
+               move_write(io);
+       }
+}
+
+void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt)
+{
+       unsigned sectors_pending = atomic_read(&ctxt->write_sectors);
+
+       move_ctxt_wait_event(ctxt,
+               !atomic_read(&ctxt->write_sectors) ||
+               atomic_read(&ctxt->write_sectors) != sectors_pending);
+}
+
+void bch2_moving_ctxt_flush_all(struct moving_context *ctxt)
+{
+       move_ctxt_wait_event(ctxt, list_empty(&ctxt->reads));
+       bch2_trans_unlock_long(ctxt->trans);
+       closure_sync(&ctxt->cl);
+}
+
+void bch2_moving_ctxt_exit(struct moving_context *ctxt)
+{
+       struct bch_fs *c = ctxt->trans->c;
+
+       bch2_moving_ctxt_flush_all(ctxt);
+
+       EBUG_ON(atomic_read(&ctxt->write_sectors));
+       EBUG_ON(atomic_read(&ctxt->write_ios));
+       EBUG_ON(atomic_read(&ctxt->read_sectors));
+       EBUG_ON(atomic_read(&ctxt->read_ios));
+
+       mutex_lock(&c->moving_context_lock);
+       list_del(&ctxt->list);
+       mutex_unlock(&c->moving_context_lock);
+
+       bch2_trans_put(ctxt->trans);
+       memset(ctxt, 0, sizeof(*ctxt));
+}
+
+void bch2_moving_ctxt_init(struct moving_context *ctxt,
+                          struct bch_fs *c,
+                          struct bch_ratelimit *rate,
+                          struct bch_move_stats *stats,
+                          struct write_point_specifier wp,
+                          bool wait_on_copygc)
+{
+       memset(ctxt, 0, sizeof(*ctxt));
+
+       ctxt->trans     = bch2_trans_get(c);
+       ctxt->fn        = (void *) _RET_IP_;
+       ctxt->rate      = rate;
+       ctxt->stats     = stats;
+       ctxt->wp        = wp;
+       ctxt->wait_on_copygc = wait_on_copygc;
+
+       closure_init_stack(&ctxt->cl);
+
+       mutex_init(&ctxt->lock);
+       INIT_LIST_HEAD(&ctxt->reads);
+       INIT_LIST_HEAD(&ctxt->ios);
+       init_waitqueue_head(&ctxt->wait);
+
+       mutex_lock(&c->moving_context_lock);
+       list_add(&ctxt->list, &c->moving_context_list);
+       mutex_unlock(&c->moving_context_lock);
+}
+
+void bch2_move_stats_exit(struct bch_move_stats *stats, struct bch_fs *c)
+{
+       trace_move_data(c, stats);
+}
+
+void bch2_move_stats_init(struct bch_move_stats *stats, const char *name)
+{
+       memset(stats, 0, sizeof(*stats));
+       stats->data_type = BCH_DATA_user;
+       scnprintf(stats->name, sizeof(stats->name), "%s", name);
+}
+
+int bch2_move_extent(struct moving_context *ctxt,
+                    struct move_bucket_in_flight *bucket_in_flight,
+                    struct btree_iter *iter,
+                    struct bkey_s_c k,
+                    struct bch_io_opts io_opts,
+                    struct data_update_opts data_opts)
+{
+       struct btree_trans *trans = ctxt->trans;
+       struct bch_fs *c = trans->c;
+       struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+       struct moving_io *io;
+       const union bch_extent_entry *entry;
+       struct extent_ptr_decoded p;
+       unsigned sectors = k.k->size, pages;
+       int ret = -ENOMEM;
+
+       if (ctxt->stats)
+               ctxt->stats->pos = BBPOS(iter->btree_id, iter->pos);
+       trace_move_extent2(c, k);
+
+       bch2_data_update_opts_normalize(k, &data_opts);
+
+       if (!data_opts.rewrite_ptrs &&
+           !data_opts.extra_replicas) {
+               if (data_opts.kill_ptrs)
+                       return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
+               return 0;
+       }
+
+       /*
+        * Before memory allocations & taking nocow locks in
+        * bch2_data_update_init():
+        */
+       bch2_trans_unlock(trans);
+
+       /* write path might have to decompress data: */
+       bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+               sectors = max_t(unsigned, sectors, p.crc.uncompressed_size);
+
+       pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
+       io = kzalloc(sizeof(struct moving_io) +
+                    sizeof(struct bio_vec) * pages, GFP_KERNEL);
+       if (!io)
+               goto err;
+
+       INIT_LIST_HEAD(&io->io_list);
+       io->write.ctxt          = ctxt;
+       io->read_sectors        = k.k->size;
+       io->write_sectors       = k.k->size;
+
+       bio_init(&io->write.op.wbio.bio, NULL, io->bi_inline_vecs, pages, 0);
+       bio_set_prio(&io->write.op.wbio.bio,
+                    IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
+
+       if (bch2_bio_alloc_pages(&io->write.op.wbio.bio, sectors << 9,
+                                GFP_KERNEL))
+               goto err_free;
