X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fdata_update.c;h=4150feca42a2e65e63a59234a3e806ebbd09e1ac;hb=d1900b637ee57c2e523e308cc335deee0164b882;hp=7ef7bb613347ea85bda4fafd1ae6ebcad94f0075;hpb=71111771690f244d13650c73d52ff601ad914d95;p=bcachefs-tools-debian

diff --git a/libbcachefs/data_update.c b/libbcachefs/data_update.c
index 7ef7bb6..4150fec 100644
--- a/libbcachefs/data_update.c
+++ b/libbcachefs/data_update.c
@@ -7,95 +7,85 @@
 #include "buckets.h"
 #include "data_update.h"
 #include "ec.h"
+#include "error.h"
 #include "extents.h"
-#include "io.h"
+#include "io_write.h"
 #include "keylist.h"
 #include "move.h"
 #include "nocow_locking.h"
+#include "rebalance.h"
 #include "subvolume.h"
+#include "trace.h"
 
-#include <trace/events/bcachefs.h>
-
-static int insert_snapshot_whiteouts(struct btree_trans *trans,
-				     enum btree_id id,
-				     struct bpos old_pos,
-				     struct bpos new_pos)
+static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k)
 {
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter, update_iter;
-	struct bkey_s_c k;
-	snapshot_id_list s;
-	int ret;
+	if (trace_move_extent_finish_enabled()) {
+		struct printbuf buf = PRINTBUF;
 
-	if (!btree_type_has_snapshots(id))
-		return 0;
+		bch2_bkey_val_to_text(&buf, c, k);
+		trace_move_extent_finish(c, buf.buf);
+		printbuf_exit(&buf);
+	}
+}
 
-	darray_init(&s);
+static void trace_move_extent_fail2(struct data_update *m,
+			 struct bkey_s_c new,
+			 struct bkey_s_c wrote,
+			 struct bkey_i *insert,
+			 const char *msg)
+{
+	struct bch_fs *c = m->op.c;
+	struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
+	const union bch_extent_entry *entry;
+	struct bch_extent_ptr *ptr;
+	struct extent_ptr_decoded p;
+	struct printbuf buf = PRINTBUF;
+	unsigned i, rewrites_found = 0;
 
-	if (bkey_eq(old_pos, new_pos))
-		return 0;
+	if (!trace_move_extent_fail_enabled())
+		return;
 
-	if (!snapshot_t(c, old_pos.snapshot)->children[0])
-		return 0;
+	prt_str(&buf, msg);
 
-	bch2_trans_iter_init(trans, &iter, id, old_pos,
-			     BTREE_ITER_NOT_EXTENTS|
-			     BTREE_ITER_ALL_SNAPSHOTS);
-	while (1) {
-		k = bch2_btree_iter_prev(&iter);
-		ret = bkey_err(k);
-		if (ret)
-			break;
+	if (insert) {
+		i = 0;
+		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
+			if (((1U << i) & m->data_opts.rewrite_ptrs) &&
+			    (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
+			    !ptr->cached)
+				rewrites_found |= 1U << i;
+			i++;
+		}
+	}
 
-		if (!bkey_eq(old_pos, k.k->p))
-			break;
+	prt_printf(&buf, "\nrewrite ptrs:   %u%u%u%u",
+		   (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
+		   (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
+		   (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
+		   (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
 
-		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
-			struct bkey_i *update;
+	prt_printf(&buf, "\nrewrites found: %u%u%u%u",
+		   (rewrites_found & (1 << 0)) != 0,
+		   (rewrites_found & (1 << 1)) != 0,
+		   (rewrites_found & (1 << 2)) != 0,
+		   (rewrites_found & (1 << 3)) != 0);
 
-			if (snapshot_list_has_ancestor(c, &s, k.k->p.snapshot))
-				continue;
+	prt_str(&buf, "\nold:    ");
+	bch2_bkey_val_to_text(&buf, c, old);
 
-			update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
+	prt_str(&buf, "\nnew:    ");
+	bch2_bkey_val_to_text(&buf, c, new);
 
