Merge tag 'bcachefs-2023-12-10' of https://evilpiepirate.org/git/bcachefs

Pull more bcachefs bugfixes from Kent Overstreet:

 - Fix a rare emergency shutdown path bug: dropping journal pins after
   the filesystem has mostly been torn down is not what we want.

 - Fix some concurrency issues with the btree write buffer and journal
   replay by not using the btree write buffer until journal replay is
   finished

 - A fixup from the prior patch to kill journal pre-reservations: at the
   start of the btree update path, where previously we took a
   pre-reservation, we do at least want to check the journal watermark.

 - Fix a race between dropping device metadata and btree node writes,
   which would re-add a pointer to a device that had just been dropped

 - Fix one of the SCRU lock warnings, in
   bch2_compression_stats_to_text().

 - Partial fix for a rare transaction paths overflow, when indirect
   extents had been split by background tasks, by not running certain
   triggers when they're not needed.

 - Fix for creating a snapshot with implicit source in a subdirectory of
   the containing subvolume

 - Don't unfreeze when we're emergency read-only

 - Fix for rebalance spinning trying to compress unwritten extentns

 - Another deleted_inodes fix, for directories

 - Fix a rare deadlock (usually just an unecessary wait) when flushing
   the journal with an open journal entry.

* tag 'bcachefs-2023-12-10' of https://evilpiepirate.org/git/bcachefs:
  bcachefs: Close journal entry if necessary when flushing all pins
  bcachefs: Fix uninitialized var in bch2_journal_replay()
  bcachefs: Fix deleted inode check for dirs
  bcachefs: rebalance shouldn't attempt to compress unwritten extents
  bcachefs: don't attempt rw on unfreeze when shutdown
  bcachefs: Fix creating snapshot with implict source
  bcachefs: Don't run indirect extent trigger unless inserting/deleting
  bcachefs: Convert compression_stats to for_each_btree_key2
  bcachefs: Fix bch2_extent_drop_ptrs() call
  bcachefs: Fix a journal deadlock in replay
  bcachefs; Don't use btree write buffer until journal replay is finished
  bcachefs: Don't drop journal pins in exit path

Author: torvalds

fs/bcachefs/btree_cache.c

@@ -9,6 +9,7 @@
 #include "debug.h"
 #include "errcode.h"
 #include "error.h"
+#include "journal.h"
 #include "trace.h"
 
 #include <linux/prefetch.h>
@@ -424,14 +425,11 @@ void bch2_fs_btree_cache_exit(struct bch_fs *c)
 		BUG_ON(btree_node_read_in_flight(b) ||
 		       btree_node_write_in_flight(b));
 
-		if (btree_node_dirty(b))
-			bch2_btree_complete_write(c, b, btree_current_write(b));
-		clear_btree_node_dirty_acct(c, b);
-
 		btree_node_data_free(c, b);
 	}
 
-	BUG_ON(atomic_read(&c->btree_cache.dirty));
+	BUG_ON(!bch2_journal_error(&c->journal) &&
+	       atomic_read(&c->btree_cache.dirty));
 
 	list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 

fs/bcachefs/btree_io.c

@@ -1704,8 +1704,8 @@ int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
 	return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
 }
 
-void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
-			      struct btree_write *w)
+static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
+				      struct btree_write *w)
 {
 	unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
 

fs/bcachefs/btree_io.h

@@ -134,9 +134,6 @@ void bch2_btree_node_read(struct bch_fs *, struct btree *, bool);
 int bch2_btree_root_read(struct bch_fs *, enum btree_id,
 			 const struct bkey_i *, unsigned);
 
-void bch2_btree_complete_write(struct bch_fs *, struct btree *,
-			      struct btree_write *);
-
 bool bch2_btree_post_write_cleanup(struct bch_fs *, struct btree *);
 
 enum btree_write_flags {

fs/bcachefs/btree_key_cache.c

@@ -992,8 +992,6 @@ void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 	list_for_each_entry_safe(ck, n, &items, list) {
 		cond_resched();
 
-		bch2_journal_pin_drop(&c->journal, &ck->journal);
-
 		list_del(&ck->list);
 		kfree(ck->k);
 		six_lock_exit(&ck->c.lock);

fs/bcachefs/btree_update.c

@@ -554,6 +554,19 @@ int __must_check bch2_trans_update_seq(struct btree_trans *trans, u64 seq,
 						 BTREE_UPDATE_PREJOURNAL);
 }
 