+
+       io->rbio.c              = c;
+       io->rbio.opts           = io_opts;
+       bio_init(&io->rbio.bio, NULL, io->bi_inline_vecs, pages, 0);
+       io->rbio.bio.bi_vcnt = pages;
+       bio_set_prio(&io->rbio.bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
+       io->rbio.bio.bi_iter.bi_size = sectors << 9;
+
+       io->rbio.bio.bi_opf             = REQ_OP_READ;
+       io->rbio.bio.bi_iter.bi_sector  = bkey_start_offset(k.k);
+       io->rbio.bio.bi_end_io          = move_read_endio;
+
+       ret = bch2_data_update_init(trans, iter, ctxt, &io->write, ctxt->wp,
+                                   io_opts, data_opts, iter->btree_id, k);
+       if (ret)
+               goto err_free_pages;
+
+       io->write.op.end_io = move_write_done;
+
+       if (ctxt->rate)
+               bch2_ratelimit_increment(ctxt->rate, k.k->size);
+
+       if (ctxt->stats) {
+               atomic64_inc(&ctxt->stats->keys_moved);
+               atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
+       }
+
+       if (bucket_in_flight) {
+               io->b = bucket_in_flight;
+               atomic_inc(&io->b->count);
+       }
+
+       this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
+       this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
+       trace_move_extent_read2(c, k);
+
+       mutex_lock(&ctxt->lock);
+       atomic_add(io->read_sectors, &ctxt->read_sectors);
+       atomic_inc(&ctxt->read_ios);
+
+       list_add_tail(&io->read_list, &ctxt->reads);
+       list_add_tail(&io->io_list, &ctxt->ios);
+       mutex_unlock(&ctxt->lock);
+
+       /*
+        * dropped by move_read_endio() - guards against use after free of
+        * ctxt when doing wakeup
+        */
+       closure_get(&ctxt->cl);
+       bch2_read_extent(trans, &io->rbio,
+                        bkey_start_pos(k.k),
+                        iter->btree_id, k, 0,
+                        BCH_READ_NODECODE|
+                        BCH_READ_LAST_FRAGMENT);
+       return 0;
+err_free_pages:
+       bio_free_pages(&io->write.op.wbio.bio);
+err_free:
+       kfree(io);
+err:
+       if (ret == -BCH_ERR_data_update_done)
+               return 0;
+
+       if (bch2_err_matches(ret, EROFS) ||
+           bch2_err_matches(ret, BCH_ERR_transaction_restart))
+               return ret;
+
+       count_event(c, move_extent_start_fail);
+
+       if (trace_move_extent_start_fail_enabled()) {
+               struct printbuf buf = PRINTBUF;
+
+               bch2_bkey_val_to_text(&buf, c, k);
+               prt_str(&buf, ": ");
+               prt_str(&buf, bch2_err_str(ret));
+               trace_move_extent_start_fail(c, buf.buf);
+               printbuf_exit(&buf);
+       }
+       return ret;
+}
+
+struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
+                         struct per_snapshot_io_opts *io_opts,
+                         struct bkey_s_c extent_k)
+{
+       struct bch_fs *c = trans->c;
+       u32 restart_count = trans->restart_count;
+       int ret = 0;
+
+       if (io_opts->cur_inum != extent_k.k->p.inode) {
+               io_opts->d.nr = 0;
+
+               ret = for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
+                                        BTREE_ITER_ALL_SNAPSHOTS, k, ({
+                       if (k.k->p.offset != extent_k.k->p.inode)
+                               break;
+
+                       if (!bkey_is_inode(k.k))
+                               continue;
+
+                       struct bch_inode_unpacked inode;
+                       BUG_ON(bch2_inode_unpack(k, &inode));
+
+                       struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
+                       bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
+
+                       darray_push(&io_opts->d, e);
+               }));
+               io_opts->cur_inum = extent_k.k->p.inode;
+       }
+
+       ret = ret ?: trans_was_restarted(trans, restart_count);
+       if (ret)
+               return ERR_PTR(ret);
+
+       if (extent_k.k->p.snapshot)
+               darray_for_each(io_opts->d, i)
+                       if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
+                               return &i->io_opts;
+
+       return &io_opts->fs_io_opts;
+}
+
+int bch2_move_get_io_opts_one(struct btree_trans *trans,
+                             struct bch_io_opts *io_opts,
+                             struct bkey_s_c extent_k)
+{
+       struct btree_iter iter;
+       struct bkey_s_c k;
+       int ret;
+
+       /* reflink btree? */