-			ret = PTR_ERR_OR_ZERO(update);
-			if (ret)
-				break;
-
-			bkey_init(&update->k);
-			update->k.p = new_pos;
-			update->k.p.snapshot = k.k->p.snapshot;
-
-			bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
-					     BTREE_ITER_NOT_EXTENTS|
-					     BTREE_ITER_ALL_SNAPSHOTS|
-					     BTREE_ITER_INTENT);
-			ret   = bch2_btree_iter_traverse(&update_iter) ?:
-				bch2_trans_update(trans, &update_iter, update,
-					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
-			bch2_trans_iter_exit(trans, &update_iter);
-			if (ret)
-				break;
+	prt_str(&buf, "\nwrote:  ");
+	bch2_bkey_val_to_text(&buf, c, wrote);
 
-			ret = snapshot_list_add(c, &s, k.k->p.snapshot);
-			if (ret)
-				break;
-		}
+	if (insert) {
+		prt_str(&buf, "\ninsert: ");
+		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
 	}
-	bch2_trans_iter_exit(trans, &iter);
-	darray_exit(&s);
 
-	return ret;
-}
-
-static void bch2_bkey_mark_dev_cached(struct bkey_s k, unsigned dev)
-{
-	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
-	struct bch_extent_ptr *ptr;
-
-	bkey_for_each_ptr(ptrs, ptr)
-		if (ptr->dev == dev)
-			ptr->cached = true;
+	trace_move_extent_fail(c, buf.buf);
+	printbuf_exit(&buf);
 }
 
 static int __bch2_data_update_index_update(struct btree_trans *trans,
@@ -120,15 +110,17 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 	while (1) {
 		struct bkey_s_c k;
 		struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
-		struct bkey_i *insert;
+		struct bkey_i *insert = NULL;
 		struct bkey_i_extent *new;
-		const union bch_extent_entry *entry;
+		const union bch_extent_entry *entry_c;
+		union bch_extent_entry *entry;
 		struct extent_ptr_decoded p;
+		struct bch_extent_ptr *ptr;
+		const struct bch_extent_ptr *ptr_c;
 		struct bpos next_pos;
-		bool did_work = false;
 		bool should_check_enospc;
 		s64 i_sectors_delta = 0, disk_sectors_delta = 0;
-		unsigned i;
+		unsigned rewrites_found = 0, durability, i;
 
 		bch2_trans_begin(trans);
 
@@ -139,8 +131,11 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 
 		new = bkey_i_to_extent(bch2_keylist_front(keys));
 
-		if (!bch2_extents_match(k, old))
-			goto nomatch;
+		if (!bch2_extents_match(k, old)) {
+			trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i),
+						NULL, "no match:");
+			goto nowork;
+		}
 
 		bkey_reassemble(_insert.k, k);
 		insert = _insert.k;
@@ -163,40 +158,65 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 		 * Fist, drop rewrite_ptrs from @new:
 		 */
 		i = 0;
-		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
+		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
 			if (((1U << i) & m->data_opts.rewrite_ptrs) &&
-			    bch2_extent_has_ptr(old, p, bkey_i_to_s_c(insert))) {
-				/*
-				 * If we're going to be adding a pointer to the
-				 * same device, we have to drop the old one -
-				 * otherwise, we can just mark it cached:
-				 */
-				if (bch2_bkey_has_device(bkey_i_to_s_c(&new->k_i), p.ptr.dev))
-					bch2_bkey_drop_device_noerror(bkey_i_to_s(insert), p.ptr.dev);
-				else
-					bch2_bkey_mark_dev_cached(bkey_i_to_s(insert), p.ptr.dev);
+			    (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
+			    !ptr->cached) {
+				bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr);
+				rewrites_found |= 1U << i;
 			}
 			i++;
 		}
 
+		if (m->data_opts.rewrite_ptrs &&
+		    !rewrites_found &&
+		    bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) {
+			trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:");
+			goto nowork;
+		}
 