+static noinline int bch2_btree_insert_clone_trans(struct btree_trans *trans,
+						  enum btree_id btree,
+						  struct bkey_i *k)
+{
+	struct bkey_i *n = bch2_trans_kmalloc(trans, bkey_bytes(&k->k));
+	int ret = PTR_ERR_OR_ZERO(n);
+	if (ret)
+		return ret;
+
+	bkey_copy(n, k);
+	return bch2_btree_insert_trans(trans, btree, n, 0);
+}
+
 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
 					    enum btree_id btree,
 					    struct bkey_i *k)
@@ -564,6 +577,9 @@ int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
 	EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
 	EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
 
+	if (unlikely(trans->journal_replay_not_finished))
+		return bch2_btree_insert_clone_trans(trans, btree, k);
+
 	trans_for_each_wb_update(trans, i) {
 		if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
 			bkey_copy(&i->k, k);

fs/bcachefs/btree_update_interior.c

@@ -1056,6 +1056,17 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 	flags &= ~BCH_WATERMARK_MASK;
 	flags |= watermark;
 
+	if (!(flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
+	    watermark < c->journal.watermark) {
+		struct journal_res res = { 0 };
+
+		ret = drop_locks_do(trans,
+			bch2_journal_res_get(&c->journal, &res, 1,
+					     watermark|JOURNAL_RES_GET_CHECK));
+		if (ret)
+			return ERR_PTR(ret);
+	}
+
 	while (1) {
 		nr_nodes[!!update_level] += 1 + split;
 		update_level++;

fs/bcachefs/data_update.c

@@ -471,7 +471,7 @@ int bch2_extent_drop_ptrs(struct btree_trans *trans,
 	 * we aren't using the extent overwrite path to delete, we're
 	 * just using the normal key deletion path:
 	 */
-	if (bkey_deleted(&n->k))
+	if (bkey_deleted(&n->k) && !(iter->flags & BTREE_ITER_IS_EXTENTS))
 		n->k.size = 0;
 
 	return bch2_trans_relock(trans) ?:
@@ -591,7 +591,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 		m->data_opts.rewrite_ptrs = 0;
 		/* if iter == NULL, it's just a promote */
 		if (iter)
-			ret = bch2_extent_drop_ptrs(trans, iter, k, data_opts);
+			ret = bch2_extent_drop_ptrs(trans, iter, k, m->data_opts);
 		goto done;
 	}
 

fs/bcachefs/dirent.c

@@ -485,20 +485,15 @@ u64 bch2_dirent_lookup(struct bch_fs *c, subvol_inum dir,
 	return ret;
 }
 
-int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
+int bch2_empty_dir_snapshot(struct btree_trans *trans, u64 dir, u32 snapshot)
 {
 	struct btree_iter iter;
 	struct bkey_s_c k;
-	u32 snapshot;
 	int ret;
 
-	ret = bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot);
-	if (ret)
-		return ret;
-
 	for_each_btree_key_upto_norestart(trans, iter, BTREE_ID_dirents,
-			   SPOS(dir.inum, 0, snapshot),
-			   POS(dir.inum, U64_MAX), 0, k, ret)
+			   SPOS(dir, 0, snapshot),
+			   POS(dir, U64_MAX), 0, k, ret)
 		if (k.k->type == KEY_TYPE_dirent) {
 			ret = -ENOTEMPTY;
 			break;
@@ -508,6 +503,14 @@ int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
 	return ret;
 }
 
+int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
+{
+	u32 snapshot;
+
+	return bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot) ?:
+		bch2_empty_dir_snapshot(trans, dir.inum, snapshot);
+}
+
 int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
 {
 	struct btree_trans *trans = bch2_trans_get(c);

fs/bcachefs/dirent.h

@@ -64,6 +64,7 @@ u64 bch2_dirent_lookup(struct bch_fs *, subvol_inum,
 		       const struct bch_hash_info *,
 		       const struct qstr *, subvol_inum *);
 
+int bch2_empty_dir_snapshot(struct btree_trans *, u64, u32);
 int bch2_empty_dir_trans(struct btree_trans *, subvol_inum);
 int bch2_readdir(struct bch_fs *, subvol_inum, struct dir_context *);
 

fs/bcachefs/extents.c

@@ -1294,7 +1294,8 @@ unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
 		unsigned i = 0;
 
 		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible) {
+			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
+			    p.ptr.unwritten) {
 				rewrite_ptrs = 0;
 				goto incompressible;
 			}