+       if (!extent_k.k->p.inode) {
+               *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
+               return 0;
+       }
+
+       k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
+                              SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
+                              BTREE_ITER_CACHED);
+       ret = bkey_err(k);
+       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+               return ret;
+
+       if (!ret && bkey_is_inode(k.k)) {
+               struct bch_inode_unpacked inode;
+               bch2_inode_unpack(k, &inode);
+               bch2_inode_opts_get(io_opts, trans->c, &inode);
+       } else {
+               *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
+       }
+
+       bch2_trans_iter_exit(trans, &iter);
+       return 0;
+}
+
+int bch2_move_ratelimit(struct moving_context *ctxt)
+{
+       struct bch_fs *c = ctxt->trans->c;
+       bool is_kthread = current->flags & PF_KTHREAD;
+       u64 delay;
+
+       if (ctxt->wait_on_copygc && c->copygc_running) {
+               bch2_moving_ctxt_flush_all(ctxt);
+               wait_event_killable(c->copygc_running_wq,
+                                   !c->copygc_running ||
+                                   (is_kthread && kthread_should_stop()));
+       }
+
+       do {
+               delay = ctxt->rate ? bch2_ratelimit_delay(ctxt->rate) : 0;
+
+               if (is_kthread && kthread_should_stop())
+                       return 1;
+
+               if (delay)
+                       move_ctxt_wait_event_timeout(ctxt,
+                                       freezing(current) ||
+                                       (is_kthread && kthread_should_stop()),
+                                       delay);
+
+               if (unlikely(freezing(current))) {
+                       bch2_moving_ctxt_flush_all(ctxt);
+                       try_to_freeze();
+               }
+       } while (delay);
+
+       /*
+        * XXX: these limits really ought to be per device, SSDs and hard drives
+        * will want different limits
+        */
+       move_ctxt_wait_event(ctxt,
+               atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
+               atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
+               atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
+               atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
+
+       return 0;
+}
+
+static int bch2_move_data_btree(struct moving_context *ctxt,
+                               struct bpos start,
+                               struct bpos end,
+                               move_pred_fn pred, void *arg,
+                               enum btree_id btree_id)
+{
+       struct btree_trans *trans = ctxt->trans;
+       struct bch_fs *c = trans->c;
+       struct per_snapshot_io_opts snapshot_io_opts;
+       struct bch_io_opts *io_opts;
+       struct bkey_buf sk;
+       struct btree_iter iter;
+       struct bkey_s_c k;
+       struct data_update_opts data_opts;
+       int ret = 0, ret2;
+
+       per_snapshot_io_opts_init(&snapshot_io_opts, c);
+       bch2_bkey_buf_init(&sk);
+
+       if (ctxt->stats) {
+               ctxt->stats->data_type  = BCH_DATA_user;
+               ctxt->stats->pos        = BBPOS(btree_id, start);
+       }
+
+       bch2_trans_iter_init(trans, &iter, btree_id, start,
+                            BTREE_ITER_PREFETCH|
+                            BTREE_ITER_ALL_SNAPSHOTS);
+
+       if (ctxt->rate)
+               bch2_ratelimit_reset(ctxt->rate);
+
+       while (!bch2_move_ratelimit(ctxt)) {
+               bch2_trans_begin(trans);
+
+               k = bch2_btree_iter_peek(&iter);
+               if (!k.k)
+                       break;
+
+               ret = bkey_err(k);
+               if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                       continue;
+               if (ret)
+                       break;
+
+               if (bkey_ge(bkey_start_pos(k.k), end))
+                       break;
+
+               if (ctxt->stats)
+                       ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
+
+               if (!bkey_extent_is_direct_data(k.k))
+                       goto next_nondata;
+
+               io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
+               ret = PTR_ERR_OR_ZERO(io_opts);
+               if (ret)
+                       continue;
+
+               memset(&data_opts, 0, sizeof(data_opts));
+               if (!pred(c, arg, k, io_opts, &data_opts))