-		/* Add new ptrs: */
-		extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
-			if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
-				/*
-				 * raced with another move op? extent already
-				 * has a pointer to the device we just wrote
-				 * data to
-				 */
-				continue;
+		/*
+		 * A replica that we just wrote might conflict with a replica
+		 * that we want to keep, due to racing with another move:
+		 */
+restart_drop_conflicting_replicas:
+		extent_for_each_ptr(extent_i_to_s(new), ptr)
+			if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) &&
+			    !ptr_c->cached) {
+				bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr);
+				goto restart_drop_conflicting_replicas;
 			}
 
-			bch2_extent_ptr_decoded_append(insert, &p);
-			did_work = true;
+		if (!bkey_val_u64s(&new->k)) {
+			trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:");
+			goto nowork;
+		}
+
+		/* Now, drop pointers that conflict with what we just wrote: */
+		extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
+			if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev)))
+				bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr);
+
+		durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) +
+			bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i));
+
+		/* Now, drop excess replicas: */
+restart_drop_extra_replicas:
+		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) {
+			unsigned ptr_durability = bch2_extent_ptr_durability(c, &p);
+
+			if (!p.ptr.cached &&
+			    durability - ptr_durability >= m->op.opts.data_replicas) {
+				durability -= ptr_durability;
+
+				bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr);
+				goto restart_drop_extra_replicas;
+			}
 		}
 
-		if (!did_work)
-			goto nomatch;
+		/* Finally, add the pointers we just wrote: */
+		extent_for_each_ptr_decode(extent_i_to_s(new), p, entry)
+			bch2_extent_ptr_decoded_append(insert, &p);
 
 		bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 		bch2_extent_normalize(c, bkey_i_to_s(insert));
@@ -219,19 +239,65 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 
 		next_pos = insert->k.p;
 
-		ret   = insert_snapshot_whiteouts(trans, m->btree_id,
-						  k.k->p, insert->k.p) ?:
+		/*
+		 * Check for nonce offset inconsistency:
+		 * This is debug code - we've been seeing this bug rarely, and
+		 * it's been hard to reproduce, so this should give us some more
+		 * information when it does occur:
+		 */
+		struct printbuf err = PRINTBUF;
+		int invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id), 0, &err);
+		printbuf_exit(&err);
+
+		if (invalid) {
+			struct printbuf buf = PRINTBUF;
+
+			prt_str(&buf, "about to insert invalid key in data update path");
+			prt_str(&buf, "\nold: ");
+			bch2_bkey_val_to_text(&buf, c, old);
+			prt_str(&buf, "\nk:   ");
+			bch2_bkey_val_to_text(&buf, c, k);
+			prt_str(&buf, "\nnew: ");
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
+
+			bch2_print_string_as_lines(KERN_ERR, buf.buf);
+			printbuf_exit(&buf);
+
+			bch2_fatal_error(c);
+			goto out;
+		}
+
+		if (trace_data_update_enabled()) {
+			struct printbuf buf = PRINTBUF;
+
+			prt_str(&buf, "\nold: ");
+			bch2_bkey_val_to_text(&buf, c, old);
+			prt_str(&buf, "\nk:   ");
+			bch2_bkey_val_to_text(&buf, c, k);
+			prt_str(&buf, "\nnew: ");
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
+
+			trace_data_update(c, buf.buf);
+			printbuf_exit(&buf);
+		}
+
+		ret =   bch2_insert_snapshot_whiteouts(trans, m->btree_id,
+						k.k->p, bkey_start_pos(&insert->k)) ?:
+			bch2_insert_snapshot_whiteouts(trans, m->btree_id,
+						k.k->p, insert->k.p) ?:
+			bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
 			bch2_trans_update(trans, &iter, insert,
 				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 			bch2_trans_commit(trans, &op->res,
 				NULL,
-				BTREE_INSERT_NOFAIL|
+				BCH_TRANS_COMMIT_no_check_rw|
+				BCH_TRANS_COMMIT_no_enospc|
 				m->data_opts.btree_insert_flags);
 		if (!ret) {
 			bch2_btree_iter_set_pos(&iter, next_pos);
 