+                       goto next;
+
+               /*
+                * The iterator gets unlocked by __bch2_read_extent - need to
+                * save a copy of @k elsewhere:
+                */
+               bch2_bkey_buf_reassemble(&sk, c, k);
+               k = bkey_i_to_s_c(sk.k);
+
+               ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
+               if (ret2) {
+                       if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
+                               continue;
+
+                       if (ret2 == -ENOMEM) {
+                               /* memory allocation failure, wait for some IO to finish */
+                               bch2_move_ctxt_wait_for_io(ctxt);
+                               continue;
+                       }
+
+                       /* XXX signal failure */
+                       goto next;
+               }
+next:
+               if (ctxt->stats)
+                       atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
+next_nondata:
+               bch2_btree_iter_advance(&iter);
+       }
+
+       bch2_trans_iter_exit(trans, &iter);
+       bch2_bkey_buf_exit(&sk, c);
+       per_snapshot_io_opts_exit(&snapshot_io_opts);
+
+       return ret;
+}
+
+int __bch2_move_data(struct moving_context *ctxt,
+                    struct bbpos start,
+                    struct bbpos end,
+                    move_pred_fn pred, void *arg)
+{
+       struct bch_fs *c = ctxt->trans->c;
+       enum btree_id id;
+       int ret = 0;
+
+       for (id = start.btree;
+            id <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
+            id++) {
+               ctxt->stats->pos = BBPOS(id, POS_MIN);
+
+               if (!btree_type_has_ptrs(id) ||
+                   !bch2_btree_id_root(c, id)->b)
+                       continue;
+
+               ret = bch2_move_data_btree(ctxt,
+                                      id == start.btree ? start.pos : POS_MIN,
+                                      id == end.btree   ? end.pos   : POS_MAX,
+                                      pred, arg, id);
+               if (ret)
+                       break;
+       }
+
+       return ret;
+}
+
+int bch2_move_data(struct bch_fs *c,
+                  struct bbpos start,
+                  struct bbpos end,
+                  struct bch_ratelimit *rate,
+                  struct bch_move_stats *stats,
+                  struct write_point_specifier wp,
+                  bool wait_on_copygc,
+                  move_pred_fn pred, void *arg)
+{
+
+       struct moving_context ctxt;
+       int ret;
+
+       bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
+       ret = __bch2_move_data(&ctxt, start, end, pred, arg);
+       bch2_moving_ctxt_exit(&ctxt);
+
+       return ret;
+}
+
+int bch2_evacuate_bucket(struct moving_context *ctxt,
+                          struct move_bucket_in_flight *bucket_in_flight,
+                          struct bpos bucket, int gen,
+                          struct data_update_opts _data_opts)
+{
+       struct btree_trans *trans = ctxt->trans;
+       struct bch_fs *c = trans->c;
+       bool is_kthread = current->flags & PF_KTHREAD;
+       struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
+       struct btree_iter iter;
+       struct bkey_buf sk;
+       struct bch_backpointer bp;
+       struct bch_alloc_v4 a_convert;
+       const struct bch_alloc_v4 *a;
+       struct bkey_s_c k;
+       struct data_update_opts data_opts;
+       unsigned dirty_sectors, bucket_size;
+       u64 fragmentation;
+       struct bpos bp_pos = POS_MIN;
+       int ret = 0;
+
+       trace_bucket_evacuate(c, &bucket);
+
+       bch2_bkey_buf_init(&sk);
+
+       /*
+        * We're not run in a context that handles transaction restarts:
+        */
+       bch2_trans_begin(trans);
+
+       bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
+                            bucket, BTREE_ITER_CACHED);
+       ret = lockrestart_do(trans,
+                       bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
+       bch2_trans_iter_exit(trans, &iter);
+
+       bch_err_msg(c, ret, "looking up alloc key");
+       if (ret)
+               goto err;
+
+       a = bch2_alloc_to_v4(k, &a_convert);
+       dirty_sectors = bch2_bucket_sectors_dirty(*a);
+       bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
+       fragmentation = a->fragmentation_lru;
+
+       ret = bch2_btree_write_buffer_tryflush(trans);
+       bch_err_msg(c, ret, "flushing btree write buffer");
+       if (ret)