 			this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
-			trace_move_extent_finish(&new->k);
+			trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i));
 		}
 err:
 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
@@ -245,16 +311,15 @@ next:
 				goto out;
 		}
 		continue;
-nomatch:
-		if (m->ctxt) {
+nowork:
+		if (m->stats) {
 			BUG_ON(k.k->p.offset <= iter.pos.offset);
-			atomic64_inc(&m->ctxt->stats->keys_raced);
+			atomic64_inc(&m->stats->keys_raced);
 			atomic64_add(k.k->p.offset - iter.pos.offset,
-				     &m->ctxt->stats->sectors_raced);
+				     &m->stats->sectors_raced);
 		}
 
-		this_cpu_add(c->counters[BCH_COUNTER_move_extent_race], new->k.size);
-		trace_move_extent_race(&new->k);
+		count_event(c, move_extent_fail);
 
 		bch2_btree_iter_advance(&iter);
 		goto next;
@@ -269,15 +334,7 @@ out:
 
 int bch2_data_update_index_update(struct bch_write_op *op)
 {
-	struct bch_fs *c = op->c;
-	struct btree_trans trans;
-	int ret;
-
-	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
-	ret = __bch2_data_update_index_update(&trans, op);
-	bch2_trans_exit(&trans);
-
-	return ret;
+	return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op));
 }
 
 void bch2_data_update_read_done(struct data_update *m,
@@ -297,25 +354,26 @@ void bch2_data_update_exit(struct data_update *update)
 	struct bch_fs *c = update->op.c;
 	struct bkey_ptrs_c ptrs =
 		bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
-	const struct bch_extent_ptr *ptr;
 
-	bkey_for_each_ptr(ptrs, ptr)
-		bch2_bucket_nocow_unlock(&c->nocow_locks,
-				       PTR_BUCKET_POS(c, ptr), 0);
+	bkey_for_each_ptr(ptrs, ptr) {
+		if (c->opts.nocow_enabled)
+			bch2_bucket_nocow_unlock(&c->nocow_locks,
+						 PTR_BUCKET_POS(c, ptr), 0);
+		percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref);
+	}
 
 	bch2_bkey_buf_exit(&update->k, c);
 	bch2_disk_reservation_put(c, &update->op.res);
 	bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 }
 
-void bch2_update_unwritten_extent(struct btree_trans *trans,
+static void bch2_update_unwritten_extent(struct btree_trans *trans,
 				  struct data_update *update)
 {
 	struct bch_fs *c = update->op.c;
 	struct bio *bio = &update->op.wbio.bio;
 	struct bkey_i_extent *e;
 	struct write_point *wp;
-	struct bch_extent_ptr *ptr;
 	struct closure cl;
 	struct btree_iter iter;
 	struct bkey_s_c k;
@@ -348,7 +406,7 @@ void bch2_update_unwritten_extent(struct btree_trans *trans,
 				&update->op.devs_have,
 				update->op.nr_replicas,
 				update->op.nr_replicas,
-				update->op.alloc_reserve,
+				update->op.watermark,
 				0, &cl, &wp);
 		if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
 			bch2_trans_unlock(trans);
@@ -356,6 +414,8 @@ void bch2_update_unwritten_extent(struct btree_trans *trans,
 			continue;
 		}
 
+		bch_err_fn_ratelimited(c, ret);
+
 		if (ret)
 			return;
 
@@ -382,58 +442,116 @@ void bch2_update_unwritten_extent(struct btree_trans *trans,
 			break;
 	}
 
-	if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
+	if (closure_nr_remaining(&cl) != 1) {
 		bch2_trans_unlock(trans);
 		closure_sync(&cl);
 	}
 }
 