+               goto err;
+
+       while (!(ret = bch2_move_ratelimit(ctxt))) {
+               if (is_kthread && kthread_should_stop())
+                       break;
+
+               bch2_trans_begin(trans);
+
+               ret = bch2_get_next_backpointer(trans, bucket, gen,
+                                               &bp_pos, &bp,
+                                               BTREE_ITER_CACHED);
+               if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                       continue;
+               if (ret)
+                       goto err;
+               if (bkey_eq(bp_pos, POS_MAX))
+                       break;
+
+               if (!bp.level) {
+                       k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
+                       ret = bkey_err(k);
+                       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                               continue;
+                       if (ret)
+                               goto err;
+                       if (!k.k)
+                               goto next;
+
+                       bch2_bkey_buf_reassemble(&sk, c, k);
+                       k = bkey_i_to_s_c(sk.k);
+
+                       ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
+                       if (ret) {
+                               bch2_trans_iter_exit(trans, &iter);
+                               continue;
+                       }
+
+                       data_opts = _data_opts;
+                       data_opts.target        = io_opts.background_target;
+                       data_opts.rewrite_ptrs = 0;
+
+                       unsigned i = 0;
+                       bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
+                               if (ptr->dev == bucket.inode) {
+                                       data_opts.rewrite_ptrs |= 1U << i;
+                                       if (ptr->cached) {
+                                               bch2_trans_iter_exit(trans, &iter);
+                                               goto next;
+                                       }
+                               }
+                               i++;
+                       }
+
+                       ret = bch2_move_extent(ctxt, bucket_in_flight,
+                                              &iter, k, io_opts, data_opts);
+                       bch2_trans_iter_exit(trans, &iter);
+
+                       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                               continue;
+                       if (ret == -ENOMEM) {
+                               /* memory allocation failure, wait for some IO to finish */
+                               bch2_move_ctxt_wait_for_io(ctxt);
+                               continue;
+                       }
+                       if (ret)
+                               goto err;
+
+                       if (ctxt->stats)
+                               atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
+               } else {
+                       struct btree *b;
+
+                       b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
+                       ret = PTR_ERR_OR_ZERO(b);
+                       if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
+                               continue;
+                       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                               continue;
+                       if (ret)
+                               goto err;
+                       if (!b)
+                               goto next;
+
+                       ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
+                       bch2_trans_iter_exit(trans, &iter);
+
+                       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                               continue;
+                       if (ret)
+                               goto err;
+
+                       if (ctxt->rate)
+                               bch2_ratelimit_increment(ctxt->rate,
+                                                        c->opts.btree_node_size >> 9);
+                       if (ctxt->stats) {
+                               atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
+                               atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
+                       }
+               }
+next:
+               bp_pos = bpos_nosnap_successor(bp_pos);
+       }
+
+       trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
+err:
+       bch2_bkey_buf_exit(&sk, c);
+       return ret;
+}
+
+typedef bool (*move_btree_pred)(struct bch_fs *, void *,
+                               struct btree *, struct bch_io_opts *,
+                               struct data_update_opts *);
+
+static int bch2_move_btree(struct bch_fs *c,
+                          struct bbpos start,
+                          struct bbpos end,
+                          move_btree_pred pred, void *arg,
+                          struct bch_move_stats *stats)
+{
+       bool kthread = (current->flags & PF_KTHREAD) != 0;
+       struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
+       struct moving_context ctxt;
+       struct btree_trans *trans;
+       struct btree_iter iter;
+       struct btree *b;
+       enum btree_id btree;
+       struct data_update_opts data_opts;
+       int ret = 0;
+
+       bch2_moving_ctxt_init(&ctxt, c, NULL, stats,
+                             writepoint_ptr(&c->btree_write_point),
+                             true);
+       trans = ctxt.trans;
+
+       stats->data_type = BCH_DATA_btree;
+
+       for (btree = start.btree;
+            btree <= min_t(unsigned, end.btree, btree_id_nr_alive(c) - 1);
+            btree ++) {
+               stats->pos = BBPOS(btree, POS_MIN);
+
+               if (!bch2_btree_id_root(c, btree)->b)
+                       continue;
+
+               bch2_trans_node_iter_init(trans, &iter, btree, POS_MIN, 0, 0,
+                                         BTREE_ITER_PREFETCH);
+retry:
+               ret = 0;
+               while (bch2_trans_begin(trans),
+                      (b = bch2_btree_iter_peek_node(&iter)) &&
+                      !(ret = PTR_ERR_OR_ZERO(b))) {
+                       if (kthread && kthread_should_stop())
+                               break;
+
+                       if ((cmp_int(btree, end.btree) ?:
+                            bpos_cmp(b->key.k.p, end.pos)) > 0)
+                               break;
+
+                       stats->pos = BBPOS(iter.btree_id, iter.pos);
+
+                       if (!pred(c, arg, b, &io_opts, &data_opts))
+                               goto next;
+
+                       ret = bch2_btree_node_rewrite(trans, &iter, b, 0) ?: ret;
+                       if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                               continue;
+                       if (ret)
+                               break;
+next:
+                       bch2_btree_iter_next_node(&iter);
+               }
+               if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+                       goto retry;
+
+               bch2_trans_iter_exit(trans, &iter);
+
+               if (kthread && kthread_should_stop())
+                       break;
+       }
+
+       bch_err_fn(c, ret);
+       bch2_moving_ctxt_exit(&ctxt);
+       bch2_btree_interior_updates_flush(c);
+
+       return ret;
+}
+
+static bool rereplicate_pred(struct bch_fs *c, void *arg,
+                            struct bkey_s_c k,
+                            struct bch_io_opts *io_opts,
+                            struct data_update_opts *data_opts)
+{
+       unsigned nr_good = bch2_bkey_durability(c, k);
+       unsigned replicas = bkey_is_btree_ptr(k.k)
+               ? c->opts.metadata_replicas
+               : io_opts->data_replicas;
+
+       if (!nr_good || nr_good >= replicas)
+               return false;
+
+       data_opts->target               = 0;
+       data_opts->extra_replicas       = replicas - nr_good;
+       data_opts->btree_insert_flags   = 0;
+       return true;
+}
+
+static bool migrate_pred(struct bch_fs *c, void *arg,
+                        struct bkey_s_c k,
+                        struct bch_io_opts *io_opts,
+                        struct data_update_opts *data_opts)
+{
+       struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+       struct bch_ioctl_data *op = arg;
+       unsigned i = 0;
+
+       data_opts->rewrite_ptrs         = 0;
+       data_opts->target               = 0;
+       data_opts->extra_replicas       = 0;
+       data_opts->btree_insert_flags   = 0;
+
+       bkey_for_each_ptr(ptrs, ptr) {
+               if (ptr->dev == op->migrate.dev)
+                       data_opts->rewrite_ptrs |= 1U << i;
+               i++;
+       }
+
+       return data_opts->rewrite_ptrs != 0;
+}
+
+static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
+                                  struct btree *b,
+                                  struct bch_io_opts *io_opts,
+                                  struct data_update_opts *data_opts)