-int bch2_data_update_init(struct bch_fs *c, struct data_update *m,
+int bch2_extent_drop_ptrs(struct btree_trans *trans,
+			  struct btree_iter *iter,
+			  struct bkey_s_c k,
+			  struct data_update_opts data_opts)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_i *n;
+	int ret;
+
+	n = bch2_bkey_make_mut_noupdate(trans, k);
+	ret = PTR_ERR_OR_ZERO(n);
+	if (ret)
+		return ret;
+
+	while (data_opts.kill_ptrs) {
+		unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
+		struct bch_extent_ptr *ptr;
+
+		bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
+		data_opts.kill_ptrs ^= 1U << drop;
+	}
+
+	/*
+	 * If the new extent no longer has any pointers, bch2_extent_normalize()
+	 * will do the appropriate thing with it (turning it into a
+	 * KEY_TYPE_error key, or just a discard if it was a cached extent)
+	 */
+	bch2_extent_normalize(c, bkey_i_to_s(n));
+
+	/*
+	 * Since we're not inserting through an extent iterator
+	 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
+	 * we aren't using the extent overwrite path to delete, we're
+	 * just using the normal key deletion path:
+	 */
+	if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_IS_EXTENTS))
+		n->k.size = 0;
+
+	return bch2_trans_relock(trans) ?:
+		bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
+		bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
+}
+
+int bch2_data_update_init(struct btree_trans *trans,
+			  struct btree_iter *iter,
+			  struct moving_context *ctxt,
+			  struct data_update *m,
 			  struct write_point_specifier wp,
 			  struct bch_io_opts io_opts,
 			  struct data_update_opts data_opts,
 			  enum btree_id btree_id,
 			  struct bkey_s_c k)
 {
+	struct bch_fs *c = trans->c;
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
 	unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
-	int ret;
+	unsigned ptrs_locked = 0;
+	int ret = 0;
 
 	bch2_bkey_buf_init(&m->k);
 	bch2_bkey_buf_reassemble(&m->k, c, k);
 	m->btree_id	= btree_id;
 	m->data_opts	= data_opts;
+	m->ctxt		= ctxt;
+	m->stats	= ctxt ? ctxt->stats : NULL;
 
 	bch2_write_op_init(&m->op, c, io_opts);
 	m->op.pos	= bkey_start_pos(k.k);
 	m->op.version	= k.k->version;
 	m->op.target	= data_opts.target;
 	m->op.write_point = wp;
+	m->op.nr_replicas = 0;
 	m->op.flags	|= BCH_WRITE_PAGES_STABLE|
 		BCH_WRITE_PAGES_OWNED|
 		BCH_WRITE_DATA_ENCODED|
 		BCH_WRITE_MOVE|
 		m->data_opts.write_flags;
-	m->op.compression_type =
-		bch2_compression_opt_to_type[io_opts.background_compression ?:
-					     io_opts.compression];
-	if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
-		m->op.alloc_reserve = RESERVE_movinggc;
+	m->op.compression_opt	= background_compression(io_opts);
+	m->op.watermark		= m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
+
+	bkey_for_each_ptr(ptrs, ptr)
+		percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref);
+
+	unsigned durability_have = 0, durability_removing = 0;
 
 	i = 0;
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-		if (((1U << i) & m->data_opts.rewrite_ptrs) &&
-		    p.ptr.cached)
-			BUG();
+		bool locked;
 
-		if (!((1U << i) & m->data_opts.rewrite_ptrs))
-			bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
+		if (((1U << i) & m->data_opts.rewrite_ptrs)) {
+			BUG_ON(p.ptr.cached);
+
+			if (crc_is_compressed(p.crc))
+				reserve_sectors += k.k->size;
 
-		if (((1U << i) & m->data_opts.rewrite_ptrs) &&
-		    crc_is_compressed(p.crc))
-			reserve_sectors += k.k->size;
+			m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
+			durability_removing += bch2_extent_ptr_desired_durability(c, &p);
+		} else if (!p.ptr.cached &&
+			   !((1U << i) & m->data_opts.kill_ptrs)) {
+			bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
+			durability_have += bch2_extent_ptr_durability(c, &p);
+		}
 