+{
+       return rereplicate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
+}
+
+static bool migrate_btree_pred(struct bch_fs *c, void *arg,
+                              struct btree *b,
+                              struct bch_io_opts *io_opts,
+                              struct data_update_opts *data_opts)
+{
+       return migrate_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
+}
+
+static bool bformat_needs_redo(struct bkey_format *f)
+{
+       unsigned i;
+
+       for (i = 0; i < f->nr_fields; i++) {
+               unsigned unpacked_bits = bch2_bkey_format_current.bits_per_field[i];
+               u64 unpacked_mask = ~((~0ULL << 1) << (unpacked_bits - 1));
+               u64 field_offset = le64_to_cpu(f->field_offset[i]);
+
+               if (f->bits_per_field[i] > unpacked_bits)
+                       return true;
+
+               if ((f->bits_per_field[i] == unpacked_bits) && field_offset)
+                       return true;
+
+               if (((field_offset + ((1ULL << f->bits_per_field[i]) - 1)) &
+                    unpacked_mask) <
+                   field_offset)
+                       return true;
+       }
+
+       return false;
+}
+
+static bool rewrite_old_nodes_pred(struct bch_fs *c, void *arg,
+                                  struct btree *b,
+                                  struct bch_io_opts *io_opts,
+                                  struct data_update_opts *data_opts)
+{
+       if (b->version_ondisk != c->sb.version ||
+           btree_node_need_rewrite(b) ||
+           bformat_needs_redo(&b->format)) {
+               data_opts->target               = 0;
+               data_opts->extra_replicas       = 0;
+               data_opts->btree_insert_flags   = 0;
+               return true;
+       }
+
+       return false;
+}
+
+int bch2_scan_old_btree_nodes(struct bch_fs *c, struct bch_move_stats *stats)
+{
+       int ret;
+
+       ret = bch2_move_btree(c,
+                             BBPOS_MIN,
+                             BBPOS_MAX,
+                             rewrite_old_nodes_pred, c, stats);
+       if (!ret) {
+               mutex_lock(&c->sb_lock);
+               c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
+               c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
+               c->disk_sb.sb->version_min = c->disk_sb.sb->version;
+               bch2_write_super(c);
+               mutex_unlock(&c->sb_lock);
+       }
+
+       bch_err_fn(c, ret);
+       return ret;
+}
+
+static bool drop_extra_replicas_pred(struct bch_fs *c, void *arg,
+                            struct bkey_s_c k,
+                            struct bch_io_opts *io_opts,
+                            struct data_update_opts *data_opts)
+{
+       unsigned durability = bch2_bkey_durability(c, k);
+       unsigned replicas = bkey_is_btree_ptr(k.k)
+               ? c->opts.metadata_replicas
+               : io_opts->data_replicas;
+       const union bch_extent_entry *entry;
+       struct extent_ptr_decoded p;
+       unsigned i = 0;
+
+       bkey_for_each_ptr_decode(k.k, bch2_bkey_ptrs_c(k), p, entry) {
+               unsigned d = bch2_extent_ptr_durability(c, &p);
+
+               if (d && durability - d >= replicas) {
+                       data_opts->kill_ptrs |= BIT(i);
+                       durability -= d;
+               }
+
+               i++;
+       }
+
+       return data_opts->kill_ptrs != 0;
+}
+
+static bool drop_extra_replicas_btree_pred(struct bch_fs *c, void *arg,
+                                  struct btree *b,
+                                  struct bch_io_opts *io_opts,
+                                  struct data_update_opts *data_opts)
+{
+       return drop_extra_replicas_pred(c, arg, bkey_i_to_s_c(&b->key), io_opts, data_opts);
+}
+
+int bch2_data_job(struct bch_fs *c,
+                 struct bch_move_stats *stats,
+                 struct bch_ioctl_data op)
+{
+       struct bbpos start      = BBPOS(op.start_btree, op.start_pos);
+       struct bbpos end        = BBPOS(op.end_btree, op.end_pos);
+       int ret = 0;
+
+       if (op.op >= BCH_DATA_OP_NR)
+               return -EINVAL;
+
+       bch2_move_stats_init(stats, bch2_data_ops_strs[op.op]);
+
+       switch (op.op) {
+       case BCH_DATA_OP_rereplicate:
+               stats->data_type = BCH_DATA_journal;
+               ret = bch2_journal_flush_device_pins(&c->journal, -1);
+               ret = bch2_move_btree(c, start, end,
+                                     rereplicate_btree_pred, c, stats) ?: ret;