 		/*
 		 * op->csum_type is normally initialized from the fs/file's
@@ -448,12 +566,54 @@ int bch2_data_update_init(struct bch_fs *c, struct data_update *m,
 		if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 			m->op.incompressible = true;
 
+		if (c->opts.nocow_enabled) {
+			if (ctxt) {
+				move_ctxt_wait_event(ctxt,
+						(locked = bch2_bucket_nocow_trylock(&c->nocow_locks,
+									  PTR_BUCKET_POS(c, &p.ptr), 0)) ||
+						(!atomic_read(&ctxt->read_sectors) &&
+						 !atomic_read(&ctxt->write_sectors)));
+
+				if (!locked)
+					bch2_bucket_nocow_lock(&c->nocow_locks,
+							       PTR_BUCKET_POS(c, &p.ptr), 0);
+			} else {
+				if (!bch2_bucket_nocow_trylock(&c->nocow_locks,
+							       PTR_BUCKET_POS(c, &p.ptr), 0)) {
+					ret = -BCH_ERR_nocow_lock_blocked;
+					goto err;
+				}
+			}
+			ptrs_locked |= (1U << i);
+		}
+
 		i++;
+	}
 
-		bch2_bucket_nocow_lock(&c->nocow_locks,
-				       PTR_BUCKET_POS(c, &p.ptr), 0);
+	/*
+	 * If current extent durability is less than io_opts.data_replicas,
+	 * we're not trying to rereplicate the extent up to data_replicas here -
+	 * unless extra_replicas was specified
+	 *
+	 * Increasing replication is an explicit operation triggered by
+	 * rereplicate, currently, so that users don't get an unexpected -ENOSPC
+	 */
+	if (!(m->data_opts.write_flags & BCH_WRITE_CACHED) &&
+	    durability_have >= io_opts.data_replicas) {
+		m->data_opts.kill_ptrs |= m->data_opts.rewrite_ptrs;
+		m->data_opts.rewrite_ptrs = 0;
+		/* if iter == NULL, it's just a promote */
+		if (iter)
+			ret = bch2_extent_drop_ptrs(trans, iter, k, m->data_opts);
+		goto done;
 	}
 
+	m->op.nr_replicas = min(durability_removing, io_opts.data_replicas - durability_have) +
+		m->data_opts.extra_replicas;
+	m->op.nr_replicas_required = m->op.nr_replicas;
+
+	BUG_ON(!m->op.nr_replicas);
+
 	if (reserve_sectors) {
 		ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 				m->data_opts.extra_replicas
@@ -463,29 +623,33 @@ int bch2_data_update_init(struct bch_fs *c, struct data_update *m,
 			goto err;
 	}
 
-	m->op.nr_replicas = m->op.nr_replicas_required =
-		hweight32(m->data_opts.rewrite_ptrs) + m->data_opts.extra_replicas;
-
-	BUG_ON(!m->op.nr_replicas);
+	if (bkey_extent_is_unwritten(k)) {
+		bch2_update_unwritten_extent(trans, m);
+		goto done;
+	}
 
-	/* Special handling required: */
-	if (bkey_extent_is_unwritten(k))
-		return -BCH_ERR_unwritten_extent_update;
 	return 0;
 err:
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		bch2_bucket_nocow_unlock(&c->nocow_locks,
-				       PTR_BUCKET_POS(c, &p.ptr), 0);
+	i = 0;
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+		if ((1U << i) & ptrs_locked)
+			bch2_bucket_nocow_unlock(&c->nocow_locks,
+						 PTR_BUCKET_POS(c, &p.ptr), 0);
+		percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref);
+		i++;
+	}
 
 	bch2_bkey_buf_exit(&m->k, c);
 	bch2_bio_free_pages_pool(c, &m->op.wbio.bio);
 	return ret;
+done:
+	bch2_data_update_exit(m);
+	return ret ?: -BCH_ERR_data_update_done;
 }
 
 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
 	unsigned i = 0;
 
 	bkey_for_each_ptr(ptrs, ptr) {