+               ret = bch2_move_data(c, start, end,
+                                    NULL,
+                                    stats,
+                                    writepoint_hashed((unsigned long) current),
+                                    true,
+                                    rereplicate_pred, c) ?: ret;
+               ret = bch2_replicas_gc2(c) ?: ret;
+               break;
+       case BCH_DATA_OP_migrate:
+               if (op.migrate.dev >= c->sb.nr_devices)
+                       return -EINVAL;
+
+               stats->data_type = BCH_DATA_journal;
+               ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
+               ret = bch2_move_btree(c, start, end,
+                                     migrate_btree_pred, &op, stats) ?: ret;
+               ret = bch2_move_data(c, start, end,
+                                    NULL,
+                                    stats,
+                                    writepoint_hashed((unsigned long) current),
+                                    true,
+                                    migrate_pred, &op) ?: ret;
+               ret = bch2_replicas_gc2(c) ?: ret;
+               break;
+       case BCH_DATA_OP_rewrite_old_nodes:
+               ret = bch2_scan_old_btree_nodes(c, stats);
+               break;
+       case BCH_DATA_OP_drop_extra_replicas:
+               ret = bch2_move_btree(c, start, end,
+                               drop_extra_replicas_btree_pred, c, stats) ?: ret;
+               ret = bch2_move_data(c, start, end, NULL, stats,
+                               writepoint_hashed((unsigned long) current),
+                               true,
+                               drop_extra_replicas_pred, c) ?: ret;
+               ret = bch2_replicas_gc2(c) ?: ret;
+               break;
+       default:
+               ret = -EINVAL;
+       }
+
+       bch2_move_stats_exit(stats, c);
+       return ret;
+}
+
+void bch2_move_stats_to_text(struct printbuf *out, struct bch_move_stats *stats)
+{
+       prt_printf(out, "%s: data type=%s pos=",
+                  stats->name,
+                  bch2_data_types[stats->data_type]);
+       bch2_bbpos_to_text(out, stats->pos);
+       prt_newline(out);
+       printbuf_indent_add(out, 2);
+
+       prt_str(out, "keys moved:  ");
+       prt_u64(out, atomic64_read(&stats->keys_moved));
+       prt_newline(out);
+
+       prt_str(out, "keys raced:  ");
+       prt_u64(out, atomic64_read(&stats->keys_raced));
+       prt_newline(out);
+
+       prt_str(out, "bytes seen:  ");
+       prt_human_readable_u64(out, atomic64_read(&stats->sectors_seen) << 9);
+       prt_newline(out);
+
+       prt_str(out, "bytes moved: ");
+       prt_human_readable_u64(out, atomic64_read(&stats->sectors_moved) << 9);
+       prt_newline(out);
+
+       prt_str(out, "bytes raced: ");
+       prt_human_readable_u64(out, atomic64_read(&stats->sectors_raced) << 9);
+       prt_newline(out);
+
+       printbuf_indent_sub(out, 2);
+}
+
+static void bch2_moving_ctxt_to_text(struct printbuf *out, struct bch_fs *c, struct moving_context *ctxt)
+{
+       struct moving_io *io;
+
+       bch2_move_stats_to_text(out, ctxt->stats);
+       printbuf_indent_add(out, 2);
+
+       prt_printf(out, "reads: ios %u/%u sectors %u/%u",
+                  atomic_read(&ctxt->read_ios),
+                  c->opts.move_ios_in_flight,
+                  atomic_read(&ctxt->read_sectors),
+                  c->opts.move_bytes_in_flight >> 9);
+       prt_newline(out);
+
+       prt_printf(out, "writes: ios %u/%u sectors %u/%u",
+                  atomic_read(&ctxt->write_ios),
+                  c->opts.move_ios_in_flight,
+                  atomic_read(&ctxt->write_sectors),
+                  c->opts.move_bytes_in_flight >> 9);
+       prt_newline(out);
+
+       printbuf_indent_add(out, 2);
+
+       mutex_lock(&ctxt->lock);
+       list_for_each_entry(io, &ctxt->ios, io_list)
+               bch2_write_op_to_text(out, &io->write.op);
+       mutex_unlock(&ctxt->lock);
+
+       printbuf_indent_sub(out, 4);
+}
+
+void bch2_fs_moving_ctxts_to_text(struct printbuf *out, struct bch_fs *c)
+{
+       struct moving_context *ctxt;
+
+       mutex_lock(&c->moving_context_lock);
+       list_for_each_entry(ctxt, &c->moving_context_list, list)
+               bch2_moving_ctxt_to_text(out, c, ctxt);
+       mutex_unlock(&c->moving_context_lock);
+}
+
+void bch2_fs_move_init(struct bch_fs *c)
+{
+       INIT_LIST_HEAD(&c->moving_context_list);
+       mutex_init(&c->moving_context_lock);
+}